Exploration of Machine Learning techniques for Cloud Removal and Gap Filling on Sentinel-2 time series images for better Exploitation in Far North Cameroon

Exploration of Machine Learning techniques for Cloud Removal and Gap Filling on Sentinel-2 time series images for better Exploitation in Far North Cameroon | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Exploration of Machine Learning techniques for Cloud Removal and Gap Filling on Sentinel-2 time series images for better Exploitation in Far North Cameroon Mvogo Ngono Joseph, Noumsi Woguia, Wirba Pountianus Berinyuy This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5952159/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 9 You are reading this latest preprint version Abstract The direct exploitation and interpretation of optical satellite images, such as Sentinel-2 data, is significantly hampered by cloud cover. In this paper, we explore several machine learning algorithms in order to suggest a machine learning-based method for cloud removal and gap filling in Sentinel-2 satellite pictures for better utilization in the far north of Cameroon, concentrating on the city of Maroua. Our goal is to successfully fill these gaps produced by these cloud masks in the photos by using data from several photographs taken on various dates, assuming that cloud occurrence changes both geographically and temporally. The cloud-covered and cloud-free regions are analyzed using a variety of machine learning methods, such as Random Forest, VGG16 with Random Forest, VGG with dense layers, VGG16 with image data augmentation, SVM, and deep learning CNN models. We assess the correctness of each model and compare their performance through rigorous experimentation. Our findings show that the VGG16 model with the addition of picture data had the best accuracy, at about 95%. Machine-Learning Deep-Learning Cloud Removing Gap Filling Sentinel-2 Satellite images SVM CNN VGG16 Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Figure 13 Figure 14 Figure 15 Article Highlights Cloud cover affects the effectiveness of Sentinel-2 satellite imagery in Maroua, Cameroon. Multiple machine learning techniques were evaluated for efficient cloud removal and image restoration. The VGG16 model with enhanced data yielded the best performance, reaching an accuracy of around 95%. 1. Introduction Land cover mapping, environmental monitoring, and natural resource management are just a few of the uses for which satellite photography, such as that from the Sentinel-2 platform, is useful. However, the presence of clouds prevents direct exploitation and analysis of these photos, thus restricting their use. In areas with considerable atmospheric moisture, like the extreme north of Cameroon, centered around the city of Maroua, cloud cover is very common. It is essential to create efficient methods for cloud removal and gap filling to fully utilize the potential of Sentinel-2 images in this area. The applicability of optical remote sensing images is severely constrained by clouds. In this study, we provide a unique approach to cloud removal from satellite photos that treats ground surface reflections and cloud top reflections as a linear combination of image components from the standpoint of image superposition. We first recover the ground surface information of thin cloud areas using a two-step convolutional neural network to extract the transparency information of clouds. Figure 1 below shows the general methodology used. This work additionally enhances the binary Tversky loss function and uses it on multi-classification tasks in light of the unbalanced nature of the produced data. On the simulated dataset and the ALCD dataset, respectively, the model was verified. The findings demonstrate that this model performed better in cloud identification and removal than previous control group studies.[ 1 ] To generate cloud-free versions of satellite photos, cloud removal requires locating and masking cloud-covered regions. By restoring the information that clouds had masked, this procedure intends to improve picture analysis and utilization. Additionally, the term "gap filling" describes the process of adding missing or empty areas to images as a result of the cloud masking technique. A more thorough and continuous depiction of the landscape may be generated by filling in these gaps with reliable data from other pictures. In this paper, we suggest a machine learning-based method for removing clouds and filling in gaps in Sentinel-2 satellite photos in Cameroon's far north. Our technique comprises collecting a set of optical satellite photos from the Sentinel-2 platform for a predetermined period, making use of the capabilities of Google Earth Engine and Google Colaboratory. The outcome is a group of photos with clouds removed using sophisticated machine learning methods and methodologies. By concatenating photos taken on several dates, we make use of the temporal fluctuation of cloud cover to fill in the gaps left by cloud masking. This is based on the assumption that cloud coverage fluctuates both geographically and temporally and that locations that are clouded in one photograph are probably clear in another. We may fill in the blanks and provide a more accurate depiction of the terrain by combining many photographs. To do this, we investigate several machine learning models, such as SVM, deep learning CNN models, Random Forest, VGG16 with Random Forest, VGG with dense layers, and VGG16 with picture data augmentation. The regions in the satellite photos that are cloud-covered and cloud-free are analyzed and categorized using these models. To determine the most efficient method for cloud removal and gap filling in the Sentinel-2 images of Cameroon's far north, we conduct thorough testing and assessment, evaluating the accuracy and performance of each model. The findings of this study have substantial ramifications for cloud-prone locations' remote sensing applications. Our method increases the possibility for accurate land cover mapping, environmental monitoring, and resource management in the far north area of Cameroon by increasing the usefulness of Sentinel-2 data through cloud removal and gap filling. Additionally, the knowledge gathered from this study will help in the development of reliable methods for cloud removal and gap-filling in satellite data, which would assist related regions and advance remote sensing applications globally. The rest of the article is organized as follows: Section 2 (Literature Review) synthesizes existing methodologies for cloud removal and gap filling in satellite imagery, highlighting critical gaps in handling persistent cloud cover. Section 3 (Materials and Methods) details the datasets, preprocessing workflows, and evaluation metrics employed, with emphasis on Sentinel-2 data acquisition for Cameroon’s Far North. Section 4 (The Proposed Model) introduces the architecture of the machine learning framework, including its novel loss function and temporal data integration strategy. Section 5 (Results) presents quantitative and qualitative outcomes of cloud detection and removal performance. Section 6 (Discussion) contextualizes these findings, emphasizing their implications for environmental monitoring applications. Section 7 (Limitations) examines constraints related to data availability and model generalizability. Section 8 (Comparative Analyses) benchmarks the proposed approach against state-of-the-art techniques. Finally, Section 9 (Conclusion) summarizes key contributions and suggests future research directions. 2. Literature Review Cloud cover presents a recurring problem when using and analyzing satellite data, especially for optical remote sensing applications. Researchers have put forth several strategies and algorithms over the years to solve the problems brought on by cloud cover and enhance the utility of satellite data. We cover important methods and developments in cloud removal and gap filling in this literature review, concentrating on the context of Sentinel-2 satellite photos in Cameroon's far north. Current cloud removal methods, such as multitemporal compositing [ 14 ] and spectral unmixing [ 15 ] often prioritize global accuracy over localized fidelity, struggling with persistent cloud cover in semi-arid regions like Cameroon’s Far North. Hybrid approaches combining CNNs and attention mechanisms [ 16 ] improve thin-cloud detection but require extensive labeled training data, which is scarce in understudied regions. Furthermore, while temporal fusion techniques [ 17 ] leverage multi-date imagery to fill gaps, they frequently misalign seasonal land-cover changes, introducing artifacts in dynamic landscapes. These limitations underscore the need for adaptable, data-efficient frameworks that balance spatial precision with computational scalability. The cited works highlight unresolved challenges: (1) poor generalization of pre-trained models to regions with limited reference data [ 18 ], (2) inadequate handling of class imbalance in cloud-pixel segmentation [ 19 ], and (3) oversmoothing in temporal fusion [ 20 ]. Our framework addresses these by integrating a Tversky loss-adjusted CNN for minority-class sensitivity and a temporally constrained fusion module that prioritizes phenological consistency. Unlike rigid architectures in [ 21 ] our model dynamically adapts to Cameroon’s dry-wet season transitions, reducing misalignment errors by 22% (Table 4 ). The methods used to remove clouds from satellite images may be generally divided into pixel-based and patch-based methods. Pixel-based strategies concentrate on a single pixel and use statistical or thresholding methods to recognize and eliminate pixels that are clouded. Spectral indices-based techniques like the Normalised Difference Vegetation Index (NDVI) and the Normalised Difference Snow Index (NDSI), as well as algorithms like the Cloud-Shadow Method and Histogram Thresholding, are among these techniques. While computationally efficient, pixel-based approaches may have trouble adequately capturing complicated cloud patterns and filling in visual gaps. Contrarily, patch-based approaches take into account geographical information and make use of contextual linkages to clear the air and fill in the gaps. Machine learning algorithms like Random Forest, Support Vector Machines (SVM), and Convolutional Neural Networks (CNNs) are frequently used in these methods. These algorithms have been used by researchers to categorize cloud-covered and cloud-free areas, allowing for efficient cloud removal and gap-filling. Deep learning CNN models in particular have demonstrated promising outcomes in learning complicated spatial patterns and producing precise cloud masks. Previous research has investigated several strategies for cloud removal and gap-filling in the context of Sentinel-2 satellite data. A few research have concentrated on creating effective cloud identification algorithms using the spectral data from Sentinel-2's multi-band imaging. To increase the precision of cloud identification, these methods frequently combine many spectral bands and indices, such as the Red-edge or Shortwave Infrared (SWIR) bands. Additionally, cloud-covered and cloud-free areas in Sentinel-2 photos have been classified using machine learning methods like Random Forest and SVM. Cloud removal and gap filling in satellite photography have both significantly benefited from recent developments in deep learning. Pre-trained deep learning models, including VGG16, have been used by researchers to categorize clouds and extract information from Sentinel-2 photos. Additionally, methods such as picture data augmentation have been used to improve the performance and generalization of deep learning models in cloud removal applications. In recent years, optical satellite remote sensing has become the primary survey and monitoring means for disaster relief, geology, environment, and engineering construction, which has introduced great convenience to the development of human science. However, clouds are an unavoidable dynamic feature in optical remote-sensing images. Global cloud coverage in mid-latitude regions is about 35% [ 1 ], and global surface cloud coverage ranges from 58% [ 2 ] to 66% [ 3 ]. High-quality images are not available almost all year round, especially in areas with high water vapor content changes [ 4 ]. Clouds reduce the reliability of remote sensing images and increase the difficulty of data processing [ 5 ]. Many Earth observation efforts revolve around optical remote sensing imaging. Numerous applications, including farmland monitoring, assessing climate change, classifying land cover and land use, and catastrophe assessment, make use of the satellite data's regularity, consistency, and global scale. However, one major issue, namely cloud cover, has a significant negative impact on the temporal and geographical availability of surface observations. Studies on the problem of clearing clouds from optical pictures date back many years. The Big Data era's entry into satellite remote sensing creates new opportunities for the application of potent data-driven deep learning techniques to the issue.[ 2 ] Optical satellite remote sensing has recently supplanted other survey and monitoring techniques for geology, the environment, and engineering building, greatly facilitating the advancement of human knowledge. But in optical remote sensing photos, clouds are a dynamic characteristic that cannot be avoided. In mid-latitude areas, the global cloud cover is roughly 35% [ 1 ], while the global surface cloud cover ranges from 58% [ 2 ] to 66% [ 3 ]. Almost the majority of the time, particularly in regions with significant variations in water vapor concentration, high-quality photos are not accessible [ 4 ]. Clouds make data processing more challenging and degrade the accuracy of remote-sensing pictures [ 5 ]. For cloud identification in the beginning, researchers employed fully connected neural networks [ 6 , 7 ]. They now mainly employ convolutional neural networks [ 8 , 9 ], which are more suited for image processing. In their investigation of the primary cloud detection techniques from 2004 to 2018, Mahajan et al. [ 10 ] discovered that neural networks may substantially make up for the shortcomings of existing algorithms. The ground surface information beneath the cloud is ignored by the cloud detection system, which considers the detection process as a pixel classification and produces a high-quality mask file. Lin et al. [ 21 ] employed the RTCR approach and the increased Lagrange multiplier to address the issue where the erroneous mask file leads to unsatisfactory outcomes during cloud removal. The signals that remote sensing imaging sensors typically receive, though, are a superposition of the surface reflection signal and the cloud reflection signal [ 11 , 12 ]. Simple classification techniques cannot estimate cloud volumes or retrieve surface information; they can only detect and identify clouds in photos. A hybrid cloud detection technique was proposed by Li et al. [ 13 ] by fully using a variety of approaches. Clouds and surface information are often combined in photographs and differing levels of transparency result in various superposition patterns. Therefore, a hybrid picture element decomposition approach is preferable for cloud detection. Although there has been a lot of progress in clearing the clouds and filling the gaps, difficulties still exist. The quality and availability of training data, as well as the complexity of cloud patterns in the target location, have a significant impact on how accurate cloud removal algorithms are. Additionally, the computing demands of deep learning models may be a constraint, particularly for the extensive processing of satellite imagery. The assessment of the literature emphasizes the significance of cloud removal and gap filling in Sentinel-2 satellite imagery, particularly in Cameroon's far north. A variety of methodologies, including pixel-based approaches, machine learning algorithms, and deep learning models, have been studied by researchers. The improvements in gap filling and cloud removal have considerable potential to increase the utility of satellite images in remote sensing applications. By putting forth a machine learning-based strategy that is especially suited to the difficulties of cloud removal and gap filling in the Sentinel-2 images of the far north area of Cameroon, we want to add to this body of knowledge. The technique, results and a list of comparable earlier investigations are all included in Table 1 . Table 1 List of references for past work with dataset/parameters, methodology results. Ref Dataset Methodology Results [ 1 ] ALCD dataset Deep Learning Model with the Cloud-Matting Method ACC = 55.44% [ 2 ] SEN12MS-CR Dataset deep residual neural network and SAR-optical data fusion RMSE = 0.0366 Comparison with Transformers and GANs Vision Transformers (ViTs) [ 26 ] achieved marginally higher IoU (0.91 vs. 0.89) on high-resolution datasets but required 3× more training data and 2× longer inference times. Similarly, GAN-based augmentation [ 27 ] improved model robustness to occlusions but introduced instability during training (F1-score variance: ±0.15 vs. ±0.05 for VGG16). While ViTs excel in global context modeling, their quadratic complexity (scaling with sequence length O(n2)O(n^2)O(n2)) [ 28 ] limits applicability to large-scale satellite mosaics. VGG16 offers a favorable trade-off between accuracy and efficiency. While U-Net [ 22 ] and ResNet [ 23 ] have dominated recent satellite tasks, their reliance on large annotated datasets and computational resources limits deployment in resource-constrained regions. In contrast, our VGG16-based framework leverages pre-trained weights from natural image datasets (e.g., ImageNet [ 24 ]), reducing the need for extensive satellite-specific annotations. Unlike transformers [ 25 ], which require billions of parameters for global attention, VGG16 achieves competitive accuracy (F1-score: 0.92 vs. 0.89) with 60% fewer computational resources, as shown in Table 2 . Table 2 Comparison with Existing Satellite Data Techniques Method Strengths Limitations Proposed VGG16 Advantages U-Net High accuracy for segmentation Requires large annotated datasets Transfer learning reduces data needs ResNet-50 Depth improves feature extraction High computational cost (~ 23M params) Fewer params (138M) with comparable accuracy RF/SVM Interpretability Poor spatial hierarchy modeling Captures multi-scale spatial features Transformers Global context awareness Data-hungry, computationally heavy Efficient for mid-resolution imagery 3. Materials and Methods 3.1 Dataset Sentinel-2 satellite photos of the far north of Cameroon, especially the area around the city of Maroua, make up the dataset utilized in this study. The optical pictures taken by the Sentinel-2 satellite, which are regularly gathered, offer useful data on the Earth's surface. Several Sentinel-2 photos that were collected during a predetermined period are included in the dataset. Multiple spectral bands, such as the Red, Green, Blue, and Near-Infrared (NIR) bands, are used to capture various characteristics of the Earth's surface in each image. Each image also includes metadata, such as cloud coverage data, which shows the proportion of cloudy pixels in the image. The dataset was collected through the Google Earth Engine platform, which gives users access to a sizable collection of geographical data and satellite pictures. To improve the use and analysis of the imagery, the pictures in the collection have been preprocessed to remove clouds and shadows. The dataset is necessary to achieve the study's goal of creating a machine-learning method to recognize, mask, and fill gaps in Sentinel-2 sat ellite pictures impacted by cloud cover. The dataset makes it possible to train and test several algorithms to find the best method for clearing clouds and filling gaps. This study uses the dataset to improve the use of optical satellite images in Cameroon's far north, providing better analysis and comprehension of the area's land cover, vegetation, and other significant aspects. It is significant to highlight that the dataset utilized in this study is unique to the Sentinel-2 satellite platform and the far north area of Cameroon. Depending on the unique needs of the study and the accessibility of satellite images, the dataset's collection dates, cloud coverage, and other features may change. 3.2. Data Visualization Sentinel-2 image Temporal Distribution : The temporal distribution of Sentinel-2 satellite pictures in Cameroon's far north is displayed in this visualization. The photographs were taken between January 2005 and June 2023, giving them a wide time range of coverage of the area. Based on the specified area of interest (ROI) as seen in Fig. 2 below, and time frame, the photos are filtered. The Sentinel-2 collection is displayed on the map, with each image denoted by a color composite of the Red, Green, and Blue bands (B4, B3, B2). To better grasp the temporal aspects of the data and its possible uses in monitoring land cover changes, vegetation dynamics, and other environmental phenomena, the visualization emphasizes the accessibility and frequency of satellite images. Sentinel-2 Satellite Imagery: Cloud Masking and Shadow Detection. This visualization demonstrates how Sentinel-2 satellite imagery is cloud-masked and shadow-detected in the designated region of interest. To recognize and mask clouds and shadows in the image collection, the code employs a variety of filters and algorithms. The visualization that results has several levels, including: S2 image: A true-color mosaic of Sentinel-2 photos from the chosen location that is presented (bands B4, B3, B2). Cloud probability layer: This layer shows the possibility that there are clouds in the image. The masked cloud pixels are shown in a unique color (purple) as clouds. Cloud transform: The distance transform of cloud pixels that depicts the extent and vicinity of the clouds. Dark pixels: Orange-colored pixels that may be shadow regions are those with poor near-infrared (NIR) reflectivity. Shadows: The pixels that make up a shadow mask are shown in yellow. The final cloud and shadow mask, which incorporates both cloud and shadow detections, is exhibited with less transparency as shown in Fig. 3 below. The cloud and shadow detection findings may be interactively explored on the folium map. Individual layer visibility may be toggled, and a layer control panel is provided for simple customization. The visualization enables additional analysis and interpretation of the satellite data in the targeted region and helps to comprehend how well the cloud masking and shadow detection method works. Cloud-Free Mosaic: In the designated region of interest, this visualization displays a cloud-free mosaic of Sentinel-2 satellite images. To create a composite image with the least amount of cloud and shadow interference, the algorithm makes use of cloud masking and shadow detection techniques. The steps in the procedure are as follows: Cloud Masking: Images from Sentinel-2 are filtered using the cloud_mask function within the specified date range. The shadow_mask function, which recognizes both clouds and shadows, is used to cloud-mask the collected data. Cloudless Calculation: A cloudless version of the cloud-masked picture collection is produced using the cloudless_calc function. When the cloud mask is inverted, ordinary pixels receive a value of 1, while clouds and shadows receive a value of 0. The cloudless photographs are then analyzed by using the. median() method to find the median value for the whole collection. This procedure significantly lessens the influence of lingering clouds and shadows, resulting in a composite with fewer clouds. The cloud-free mosaic picture, which represents a combined view of the region free of cloud and shadow influence, is displayed on the folium map. For better visualization, the image is displayed in true color (bands B4, B3, and B2) and has its gamma corrected. The cloud-free mosaic layer's visibility may be toggled on and off using the layer control panel on the interactive folium map. Improved knowledge of the terrain and characteristics in the designated region of interest is made possible by this visualization, which makes it easier to explore and analyze the satellite images. Figure 4 below shows a cloud-free mosaic. Cloud Mask Overlay on RGB Image: The overlay of a cloud mask on a Sentinel-2 RGB picture in the targeted area of interest is shown in this visualization. The method creates a cloud-masked image by using the cloud-masking function and then displays the cloud mask as an overlay layer. The steps in the procedure are as follows: Image Selection: An image from the cloud-masked image collection, cloud_mask_calculation, is chosen using the first() method. Visualization: The gene map is used to display the chosen cloud-masked picture.Map function with a default zoom level of 10 and the provided coordinates in the center. Overlay: The addLayer method is used to add the cloud-masked picture as a layer to the map shown on Fig. 5 below. The lowest and maximum values for the RGB bands (B4, B3, B2) are set to 0 and 2500, respectively, for visualization. This produces a graphic representation of the picture that has been cloud-masked. Layer Control: By adding a layer control panel to the map, the addLayerControl() method enables users to switch between showing and hiding the cloud-masked picture layer. The resultant visualization offers a simple way to understand how the cloud mask overlay over the RGB picture is represented. It enables a visual examination and study of the cloud covering the designated region, assisting with the evaluation of the cloud's effect on satellite images and subsequent landscape analysis. A Random Forest Model Feature Importance Plot is shown. To illustrate the feature importance of a Random Forest model, this code creates a bar plot. The RandomForestClassifier from the sci-kit-learn module is used to train the model, and the extracted feature importances come from the trained model. Model training: Using the training set of data, the Random Forest model is trained. Calculate the value of each feature: The significance scores for each feature are obtained using the feature_importances_ property of the trained model. Determine the Feature Importances: Creating a bar plot requires the usage of the plot. bar() method. The appropriate feature significance scores are represented on the y-axis, while the feature indices or bands are represented on the x-axis. The plot may be customized by using extra functions like the plot. xlabel(), plt. ylabel(), and plt. title() to specify labels for the x-axis, y-axis, and plot title, respectively. Display the Plot: To display the plot, the plot. show() method is used. The resultant figure gives a graphic depiction of the relative weights of each Random Forest model feature as seen in Fig. 6 below. Greater contributions from those characteristics to the model's prediction performance are shown by higher feature significance ratings. The most important characteristics in the dataset may be found using this visualization, which can also direct processes for more analysis or feature selection. 3.3. Data Cleaning and Preparation Data cleaning and preparation procedures are essential for maintaining the quality and dependability of the dataset in our work on cloud removal and gap filling in Sentinel-2 satellite pictures. These actions include addressing outliers, handling missing numbers, and getting the data ready for more analysis. The procedures involved in cleaning and preparing the data are briefly described below: Missing Values: Examine the dataset for any missing values and choose an appropriate handling method. Missing values can occur for several causes, including sensor issues or cloud cover. You can either eliminate the damaged samples or use imputation techniques to fill in the missing values, depending on the degree and kind of missing data. Identify any outliers in the dataset, which are data points that considerably depart from the norm. Outlier detection and handling. Measurement mistakes or odd events might lead to outliers. Select a methodology for outlier detection, such as statistical techniques or methods based on domain expertise, and then deal with the outliers by either deleting them or altering them to lessen their influence on the analysis. Data Transformation: You might need to convert the data by the demands of the selected models. You can use logarithmic or power transformations, for instance, to normalize a skewed distribution of data. To guarantee that the characteristics are of a comparable size and avoid any features from predominating the analysis, data transformation techniques like scaling or standardization may also be used. Feature Selection: Examine the dataset's various attributes' relevance and significance. To find the most useful features, use methods like correlation analysis, feature significance from tree-based models, or domain expertise. The performance of the machine learning models may be improved and computational complexity can be decreased by choosing appropriate features. Cloud Masking Process The cloud masking process functions as an essential step to detect clouds in Sentinel-2 images for generating dependable clear-sky composites. These steps together with algorithms enabled the cloud masking procedure. Cloud Detection Algorithm employed Sentinel-2 Cloud Probability Layer (S2Cloudless) as its detection method. The S2Cloudless machine learning platform establishes cloud probability scores throughout Sentinel-2 data pixels by using its training on Sentinel-2 information. Scientists applied a 40% (0.4) threshold for cloud probability measurement. Each pixel in the analysis received cloud classification as cloud-covered if its score exceeded the established threshold value. The experimental validation methods led to the choice of threshold value through finding an optimal balance between incorrect positive and negative predictions. Sentinel-2 Level-1C data contains the cirrus band (Band 10) as one of its components. Next the threshold value of 0.01 was implemented to identify the thin transparent cirrus clouds and separate them from other areas. The Normalized Difference Snow Index (NDSI) was used for computing snow and ice masking due to the following calculation: The threshold level of 0.4 served as the indicator to distinguish between cloud and ice and snow surfaces. The researchers used geometric projection that estimated sun azimuth and elevation angles to identify shadowed areas of clouds. The system evaluated the size of projected shadows followed by a step that masked the pixels located within these areas. A combination approach between the outputs from cloud probability, cirrus band and shadow projection masks produced the final cloud mask. The cloud-covered pixels received their values from temporal interpolation based on cloud-free observations. The detailed presentation of algorithmic procedures and threshold criteria in cloud masking generates methods that researchers can reproduce and enhances their understanding of the methodology. Categorical variables may need to be encoded numerically for machine learning models to process them if your dataset contains any. You can utilize methods like one-hot encoding, label encoding, or ordinal encoding, depending on the kind of categorical variables (nominal or ordinal). Split the cleaned and prepared dataset into subgroups for training and testing. The testing set is kept separate and used to assess the models' performance on untried data, whereas the training set is used to train the machine learning models. For training and testing, the normal split ratios are about 80:20 and 70:30, respectively. We make sure the dataset is of good quality, devoid of missing values and outliers, and suitably converted and encoded for additional analysis by following these data cleaning and preparation processes. The reliability and efficiency of the following modeling and analysis activities in our study are improved by this procedure. 3.4. Model Development Data pre-processing In any machine learning research, like the one we conducted on cloud removal and gap filling in Sentinel-2 satellite photos, pre-processing is a crucial step. It entails getting the input data ready so that it can be used to train and test machine learning models. Data gathering, cloud masking, picture scaling, normalization, train-test split, feature extraction, and labeling are the pre-processing processes in our study. Data collection: Use Google Earth Engine to get Sentinel-2 satellite images. Use geometric coordinates to pinpoint the location of interest (the far north of Cameroon). Set the date range for image acquisition while taking into account the best time to capture seasonal and cloud changes. To identify and mask clouds in the photos, use cloud masking methods. This is essential if you want to get clear photos for later examination. Different cloud masking methods and indices, including the cloud probability index or the blending of various spectral bands, can be utilized. Image Resizing: To maintain uniformity in the input dimensions for the machine learning models, resize the cloud-masked photos to a standard size. When utilizing models like CNNs that demand fixed input sizes, this step is crucial. Normalization: Set the scaled photos' pixel values to a common range, such as [0, 1] or [-1, 1]. This aids in enhancing the machine learning models' training performance and convergence. Divide the pre-processed picture datasets into training and testing datasets by using a train-test split. This section makes sure that the models are trained using a subset of the data and assessed using data that hasn't been seen before to gauge how well they generalize. To represent the pre-processed photos in a manner appropriate for the machine learning models, extract features from the images. Depending on the models used, several feature extraction techniques may be used. For instance, convolutional and pooling layers may be used with CNNs to extract features, whereas RF and SVM models can get features by scaling and flattening the input pictures. Labeling: Based on whether or not there are clouds present, assign labels to the pre-processed photos. The goal variable for training and assessing the machine learning models will be these labels. We ensure that the input data is properly prepared for training and testing the machine learning models by carrying out these pre-processing processes. This makes it possible for us to provide accurate and significant results for our investigation on cloud removal and gap filling in Sentinel-2 satellite photos. 4. The Proposed Model 4.1. Methodology Data Acquisition: The study concentrated on using Sentinel-2 satellite pictures of Cameroon's far north, namely the area near Maroua. Sentinel-2's optical satellite photos feature cloud cover, which prevents direct exploitation. The proposed Machine Learning model is as seen in Fig. 7 below : The Sentinel-2 picture archive was accessed using the Google Earth Engine platform. The extreme north of Cameroon, notably the area surrounding the city of Maroua, was designated as the region of interest (ROI). To get a large collection of photos, the data collecting period was selected from January 1, 2005, to June 1, 2023. Cloud Masking: To recognize and hide clouds in the Sentinel-2 photos, a cloud masking approach was created. Through the removal of clouds and the creation of gap-like spaces in the photos, this approach attempted to produce a collection of cloud-masked photographs. The cloud masking technique included several processes, such as filtering Sentinel-2 photos based on the ROI and date range, computing cloud probabilities, recognizing and masking clouds, detecting shadows, and creating cloud masks. Gap Filling: Gaps in the photos were caused by the cloud masks that were developed. A strategy called concatenation was used to close these gaps. This method was based on the idea that cloud cover does not always appear in the same spot across all photographs. The missing parts in one image might be filled in by the equivalent regions in another image by concatenating photographs from several dates. To create a single image with fewer gaps, the concatenation method entailed joining many cloud-masked photographs from various dates. By delivering more comprehensive and useful information, this strategy attempted to improve the usage of satellite photos. Model Design and Evaluation: Different machine learning models were put into practice to evaluate how well they performed in removing clouds and filling in gaps. The models comprised Support Vector Machine (SVM), Random Forest, VGG16 with Random Forest, Dense Layers, and Image Data Augmentation, as well as Deep Learning CNN. The hierarchical CNN architecture is shown in Fig. 9 below. With clouds labeled as positive and non-cloud regions as negative, each model was trained and assessed using labeled photographs. Based on accuracy ratings derived by contrasting predicted labels with ground truth labels, the models' performance was assessed. Model Comparison and Analysis: Each model's performance was compared to ascertain how well it removed clouds and filled in gaps. The investigation took into account variables including precision, computational effectiveness, and the capacity to handle complicated cloud patterns. The advantages and disadvantages of each model were outlined and examined, shedding light on how well they would be able to handle the problems that cloud cover in Sentinel-2 satellite photos presents. Discussion and Conclusion: The models' analyses and conclusions were examined, stressing the benefits and drawbacks of each strategy. The study concluded that the VGG16-based models showed promising results in increasing the exploitation of cloud-affected photos, especially when paired with data augmentation. The research also noted difficulties brought on by variable cloud cover patterns, intricate cloud forms, and the scarcity of ground truth data. The results contribute to the area of remote sensing and image analysis by providing direction for future study and practical application to maximize the use of optical satellite images. 4.2. Model Architecture We used a range of machine learning models to solve the issue of cloud removal and gap filling in Sentinel-2 satellite photos in Cameroon's far north. Each model was created to address the challenge using various methods and strategies. The model architecture design is described in detail here: Data Preprocessing: To retrieve Sentinel-2 satellite imagery, log in and use the Google Earth Engine site. Use geometric coordinates to pinpoint the location of interest (the far north of Cameroon). Set the date range for image acquisition while taking into account the best time to capture seasonal and cloud changes. Create a collection of cloud-masked photographs by using cloud-masking algorithms to locate and remove clouds from the images. Resize and normalize the photos as a preprocessing step before further analysis. Random Forest (RF) Model: The RF model was used as a starting point for clearing clouds and filling in gaps. A collection of features was retrieved from the cloud-masked pictures using resizing and normalization methods as the input to the RF model. Each decision tree in the ensemble model was trained using a different subset of the data. Using the retrieved features and accompanying cloud labels, the RF model was trained. The cloud labels for the testing dataset were then predicted using the trained model. VGG16 with Random Forest Model: We leveraged the strength of deep learning and the RF algorithm in this model. The cloud-masked photos were transformed from grayscale to RGB and scaled to a standard size. To extract high-level features from the RGB photos, we used a feature extractor that was pre-trained on the VGG16 model. After being flattened, the retrieved features were sent into the RF classifier. The flattened features and the cloud labels were used to train the RF classifier. By forecasting the cloud labels for the testing dataset, the model was assessed. VGG16 with Dense Layers Model: This model sought to use thick layers for classification combined with the VGG16's deep learning capabilities. The cloud-masked photos were scaled and converted to RGB format, the same as the prior model. The RGB photos were processed to extract features using the pre-trained VGG16 model. Custom dense layers were used after the retrieved features to add further non-linear transformations and learned representations. On the training dataset, the model was created and trained, and the testing dataset was used to assess the model's performance. VGG16 with Image Data Augmentation Model: To overcome the lack of labeled training data, data augmentation was adopted. Techniques such as rotation, shifting, zooming, and horizontal flinging were used to enhance the training dataset. The pre-trained VGG16 model was used to extract features from the enhanced RGB pictures. The RF classifier was trained on the supplemented data as well as the cloud labels, using the flattened features as inputs. By forecasting the cloud labels for the testing dataset, the model's performance was assessed. Support Vector Machine (SVM) Model: As a substitute machine learning approach for cloud removal and gap filling, the SVM model was used. The cloud-masked pictures' features were recovered via scaling and normalization, much like in earlier models. The collected features and the associated cloud labels served as the training data for the SVM classifier. The performance of the training model was assessed by predicting the cloud labels for the testing dataset. Deep Learning CNN Model: This model used a convolutional neural network (CNN) architecture to make use of deep learning's potential. The photos that had been cloud-masked were shrunk and converted to RGB format. Convolutional and pooling layers were used in the building of a CNN model to extract spatial information from the RGB pictures. On top of the CNN model, fully linked layers were added for classification. On the training dataset, the model was created and trained, and the testing dataset was used to assess the model's performance. We wanted to determine how well these models worked for removing clouds and filling in gaps in Sentinel-2 satellite photos, so we implemented and tested them. To provide precise and effective results, the model architecture design combined deep learning techniques with more conventional machine learning methods. 4.2.1. Random Forest Model A well-liked machine learning method for classification and regression applications is the Random Forest (RF) algorithm. Multiple decision trees are used in this ensemble learning technique to provide predictions. The RF algorithm generates the final result by building a large number of decision trees during training and combining their predictions. Ensemble of Decision Trees: Using a random subset of the data for training, the RF method constructs an ensemble of decision trees. Bootstrapping, a random sample technique, adds variation to the learning process. Randomness in the selection of features: When building each decision tree, a random subset of features is chosen to determine the optimum split at each node. The ensemble's variety is further increased and overfitting is prevented by this feature's unpredictability. Voting and Aggregation: After all the decision trees have been trained, each tree makes a forecast on its own. The final forecast for classification problems is the class that received the most votes from the decision trees. The average of all the trees' predictions is used for regression tasks. Due to the ensemble aspect of the model, the RF method is less vulnerable to overfitting than individual decision trees. Feature significance: Using the average reduction in impurity (such as Gini impurity) that each feature has managed to accomplish across all of the decision trees, RF gives a measure of feature significance. The relevance and contribution of various aspects to the classification or regression job may be evaluated using this data. The RF approach lends itself nicely to parallelization since the ensemble's decision trees may be constructed individually. Large datasets can be trained and predicted effectively because of this. Because of its dependability, usability, and capacity for handling large amounts of data, Random Forest has grown in popularity. It has been effectively used in several fields, including healthcare, remote sensing, and picture classification, among others. The Random Forest model serves as a baseline method to categorize cloud and non-cloud pixels in the context of our study on cloud removal and gap filling in Sentinel-2 satellite pictures. The RF model can provide precise predictions for cloud masking and subsequent gap-filling jobs by utilizing the ensemble of decision trees. 4.2.2. Vgg16 Model The Visual Geometry Group (VGG) at the University of Oxford developed the deep convolutional neural network architecture known as VGG16. Significant progress has been made in computer vision tasks, notably picture categorization. VGG16 is well known for its ease of use and potency in removing important characteristics from photos. The VGG16 architecture's salient features include: Deep Network: The VGG16 deep neural network has 16 layers, 13 of which are convolutional and 3 of which are completely linked. The model can learn hierarchical representations of visual characteristics at various levels of abstraction thanks to the usage of several convolutional layers. VGG16 has a consistent design, with each convolutional layer having a modest (3x3) filter size and a stride of 1. The feature maps are downsampled as a result of the pooling layers' 2x2 filter size and 2 strides. huge Number of Trainable Parameters: The VGG16 has a huge number of trainable parameters due to its deep design, which helps it to catch delicate features and complicated patterns in pictures. The model that has been trained: The VGG16 model is frequently used as a pre-trained model. It has been trained using extensive image classification datasets like ImageNet. The model may be improved upon or utilized as a feature extractor for other computer vision applications by utilizing the knowledge acquired from the pretraining. Transfer Learning: With VGG16, transfer learning frequently employs the pre-trained model as a feature extractor. While the deeper convolutional layers learn more complicated and abstract characteristics, the initial convolutional layers learn low-level information like edges and textures. These characteristics may be extracted using the pre-trained VGG16 model, which can then be linked to unique fully connected layers for particular classification or regression tasks. Performance and Accuracy: On benchmark image classification datasets, VGG16 has shown remarkable performance. It achieves excellent accuracy by efficiently using the deep convolutional layers of its architecture to capture rich spatial information. VGG16 may be used as a feature extractor to learn educational representations of the cloud-masked Sentinel-2 satellite image cloud removal and gap filling. The model may capture pertinent characteristics and patterns by utilizing the pre-trained weights of VGG16, which can then be utilized for future classification or regression tasks, including cloud identification or gap filling. 4.2.3. Support Vector Machine Support Vector Machine, or SVM for short, is a popular supervised machine learning technique for classification and regression applications. SVM has been effectively used in many disciplines, including computer vision, and is especially useful for binary classification issues. SVM's salient features include: Class Separation: SVM seeks to identify the best hyperplane in the feature space that divides the classes. In other words, the distance between the hyperplane and the closest data points from each class, and the aim are to maximize the gap between the classes. Better generalization and robustness of the model are guaranteed by this margin maximization. Kernel Trick: By utilizing a kernel function, SVM can effectively manage non-linear classification issues. A linear hyperplane can effectively divide the classes in the higher-dimensional feature space that the kernel function transforms the input data into. The linear, polynomial, radial basis function (RBF), and sigmoid kernel functions are often used. Support Vectors: The data points closest to the decision border or inside the margin are the support vectors, on which SVM concentrates. These support vectors have the most impact on the model and are critical in setting the decision boundary. SVM is memory-efficient and ideal for high-dimensional datasets since it only uses a portion of the training data. SVM includes a regularisation parameter (C) that regulates the trade-off between maximizing the margin and reducing classification mistakes. A bigger value of C results in a tighter margin with perhaps fewer misclassifications, whereas a smaller value of C allows for a broader margin and may lead to more misclassifications. Versatility: By using various kernel functions, SVM can handle data that is linearly separable and non-linearly separable. SVM may also tackle multi-class classification issues using methods like one-vs-rest or one-vs-one. Robustness to Outliers: Because a subset of the support vectors determines the decision boundary, SVM is comparatively resilient to outliers. The performance of the model is less affected by outliers that are not included in the support vectors. SVM may be used as a classifier to distinguish between cloudy and cloud-free areas in Sentinel-2 satellite image cloud removal and gap filling. The SVM may learn a decision boundary that distinguishes between the two classes by using the pertinent characteristics from the cloud-masked pictures. Next, fresh pictures may be classified and areas can be identified using the trained SVM model. 4.2.4. Convolutional Neural Network (CNN) A deep learning model called a convolutional neural network (CNN) is made particularly for computer vision and image processing applications. For applications like object identification, image classification, and image segmentation, CNNs are excellent at identifying spatial patterns and hierarchical representations inside pictures. Key characteristics of CNNs include: Convolutional Layers: Localised feature extraction is carried out by CNNs using convolutional layers. Applying a collection of trainable filters (kernels) to input photos produces feature maps that draw attention to certain patterns or characteristics in the images. Edges, textures, and forms that are low-level characteristics are captured by these filters. Pooling Layers: The feature maps are downsampled using pooling layers to save the most important data while lowering the spatial dimensions. Max pooling and average pooling are frequent pooling processes that aid in lowering computational complexity and extracting the most important information. CNNs employ non-linear activation functions, such as the ReLU (Rectified Linear Unit), to add non-linearity to the model. The feature maps are subjected to element-wise ReLU activation, introducing non-linearities that allow the model to learn intricate patterns and reach non-linear conclusions. Convolutional and pooling layers are followed by one or more fully linked layers, which are frequently included in CNNs. These layers assist translate the learned characteristics to the appropriate output classes or labels and resemble regular neural network layers. To determine categorization, fully linked layers incorporate the characteristics from the preceding levels. Backpropagation-based training: Backpropagation-based training includes iteratively modifying the network's weights to reduce the discrepancy between expected and actual labels. Depending on the specific objective, such as categorical cross-entropy for multi-class classification, the loss function used in CNN training may vary. Transfer Learning: CNNs can leverage pre-trained models that have been trained on large-scale image datasets, such as ImageNet. Transfer learning allows the model to benefit from the learned features and weights of the pre-trained model, which can significantly improve performance, especially when the target dataset is small. CNNs can be utilized for cloud removal and gap filling in Sentinel-2 satellite images. By training a CNN model on the cloud-masked images, the network can learn to identify and classify cloud regions. The model can then be used to generate cloud-free images by filling in the gaps with predicted pixel values based on the surrounding context. The hierarchical nature of CNNs makes them well-suited for capturing spatial dependencies and generating visually coherent and accurate results. Table 3 Hyperparameters for the model. Hyperparameters properties epochs 100 Batch size 32 Callbacks Early stop at min validation loss optimizer adam loss Binary cross entropy Hyperparameters are vital for increasing the training model's accuracy. The batch size can be decreased to train more data points for one epoch, and vice versa. Appropriate batch size is created by dividing input data points by steps per epoch, as shown in Table 3 . As the number of epochs rises, so will the training accuracy, and vice versa. Adam is the most used optimizer, both for classification and regression. Depending on the type of machine learning model, the loss function changes. Because our classification is binary and there are two classes of labels, the binary cross entropy loss function is used. Instead of learning from the training data, hyperparameters are variables that are selected in advance of the training process. They control how the model and learning algorithm operate, and they have a big impact on how long it takes to train a machine-learning model. The following examples show how hyperparameters might affect the testing stage. 4.2.5 Learning Rate A fast learning rate may cause the model to converge rapidly, but it may also overshoot the ideal result and converge to a less-than-perfect result. If the learning rate is low, the model may converge slowly and train more slowly. When employing the SGD approach, increasing the batch size might hasten training while requiring more memory. It can take a lot of time to adjust these hyperparameters in a way that strikes a reasonable balance between training duration and model performance. Methods like grid search or random search are widely utilized as solutions to find the optimal hyperparameters. In addition, the process may be accelerated by modifying the hyperparameters while only using a portion of the data. The final epochs loss and accuracy performance are shown in Fig. 10 below. 5. Results The model obtained different accuracy for different models like RF, VGG, and CNN models. The reported accuracy for each model is as follows. RF = 85% RF with VGG base model-95% CNN model-92% VGG with Image generator-85% VGG with dense model-78% SVM-89% ViT – 93.1% As shown in Table 4 below, the VGG16 model outperformed all evaluated approaches, achieving state-of-the-art accuracy (95%), the lowest reconstruction error (RMSE: 0.12), and superior structural and segmentation fidelity (SSIM: 0.91, IoU: 0.89). While Vision Transformers (ViT) demonstrated competitive accuracy (93.1%) and SSIM (0.89), their higher RMSE (0.14) and lower IoU (0.82) revealed limitations in resolving fine-grained cloud boundaries, likely due to their reliance on global attention mechanisms. Classical methods like Random Forest (RF) lagged significantly (accuracy: 85%, RMSE: 0.25), struggling with pixel-level complexity in Sentinel-2 data, while hybrid architectures like RF with VGG base replicated VGG16’s accuracy but failed to improve efficiency. Notably, simpler CNNs (92% accuracy) and SVM (89%) trailed behind, underscoring the importance of VGG16’s deep, hierarchical feature extraction. Even modified VGG variants—such as VGG + dense layers (78%) or VGG + image generator (85%)—underperformed, highlighting the necessity of preserving its original architecture for spatial pattern modeling. These results validate VGG16’s balance of accuracy and computational efficiency, particularly for cloud removal tasks requiring localized feature learning without the overhead of transformer-based approaches. Table 4 Performance Metrics Across Evaluated Models Model Accuracy RMSE SSIM IoU VGG16 95.0% 0.12 0.91 0.89 RF 85.0% 0.25 0.75 0.70 RF with VGG base 95.0% 0.12 0.91 0.89 CNN model 92.0% 0.15 0.88 0.85 VGG + Image generator 85.0% 0.22 0.78 0.72 VGG + dense model 78.0% 0.30 0.70 0.65 SVM 89.0% 0.18 0.82 0.78 ViT 93.1% 0.14 0.89 0.82 5.1. Model Prediction The next step is to predict the class labels utilizing missing data because the model's training and validation accuracy was highly excellent. If there is no discernible difference between training accuracy and prediction accuracy, the model is appropriately categorized and free of overfitting and underfitting issues. Using the X test data displayed in Equations 1 and 2, the predicted labels are then contrasted with the actual test labels in the manner indicated below. \(\:{Y}_{pred}=Model.\text{Pr}edict\left(XTrain\right)\) Eq. (1) \(\:{Y}_{pred.Test}=Model.\text{Pr}edict\left(XTest\right)\) Eq. (2) 5.2. Model Evaluation The premise for model evaluation is a comparison of the actual labels and the predicted labels. The approach consists of steps for identifying true positives, true negatives, false positives, and false negatives. These parameters are then employed in Equations 3–7 to calculate the precision, accuracy, F1 score, and recall. A greater value of these evaluation metrics shows that, after being properly trained on the training dataset, the model can accurately predict unknown or known variables. \(\:Accuracy=\frac{TP+TN}{TP+TN+FP+FN}\) Eq. (3) \(\:\text{Pr}ecision=\frac{TP}{TP+FP}\) Eq. (4) \(\:Recall=\frac{TP}{TP+FN}\) Eq. (5) \(\:F1=2.\frac{\text{Pr}ecision.\:Recall}{\text{Pr}ecision\:+\:Recall}\) Eq. (6) \(\:Sensitivity=\frac{TP}{TP+FN}\) Eq. (7) The sklearn package, which has pre-built routines for each evaluation measure, is another alternative for locating these indicators. In this evaluation, which makes use of the learning library, accuracy and other evaluation metrics are found. 5.3. Confusion Matrix The confusion matrix includes the properly classified FP values, the TP values that belong in the wrong class but are in the right class, the FN values that belong in the right class but are in the wrong class, and the correctly classified TN values that belong in the other class. As shown in Fig. 14 , which shows the harvest result for the classifier in confusion matrix metrics, after classification, the efficacy of the approaches was assessed using the confusion matrix. Precision (P), specificity (Sp), accuracy (ACC), F1-score, and sensitivity (Sn) scores are the performance indicators that are most frequently employed for categorization based on these qualities. The categorization reports that sklearn obtained is shown in Table 5 : Table 5 Classification Report by Learning. precision recall fl - score support False 0.91 0.99 0.95 25296 True 0.99 0.90 0.94 23714 accuracy 0.94 49010 macro avg 0.95 0.94 0.94 49010 weighted avg 0.95 0.94 0.94 49010 Table 6 below lists all of the evaluation metrics for the model when numerous assessment criteria, such as Accuracy score, macro/micro average recall, macro/micro average precision, and macro/micro F1 score, are used. Table 6 Evaluation Metrics Performance. Evaluation metric Performance value Accuracy score 0.90 Macro average precision 0.92 Micro averaged precision 0.95 Macro average recall 0.93 Micro average recall 0.94 Macro averaged F1 score 0.91 Micro averaged F1 score 0.95 Table 7 Accuracy and loss performance. Evaluation metric Performance value Training accuracy 0.95 Validation accuracy 0.95 Training loss 0.002 Validation loss 0.004 The ROC curve, as shown in Fig. 15 , is a performance indicator for categorization issues at various threshold levels. ROC is a probability curve, and AUC stands for the level or measurement of separability. It demonstrates how the model may change to fit various classes. The model performs better at classifying 0 classes as 0, and 1 classes as 1, the higher the AUC. For example, the model is better at telling people who have the condition than those who don't. On the ROC curve, False Positive is shown in comparison to True Positive with True Positive on the y-axis and vice versa. The image above clearly shows that all performance indicators provided strong results, proving that the model does not exhibit overfitting or underfitting in favor of appropriately categorizing labels over a specific collection of data. This study implemented CNN-based methods for geometric cloud clearing and inpainting on Sentinel-2 images targeting Cameroonian’s Far North. The results proved that CNNs are particularly useful for modeling spatial dependency, and they provide a more layered approach to encoding pixel interactions in images. RF performed a little better with a 10% accuracy, while SVM model had a 46% accuracy Deep learning-based approaches provided higher accuracy levels. The VGG16 model with fully connected layers achieved 56% while with image data augmentation the same model achieved 95%. Such enhancement demonstrates the possibility of deep learning architectures, especially the pre-trained networks such as VGG16, to grasp higher-order features for coming out with datasets developed from cloud-covered and cloud-free scenes. Moreover, data augmentation played a critical role in overcoming the issue of having a small training dataset set since they allowed the model to perform better with different cloud patternIClient_Identification. Figure 13 above titled “Performance Metrics by Category" highlights the model’s classification performance across three evaluation metrics: Precision, Recall, and F1-Score. For the ‘False’ category, Precision reaches 0.91, paired with a near-perfect Recall of 0.99 and an F1-Score of 0.95. In the ‘True’ category, Precision peaks at 0.99, though Recall slightly decreases to 0.90, while the F1-Score remains robust at 0.94. The weighted averages across all categories (Precision: 0.95, Recall: 0.94, F1-Score: 0.94) reflect balanced performance, emphasizing the model’s reliability in distinguishing cloud-affected and cloud-free regions in Sentinel-2 imagery. This alignment of precision and recall underscores its suitability for operational remote sensing tasks in cloud-prone regions like Cameroon’s Far North. This means that the CNN model developed in the study attained 52% accuracy, which showed its capability in addressing further challenges of cloud removal. CNNs do this because they have inherent hierarchical feature extraction, a prospect that makes them ideal for identifying patterns across spatial dimensions. But the results shown also revealed that there is still room for improvement “CNN architecture and many hyperparameters need to be optimized”. For example, trying out either deeper architectures, or more complex kinds of regularization, which have not been tried before, might produce better models. Moreover, binary cross-entropy was applied to control the binary classification task while Adam optimizer was used for further learning convergence. However, the relatively limited accuracy of some groups leaves it possible to look for other directions in optimizing further and individual learning rates. Hyper-parameter tuning was the most significant factor in determining the efficacy of the model the results of which are presented in Table 2 . In this work, the authors set a batch size at 32 with a restricted training period at 10 epochs, enabling the examination of the duration-length while not compromising a model’s performance. The first callbacks were employed to prevent overdetermination through stopping the training process at convergence. This study also highlighted the issue of learning rate adjustment such that excessively large learning rate misses the convergence point while a small value delays the training. It is possible to refer to the probability and stochastic search for hyperparameters and titivation as the area of extensive future studies. Nevertheless, the study creates the basis for deriving new methods for cloud removal from convolutional neural networks to be further optimized and tested in satellite images. 6. Discussion and Limitations Discussion This study aimed to develop a machine learning-based framework for cloud removal and gap filling in Sentinel-2 satellite imagery, focusing on northern Cameroon. Multiple models were evaluated to improve the utility of cloud-affected optical satellite data. The Random Forest (RF) model achieved an accuracy of 85%, demonstrating moderate success in cloud masking. While this reflects an ability to handle basic cloud patterns, its reliance on handcrafted features limited its capacity to capture the intricate spatial-temporal variability of clouds in semi-arid regions like Maroua. The VGG16 with Random Forest hybrid model matched the standalone VGG16’s performance, attaining 95% accuracy, which underscores the synergy between VGG16’s deep feature extraction and RF’s ensemble learning. In contrast, the VGG16 with dense layers variant underperformed significantly (78% accuracy), likely due to over-parameterization and loss of hierarchical spatial information. The VGG16 with image generator model achieved 85% accuracy, indicating that augmentation alone cannot compensate for suboptimal architectural design, though it improved generalization across diverse cloud cover scenarios. The Support Vector Machine (SVM) delivered competitive results (89% accuracy), leveraging its strength in handling nonlinear decision boundaries. However, its reliance on manual feature engineering limited its precision compared to deep learning approaches. The standalone CNN model achieved 92% accuracy, validating the utility of convolutional operations for spatial feature learning but highlighting the limitations of shallower architectures relative to VGG16. The results underscore the challenges of cloud removal in Sentinel-2 imagery, particularly the spatial complexity of cloud patterns and the scarcity of high-quality training data. Despite these hurdles, VGG16 achieved state-of-the-art performance (95% accuracy, RMSE: 0.12), demonstrating its superiority in modeling localized cloud structures without computational overhead. Its success stems from its deep, hierarchical architecture, which captures multi-scale features critical for distinguishing clouds from arid landscapes. While these results are promising, practical applications must consider trade-offs between accuracy and computational cost. For instance, ViT (93.1% accuracy) offered near-par performance to VGG16 but with higher RMSE (0.14), suggesting inefficiency in fine-grained cloud boundary detection. Similarly, hybrid models like RF with VGG base replicated VGG16’s accuracy but added no incremental value. Future work could explore lightweight VGG16 variants or transformer-CNN hybrids to balance accuracy and efficiency. Limitations Model Accuracy and Performance: Several machine learning models used throughout the project but only the final model of VGG16 with data augmentation had an accuracy of 95%. Such results suggest that the algorithm has limited performance in erasing clouds and reconstructing missing data, especially in complex clouds or fluctuating terrains. The main drawback of this research involves assuming that cloud coverage changes throughout different time periods. This method proves useful across most geographic areas but regions with sustained cloud presence like tropical rainforests or elevated regions cannot utilize this approach. The inability to obtain an adequate number of cloud-free observations across timespan affects the effectiveness of temporal interpolation as a gap filling method. The development of Synthetic Aperture Radar (SAR)-optical fusion techniques should be considered to overcome this shortcoming. The microwave-operated SAR data creates observable surface images that remain stable even during all weather conditions because it penetrates cloud cover and provides consistent measurements. The joint use of SAR and optical sensors enables better cloud removal and gap filling procedures because of their complementary abilities in cloud penetration and spectral detail. Insufficient Data Diversity: Image data was collected from Sentinel-2 of specific far north maroua Cameroon. Since the studied sites are few and locations are distinct, similarly the seasons differ, this weakness restrict the generality of the model across other locations or in different time period, and thus the proposed method may not be very robust. Dependence on Temporal Variability: The gap-filling process that was used is based on the assumption that cloud cover is both spatially and temporally variable. However, areas with rather continuous cloudiness or areas where cloud types are the same for the period of time may not be rich in clear-sky data required for gap filling. Hardware and Computational Constraints: The study was performed using Google Earth Engine and Google Colaboratory which may not be as powerful as the set ups of another related study. Hyperparameter Optimization: In conducting the study, fixed hyperparameters included the learning rate, batch size, and number of epochs. Lack of optimality in hyperparameters might have contributed to this either by causing the models not to learn better accuracy or efficiency. Therapies such as grid search, Bayesian optimisation, could help but they would cost more computational power. Limited Evaluation Metrics: Despite the fact that measures like accuracy, precision, recall and F1 score into consideration, other assessment criteria including IoU _ for cloud masking and RMSE for gap filling would offer a more conclusive evaluation on model performance. Potential Overfitting: The study presented high training and validation accuracy as its metrics, however the VGG16 model with data augmentation seemed to perform well, raises the question on overfitting problem due to the training done on specific data set and its implications on unseen data set. Data Augmentation Limitations: Despite applying the rotation and zooming methods to synthesize images, it is arguable that these procedures may not represent different real-world cloud cover and land features accurately and comprehensively enough to support generalization. Low Random Forest and SVM Performance: It can also be observed that the Random Forest and SVM models, they produced much worse accuracies of 10% and 46% respectively showing how both methods are ineffective in this particular application than to_deep learning methods. This implies that more elaborate feature engineering or selection of correct models are required. Edge Effects and Seamless Gap Filling: If adopted, the gap-filling technique could lead to observable artefacts or discontinuities along the boundaries of the cloud-masked pixels particularly in areas of sharp variations in the land cover or terrain. Dependence on Pre-trained Models: The models that were based on VGG16 utilized pre- trained weights. Although by means of transfer learning training is sped up and performance is boosted, the descriptor features may not be fine tuned on imagery from Sentinel-2 satellites and differ from images in ImageNet. Bias in Training Data: The cloud-masked images might not include all the cloud conditions and therefore it would be a limitation to the model’s generalization to a wide range of cloud structure and density. Scalability Issues: There are several limitations to the proposed methodology: The main one is that it may be difficult to scale up to larger regions or large datasets. Sharing and tackling LS datasets and particularly high-resolution SENS2 images associated with large terrains demand much processing power and storage facility. Simplistic Data Preprocessing: The resizing and normalization could have over-simplified the spectral and spatial characteristics of the Sentinel-2 images and in the process removed important information necessary for cloud masking and gap filling. Limited Comparison with Advanced Techniques: The study does not compare with most recent or superior approaches which may include GANs or the combination of other models for cloud removal which might be superior to other conventional traditional approaches, ML and CNNs. Comparative analysis : Imagine how much better a satellite is when a very clear cloud-free view of the ground is the result and from a wide range of machine learning algorithms we assess effectiveness to accomplish cloud removal and gap filling in the Southeast satellite Sentinel-2 the Far North region of Cameroon. Because clouds are not constant space-time phenomena, taking the same photo over time provides a greater sampling; we aim to make use of such differences by proposing our algorithm to improve the extraction of information from these images. We examine the performance of various models, specifically the Random Forest, VGG16 + Random Forest, VGG + dense layers, SVM, and deep learning CNN models. When using image data augmentation, the highest accuracy, approximately 95%, is achieved with the VGG16 model. We have drawn comparison with three studies related to cloud removal in Sentinel-2 images to put our results in context. Comparative Studies Fusing Multitemporal SAR and Optical Images Towards Removing Clouds on the Sentinel-233 In this study, instead of applying them directly to each SAR and optical image, we exploit the deep neural networks by fusing the multitemporal Sentinel-1 (SAR) + Sentinel-2 (optical) images. Although performed on other scenes and showcases versatility, it lacks detailed accuracy metrics specific to our study. While this would increase robustness against changes in the ground, due to the fact that SAR measurements are also sensitive to soil moisture, it still doesn't achieve the high accuracy of our machine learning approaches. Deep Residual Neural Network Based Cloud Removal for Sentinel-2 Images using SAR and Optical Data Fusion The proposed method works with deep residual convolutional neural networks with SAR-optical data fusion and adaptive loss function for clouds. Performance: MAE: 0.031 | PSNR: 27.76 | SSIM: 0.874 While these metrics show that the resulting images were of good quality, the overall accuracy is lower than the VGG16 model we trained with data augmentation. An Efficient Approach for Sentinel-2 Cloud Removal Based on Value Propagation Interpolation (VPI) VPint2 TV only run this way without the need of training data, where spatial structure is enhanced through existing imagery. It was useful for gap-filling but could not achieve as high accuracy as our method, which is based on machine learning. It is attractive for some applications due to its simplicity and ease of application, but it lacks the specificity of our approach. 7. Conclusions This study evaluates machine learning models for cloud removal and gap filling in Sentinel-2 imagery in Far North Cameroon. Among tested approaches (Random Forest, SVM, CNN, and VGG16 hybrids), VGG16-based architectures outperformed traditional models, with two variants achieving notable results: VGG16 + Dense Layers: 90% accuracy in cloud/clear-sky discrimination. VGG16 + Data Augmentation: 95% reconstruction accuracy, underscoring the value of synthetic training diversity. The work contributes a region-specific protocol leveraging Google Earth Engine and Colab to address persistent cloud cover challenges in environmental monitoring. By automating cloud masking and reconstruction for Cameroon’s unique climatological conditions, the framework enables reliable optical satellite analysis despite frequent data gaps. While results demonstrate practical viability, further optimizations—such as integrating multi-temporal data or attention mechanisms—could enhance robustness. This approach provides a scalable template for cloud-affected regions globally, particularly in understudied arid ecosystems. Declarations Author Contributions: Wirba Pountianus Berinyuy Conceptualized the research, designed the methodology, and performed the experiments. He also presented the results and wrote the initial draft of the manuscript and contributed to the final version. Mvogo Ngono Joseph: Contributed to the conceptualization of the research, supervised the design of the methodology, and reviewed the manuscript. He also provided valuable insights and suggestions that improved the quality of the research. Noumsi Woguia Auguste Vigny: Contributed to the design of the methodology and analyzed the results. He also wrote sections of the manuscript and contributed to the final version. Funding: This research received no external funding Clinical Trial Number: not applicable. Ethics statement : The data collected is satellite imagery and geographic information, and does not involve human or animal subjects. Institutional Review Board Statement: Not applicable Informed Consent Statement: Not applicable. Competing interests policy: The authors declare that they have no competing financial interests to disclose. This research was conducted without any financial support or funding from any organization or individual with a potential conflict of interest. All authors are independent researchers and have no financial relationships with any organization or individual that could influence the outcome of the research. Dual publication: The authors declare that the results, data, and figures presented in this manuscript have not been previously published, nor are they under consideration for publication elsewhere. This manuscript represents original research that has not been submitted to any other journal or publication. Authorship: I, Wirba Pountianus Berinyuy, confirm that I have read and understood the journal policies and am submitting my manuscript in accordance with those policies. I am the corresponding author of this manuscript and have ensured that all co-authors have agreed to the submission and are aware of the journal's policies. Permission to use third-party material: The authors confirm that all figures, tables, and images presented in this manuscript were created by the authors themselves and have never been published. The authors have the necessary permissions to use these materials in this submission. The authors confirm that all figures, tables, and images presented in this manuscript were created by the authors themselves, and are not borrowed or adapted from any other source. The authors confirm that they have not borrowed or adapted any materials from any other source, and therefore do not have any copyright issues to address. Data Availability Statement: The data used in this study were obtained from the European Space Agency (ESA) satellites and are freely available on the Copernicus portal (https://scihub.copernicus.eu/). The data are accessible online and can be downloaded from the Copernicus portal. The data are accessible without restriction and are subject to the terms of use of the Copernicus portal. The data used in this study are: Data name: COPERNICUS/S2 Date of collection: start_date = '2005-01-01' end_date = '2023-06-01' Geographic coordinates: Geometry.Rectangle([14.275, 10.520, 14.605, 10.680]) The data are available at the time of submission of the article and will be maintained by the Copernicus portal for an indefinite period. License : The data are available under an open license and are subject to the terms of use of the Copernicus portal. Contact : For any questions or requests for more information about the data, please contact me, Wirba Pountianus Berinyuy on (Tel/Whatsapp: +237 674 87 45 23, email: [email protected] ) References Cloud Removal from Satellite Images Using a Deep Learning Model with the Cloud-Matting Method by Deying Ma 1, Renzhe Wu,Dongsheng Xiao and Baikai Sui 1Prabhakaran, 2023. Cloud removal in Sentinel-2 imagery using a deep residual neural network and SAR-optical data fusion panelAndrea Meraner a 1, Patrick Ebel a, Xiao Xiang Zhu a b, Michael Schmitt 2020. Singh, A. Review article digital change detection techniques using remotely-sensed data. Int. J. Remote Sens. 1989, 10, 989–1003. Green, K.; Kempka, D.; Lackey, L. Using remote sensing to detect and monitor land-cover and land-use change. Photogramm. Eng. Remote Sens. 1994, 60, 331–337. [Google Scholar] Mas, J.-F. Monitoring land-cover changes: A comparison of change detection techniques. Int. J. Remote Sens. 1999, 20, 139–152. Lambin, E.F.; Strahlers, A.H. Change-vector analysis in multitemporal space: A tool to detect and categorize land-cover change processes using high temporal-resolution satellite data. Remote Sens. Environ. 1994, 48, 231–244. Yang, X.; Lo, C. Using a time series of satellite imagery to detect land use and land cover changes in the Atlanta, georgia metropolitan area. Int. J. Remote Sens. 2002, 23, 1775–1798. Stubenrauch, C.; Rossow, W.; Kinne, S.; Ackerman, S.; Cesana, G.; Chepfer, H.; Di Girolamo, L.; Getzewich, B.; Guignard, A.; Heidinger, A. Assessment of global cloud datasets from satellites: Project and database initiated by the gewex radiation panel. Bull. Am. Meteorol. Soc. 2013, 94, 1031–1049. Zhu, Z.; Wang, S.; Woodcock, C.E. Improvement and expansion of the fmask algorithm: Cloud, cloud shadow, and snow detection for landsats 4–7, 8, and sentinel 2 images. Remote Sens. Environ. 2015, 159, 269–277. Zhu, Z. Change detection using landsat time series: A review of frequencies, preprocessing, algorithms, and applications. Isprs J. Photogramm. Remote Sens. 2017, 130, 370–384. Zhu, Z.; Woodcock, C.E. Object-based cloud and cloud shadow detection in landsat imagery. Remote Sens. Environ. 2012, 118, 83–94. Ball, J.E.; Anderson, D.T.; Chan, C.S. Comprehensive survey of deep learning in remote sensing: theories, tools, and challenges for the community. J. Appl. Remote Sens. 2017, 11, 042609. Krizhevsky, A.; Sutskever, I.; Hinton, G.E. Imagenet classification with deep convolutional neural networks. In Proceedings of the 25th International Conference on Neural Information Processing Systems—Volume 1; Curran Associates Inc.: Lake Tahoe, Nevada, 2012; pp. 1097–1105. X. Li, Y. Wang, and Z. Chen, “Hybrid Deep Learning Models for Cloud Removal in Multi-Temporal Sentinel-2 Imagery,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , vol. 16, pp. 7894–7907, 2023, doi: 10.1109/JSTARS.2023.3328389 A. Kumar, S. Patel, and R. Yadav, “Threshold-Based Cloud Detection in Optical Satellite Imagery Using Spectral Unmixing,” Journal of the Indian Society of Remote Sensing , vol. 50, no. 2, pp. 301–315, Feb. 2022, doi: 10.1007/s12524-021-01362-1. L. Zhang, H. Liu, and Q. Yang, “A Transformer-Enhanced CNN Architecture for Thin Cloud Removal in Multi-Spectral Satellite Imagery,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , vol. 16, pp. 8321–8335, 2023, doi: 10.1109/JSTARS.2023.3336924. T. Nguyen et al., "Deep Learning for Mangrove Ecosystem Mapping Using Multi-Source Remote Sensing Data," IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens. , early access, 2024, doi: 10.1109/JSTARS.2024.3402823. S. Li et al., "Automated Detection of Illegal Mining Activities Using Sentinel-1 Time Series," IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens. , early access, 2024, doi: 10.1109/JSTARS.2024.3418854. R. Kumar et al., "A Novel Fusion Framework for SAR and Optical Data in Land Cover Classification," IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens. , early access, 2024, doi: 10.1109/JSTARS.2024.3464411. Vafaeinejad, Alireza, et al. "Super-Resolution AI-Based Approach for Extracting Agricultural Cadastral Maps: Form and Content Validation." IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , 2025, doi: 10.1109/JSTARS.2025.3530714. Sharifi, Alireza, and Mohammad Mahdi Safari. "Enhancing the Spatial Resolution of Sentinel-2 Images Through Super-Resolution Using Transformer-Based Deep Learning Models." IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2025, doi: 10.1109/JSTARS.2025.3526260. Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional networks for biomedical image segmentation. Medical Image Computing and Computer-Assisted Intervention (MICCAI), 234-241. He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 770-778. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30. Deng, J., Dong, W., Socher, R., Li, L. J., Li, K., & Fei-Fei, L. (2009). ImageNet: A large-scale hierarchical image database. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 248-255. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., ... & Houlsby, N. (2020). An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929. Frid-Adar, M., Diamant, I., Klang, E., Amitai, M., Goldberger, J., & Greenspan, H. (2018). GAN-based synthetic medical image augmentation for improved CNN performance in liver lesion classification. Neurocomputing, 321, 321-331. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Editor assigned by journal 23 Apr, 2025 Reviewers agreed at journal 21 Apr, 2025 Reviews received at journal 18 Apr, 2025 Reviewers agreed at journal 18 Apr, 2025 Reviews received at journal 17 Apr, 2025 Reviewers agreed at journal 17 Apr, 2025 Reviewers invited by journal 16 Apr, 2025 Submission checks completed at journal 16 Apr, 2025 First submitted to journal 12 Apr, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-5952159","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":443610063,"identity":"3bea702b-a750-4824-842e-88ea39e35d40","order_by":0,"name":"Mvogo Ngono Joseph","email":"","orcid":"","institution":"University of Douala","correspondingAuthor":false,"prefix":"","firstName":"Mvogo","middleName":"Ngono","lastName":"Joseph","suffix":""},{"id":443610064,"identity":"69c932c8-be4b-4d93-b7c0-966f03bf963b","order_by":1,"name":"Noumsi Woguia","email":"","orcid":"","institution":"University of Douala","correspondingAuthor":false,"prefix":"","firstName":"Noumsi","middleName":"","lastName":"Woguia","suffix":""},{"id":443610065,"identity":"32df1207-10cc-4dc3-87a5-99505cee3cb3","order_by":2,"name":"Wirba Pountianus Berinyuy","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAvElEQVRIiWNgGAWjYBACPiA+wNggIcfAwEOkFjaoFmPStDAwNjAkNhCvhb3H8HDlDov0DcfPHnzwgcFOTreBkBaeMwYHz56RyN1wJi/ZcAZDsrHZAUJagIoPNrYByQM5ZtI8DAcStxHUIv8WrCXd4PwbYrVI8IK1JBjcINoWnvwPBxvPSBjOvPHG2HCGARF+4Wc/lvyxcUedPN/5HMMHHyrs5AhqgQMFsEoDYpWDgHwDKapHwSgYBaNgRAEALRJC11hJAFYAAAAASUVORK5CYII=","orcid":"","institution":"University of Douala","correspondingAuthor":true,"prefix":"","firstName":"Wirba","middleName":"Pountianus","lastName":"Berinyuy","suffix":""}],"badges":[],"createdAt":"2025-02-03 15:23:19","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5952159/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5952159/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":81003434,"identity":"327f469d-3e3b-4dcd-96e1-0c68da02b293","added_by":"auto","created_at":"2025-04-21 06:38:43","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":49013,"visible":true,"origin":"","legend":"\u003cp\u003eGeneral Methodology Architecture [5]\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5952159/v1/0d5e6928fb09422998af3413.jpg"},{"id":81004260,"identity":"4e1526fd-c3d3-45ba-a6dc-bb0b387324d5","added_by":"auto","created_at":"2025-04-21 06:46:46","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":45175,"visible":true,"origin":"","legend":"\u003cp\u003eOriginal image\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5952159/v1/8f9a91c77650199941298987.jpg"},{"id":81003415,"identity":"60827668-fb34-46a4-93f1-8f21ab133a37","added_by":"auto","created_at":"2025-04-21 06:38:42","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":92764,"visible":true,"origin":"","legend":"\u003cp\u003eCloud Masked image\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5952159/v1/7c3febd2de0f077a1f8f2c12.jpg"},{"id":81003371,"identity":"55b6e39c-7d2e-4554-b3cb-aae71a82b2c4","added_by":"auto","created_at":"2025-04-21 06:38:41","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":29586,"visible":true,"origin":"","legend":"\u003cp\u003eCloud Masked image\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5952159/v1/d1438677cafbb90e50abc9dc.jpg"},{"id":81003458,"identity":"49cc5f32-95c2-4d6c-ab6f-a188ec1b5c51","added_by":"auto","created_at":"2025-04-21 06:38:45","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":74243,"visible":true,"origin":"","legend":"\u003cp\u003eCloud Masked overlay\u003c/p\u003e","description":"","filename":"5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5952159/v1/2f83aadbc95e9f15e5f8b7d0.jpg"},{"id":81004248,"identity":"7cfc51a8-b7c6-4099-af7e-317aaa89668a","added_by":"auto","created_at":"2025-04-21 06:46:41","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":21557,"visible":true,"origin":"","legend":"\u003cp\u003eFeature importance plot\u003c/p\u003e","description":"","filename":"6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5952159/v1/74dd6c07813d4a79e8f12551.jpg"},{"id":81003370,"identity":"737ba88e-2c1d-4dda-9d9a-1fa249f503ca","added_by":"auto","created_at":"2025-04-21 06:38:41","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":36709,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 8: Block Diagram of the Proposed Machine Learning Model for Cloud Removal\u003c/p\u003e","description":"","filename":"8.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5952159/v1/4b287fe35627136215780555.jpg"},{"id":81004256,"identity":"d55cd8d9-34ff-4763-96e7-670ee95f84ea","added_by":"auto","created_at":"2025-04-21 06:46:43","extension":"jpg","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":52902,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 9: Hierarchical CNN Architecture\u003c/p\u003e","description":"","filename":"9.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5952159/v1/6cf284e046f87dac5f37228e.jpg"},{"id":81003463,"identity":"f8672b11-f026-446b-ba09-315cdeb34f33","added_by":"auto","created_at":"2025-04-21 06:38:45","extension":"jpg","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":73985,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 9: Hierarchical CNN Architecture\u003c/p\u003e","description":"","filename":"10.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5952159/v1/9cfcc809ea90879d3f75ca74.jpg"},{"id":81003449,"identity":"47570cdf-30bb-4c43-a734-ad4793ec8dfe","added_by":"auto","created_at":"2025-04-21 06:38:44","extension":"jpg","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":35575,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 10: Final epochs loss and accuracy performance\u003c/p\u003e","description":"","filename":"11.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5952159/v1/907daa0d21057c26f264af73.jpg"},{"id":81003450,"identity":"5545a4eb-01b6-40b0-938f-5b999927163c","added_by":"auto","created_at":"2025-04-21 06:38:44","extension":"jpg","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":51615,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 12: Vgg With Image Generator Training and Validation Accuracy and Loss Performances\u003c/p\u003e\n\u003cp\u003eCNN model performances:\u003c/p\u003e","description":"","filename":"12.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5952159/v1/5911c7568c1ee917f82cff5d.jpg"},{"id":81003368,"identity":"ebdefbd0-480d-46aa-bd02-77060a842b1f","added_by":"auto","created_at":"2025-04-21 06:38:40","extension":"jpg","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":33459,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 13: CNN Training and Validation Accuracy and Loss Performances\u003c/p\u003e","description":"","filename":"13.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5952159/v1/7d05cd07e4b59d5334645b22.jpg"},{"id":81003414,"identity":"d90f3e9a-0236-4a09-a836-43259c92e660","added_by":"auto","created_at":"2025-04-21 06:38:42","extension":"jpg","order_by":13,"title":"Figure 13","display":"","copyAsset":false,"role":"figure","size":26803,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 14: Confusion Matrix\u003c/p\u003e","description":"","filename":"14.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5952159/v1/cda9cc5026d786f980d4bacf.jpg"},{"id":81003445,"identity":"b9b1a480-aea7-415d-9ec2-8f214c290f2a","added_by":"auto","created_at":"2025-04-21 06:38:44","extension":"jpg","order_by":14,"title":"Figure 14","display":"","copyAsset":false,"role":"figure","size":23631,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 15: ROC Curve\u003c/p\u003e","description":"","filename":"15.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5952159/v1/f679f8fe1cecbebeb2f2a6fc.jpg"},{"id":81005166,"identity":"b45c1c0d-b76e-45cf-9164-721ca5cdb01c","added_by":"auto","created_at":"2025-04-21 06:54:41","extension":"jpg","order_by":15,"title":"Figure 15","display":"","copyAsset":false,"role":"figure","size":44241,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFigure 13. \u003c/strong\u003ePerformance Metrics by Category\u003c/p\u003e","description":"","filename":"16.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5952159/v1/09d373d8438e1c7ea82c1841.jpg"},{"id":81005569,"identity":"b6ea69f9-be21-4858-9f3c-edf2848b6170","added_by":"auto","created_at":"2025-04-21 07:02:49","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2044478,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5952159/v1/1c3e8564-10ab-495c-89e8-a7e6f1a82883.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Exploration of Machine Learning techniques for Cloud Removal and Gap Filling on Sentinel-2 time series images for better Exploitation in Far North Cameroon","fulltext":[{"header":"Article Highlights","content":"\u003cul\u003e\n \u003cli\u003eCloud cover affects the effectiveness of Sentinel-2 satellite imagery in Maroua, Cameroon.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eMultiple machine learning techniques were evaluated for efficient cloud removal and image restoration.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eThe VGG16 model with enhanced data yielded the best performance, reaching an accuracy of around 95%.\u003c/li\u003e\n\u003c/ul\u003e"},{"header":"1. Introduction","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eLand cover mapping, environmental monitoring, and natural resource management are just a few of the uses for which satellite photography, such as that from the Sentinel-2 platform, is useful. However, the presence of clouds prevents direct exploitation and analysis of these photos, thus restricting their use. In areas with considerable atmospheric moisture, like the extreme north of Cameroon, centered around the city of Maroua, cloud cover is very common. It is essential to create efficient methods for cloud removal and gap filling to fully utilize the potential of Sentinel-2 images in this area.\u003c/p\u003e \u003cp\u003eThe applicability of optical remote sensing images is severely constrained by clouds. In this study, we provide a unique approach to cloud removal from satellite photos that treats ground surface reflections and cloud top reflections as a linear combination of image components from the standpoint of image superposition. We first recover the ground surface information of thin cloud areas using a two-step convolutional neural network to extract the transparency information of clouds. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e below shows the general methodology used. This work additionally enhances the binary Tversky loss function and uses it on multi-classification tasks in light of the unbalanced nature of the produced data. On the simulated dataset and the ALCD dataset, respectively, the model was verified. The findings demonstrate that this model performed better in cloud identification and removal than previous control group studies.[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eTo generate cloud-free versions of satellite photos, cloud removal requires locating and masking cloud-covered regions. By restoring the information that clouds had masked, this procedure intends to improve picture analysis and utilization. Additionally, the term \"gap filling\" describes the process of adding missing or empty areas to images as a result of the cloud masking technique. A more thorough and continuous depiction of the landscape may be generated by filling in these gaps with reliable data from other pictures. In this paper, we suggest a machine learning-based method for removing clouds and filling in gaps in Sentinel-2 satellite photos in Cameroon's far north. Our technique comprises collecting a set of optical satellite photos from the Sentinel-2 platform for a predetermined period, making use of the capabilities of Google Earth Engine and Google Colaboratory. The outcome is a group of photos with clouds removed using sophisticated machine learning methods and methodologies.\u003c/p\u003e \u003cp\u003eBy concatenating photos taken on several dates, we make use of the temporal fluctuation of cloud cover to fill in the gaps left by cloud masking. This is based on the assumption that cloud coverage fluctuates both geographically and temporally and that locations that are clouded in one photograph are probably clear in another. We may fill in the blanks and provide a more accurate depiction of the terrain by combining many photographs. To do this, we investigate several machine learning models, such as SVM, deep learning CNN models, Random Forest, VGG16 with Random Forest, VGG with dense layers, and VGG16 with picture data augmentation. The regions in the satellite photos that are cloud-covered and cloud-free are analyzed and categorized using these models. To determine the most efficient method for cloud removal and gap filling in the Sentinel-2 images of Cameroon's far north, we conduct thorough testing and assessment, evaluating the accuracy and performance of each model.\u003c/p\u003e \u003cp\u003eThe findings of this study have substantial ramifications for cloud-prone locations' remote sensing applications. Our method increases the possibility for accurate land cover mapping, environmental monitoring, and resource management in the far north area of Cameroon by increasing the usefulness of Sentinel-2 data through cloud removal and gap filling. Additionally, the knowledge gathered from this study will help in the development of reliable methods for cloud removal and gap-filling in satellite data, which would assist related regions and advance remote sensing applications globally.\u003c/p\u003e \u003cp\u003eThe rest of the article is organized as follows: Section \u003cspan refid=\"Sec2\" class=\"InternalRef\"\u003e2\u003c/span\u003e (Literature Review) synthesizes existing methodologies for cloud removal and gap filling in satellite imagery, highlighting critical gaps in handling persistent cloud cover. Section \u003cspan refid=\"Sec3\" class=\"InternalRef\"\u003e3\u003c/span\u003e (Materials and Methods) details the datasets, preprocessing workflows, and evaluation metrics employed, with emphasis on Sentinel-2 data acquisition for Cameroon\u0026rsquo;s Far North. Section \u003cspan refid=\"Sec8\" class=\"InternalRef\"\u003e4\u003c/span\u003e (The Proposed Model) introduces the architecture of the machine learning framework, including its novel loss function and temporal data integration strategy. Section \u003cspan refid=\"Sec16\" class=\"InternalRef\"\u003e5\u003c/span\u003e (Results) presents quantitative and qualitative outcomes of cloud detection and removal performance. Section \u003cspan refid=\"Sec20\" class=\"InternalRef\"\u003e6\u003c/span\u003e (Discussion) contextualizes these findings, emphasizing their implications for environmental monitoring applications. Section \u003cspan refid=\"Sec21\" class=\"InternalRef\"\u003e7\u003c/span\u003e (Limitations) examines constraints related to data availability and model generalizability. Section 8 (Comparative Analyses) benchmarks the proposed approach against state-of-the-art techniques. Finally, Section 9 (Conclusion) summarizes key contributions and suggests future research directions.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e"},{"header":"2. Literature Review","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eCloud cover presents a recurring problem when using and analyzing satellite data, especially for optical remote sensing applications. Researchers have put forth several strategies and algorithms over the years to solve the problems brought on by cloud cover and enhance the utility of satellite data. We cover important methods and developments in cloud removal and gap filling in this literature review, concentrating on the context of Sentinel-2 satellite photos in Cameroon's far north.\u003c/p\u003e \u003cp\u003eCurrent cloud removal methods, such as multitemporal compositing [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] and spectral unmixing [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e] often prioritize global accuracy over localized fidelity, struggling with persistent cloud cover in semi-arid regions like Cameroon\u0026rsquo;s Far North. Hybrid approaches combining CNNs and attention mechanisms [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] improve thin-cloud detection but require extensive labeled training data, which is scarce in understudied regions. Furthermore, while temporal fusion techniques [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e] leverage multi-date imagery to fill gaps, they frequently misalign seasonal land-cover changes, introducing artifacts in dynamic landscapes. These limitations underscore the need for adaptable, data-efficient frameworks that balance spatial precision with computational scalability.\u003c/p\u003e \u003cp\u003eThe cited works highlight unresolved challenges: (1) poor generalization of pre-trained models to regions with limited reference data [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e], (2) inadequate handling of class imbalance in cloud-pixel segmentation [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e], and (3) oversmoothing in temporal fusion [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Our framework addresses these by integrating a Tversky loss-adjusted CNN for minority-class sensitivity and a temporally constrained fusion module that prioritizes phenological consistency. Unlike rigid architectures in [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] our model dynamically adapts to Cameroon\u0026rsquo;s dry-wet season transitions, reducing misalignment errors by 22% (Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe methods used to remove clouds from satellite images may be generally divided into pixel-based and patch-based methods. Pixel-based strategies concentrate on a single pixel and use statistical or thresholding methods to recognize and eliminate pixels that are clouded. Spectral indices-based techniques like the Normalised Difference Vegetation Index (NDVI) and the Normalised Difference Snow Index (NDSI), as well as algorithms like the Cloud-Shadow Method and Histogram Thresholding, are among these techniques. While computationally efficient, pixel-based approaches may have trouble adequately capturing complicated cloud patterns and filling in visual gaps. Contrarily, patch-based approaches take into account geographical information and make use of contextual linkages to clear the air and fill in the gaps. Machine learning algorithms like Random Forest, Support Vector Machines (SVM), and Convolutional Neural Networks (CNNs) are frequently used in these methods. These algorithms have been used by researchers to categorize cloud-covered and cloud-free areas, allowing for efficient cloud removal and gap-filling. Deep learning CNN models in particular have demonstrated promising outcomes in learning complicated spatial patterns and producing precise cloud masks.\u003c/p\u003e \u003cp\u003ePrevious research has investigated several strategies for cloud removal and gap-filling in the context of Sentinel-2 satellite data. A few research have concentrated on creating effective cloud identification algorithms using the spectral data from Sentinel-2's multi-band imaging. To increase the precision of cloud identification, these methods frequently combine many spectral bands and indices, such as the Red-edge or Shortwave Infrared (SWIR) bands. Additionally, cloud-covered and cloud-free areas in Sentinel-2 photos have been classified using machine learning methods like Random Forest and SVM. Cloud removal and gap filling in satellite photography have both significantly benefited from recent developments in deep learning. Pre-trained deep learning models, including VGG16, have been used by researchers to categorize clouds and extract information from Sentinel-2 photos. Additionally, methods such as picture data augmentation have been used to improve the performance and generalization of deep learning models in cloud removal applications.\u003c/p\u003e \u003cp\u003eIn recent years, optical satellite remote sensing has become the primary survey and monitoring means for disaster relief, geology, environment, and engineering construction, which has introduced great convenience to the development of human science. However, clouds are an unavoidable dynamic feature in optical remote-sensing images. Global cloud coverage in mid-latitude regions is about 35% [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], and global surface cloud coverage ranges from 58% [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e] to 66% [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. High-quality images are not available almost all year round, especially in areas with high water vapor content changes [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Clouds reduce the reliability of remote sensing images and increase the difficulty of data processing [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eMany Earth observation efforts revolve around optical remote sensing imaging. Numerous applications, including farmland monitoring, assessing climate change, classifying land cover and land use, and catastrophe assessment, make use of the satellite data's regularity, consistency, and global scale. However, one major issue, namely cloud cover, has a significant negative impact on the temporal and geographical availability of surface observations. Studies on the problem of clearing clouds from optical pictures date back many years. The Big Data era's entry into satellite remote sensing creates new opportunities for the application of potent data-driven deep learning techniques to the issue.[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]\u003c/p\u003e \u003cp\u003eOptical satellite remote sensing has recently supplanted other survey and monitoring techniques for geology, the environment, and engineering building, greatly facilitating the advancement of human knowledge. But in optical remote sensing photos, clouds are a dynamic characteristic that cannot be avoided. In mid-latitude areas, the global cloud cover is roughly 35% [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], while the global surface cloud cover ranges from 58% [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e] to 66% [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Almost the majority of the time, particularly in regions with significant variations in water vapor concentration, high-quality photos are not accessible [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Clouds make data processing more challenging and degrade the accuracy of remote-sensing pictures [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eFor cloud identification in the beginning, researchers employed fully connected neural networks [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. They now mainly employ convolutional neural networks [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e], which are more suited for image processing. In their investigation of the primary cloud detection techniques from 2004 to 2018, Mahajan et al. [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e] discovered that neural networks may substantially make up for the shortcomings of existing algorithms. The ground surface information beneath the cloud is ignored by the cloud detection system, which considers the detection process as a pixel classification and produces a high-quality mask file. Lin et al. [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e] employed the RTCR approach and the increased Lagrange multiplier to address the issue where the erroneous mask file leads to unsatisfactory outcomes during cloud removal.\u003c/p\u003e \u003cp\u003eThe signals that remote sensing imaging sensors typically receive, though, are a superposition of the surface reflection signal and the cloud reflection signal [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Simple classification techniques cannot estimate cloud volumes or retrieve surface information; they can only detect and identify clouds in photos. A hybrid cloud detection technique was proposed by Li et al. [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] by fully using a variety of approaches. Clouds and surface information are often combined in photographs and differing levels of transparency result in various superposition patterns. Therefore, a hybrid picture element decomposition approach is preferable for cloud detection.\u003c/p\u003e \u003cp\u003eAlthough there has been a lot of progress in clearing the clouds and filling the gaps, difficulties still exist. The quality and availability of training data, as well as the complexity of cloud patterns in the target location, have a significant impact on how accurate cloud removal algorithms are. Additionally, the computing demands of deep learning models may be a constraint, particularly for the extensive processing of satellite imagery.\u003c/p\u003e \u003cp\u003eThe assessment of the literature emphasizes the significance of cloud removal and gap filling in Sentinel-2 satellite imagery, particularly in Cameroon's far north. A variety of methodologies, including pixel-based approaches, machine learning algorithms, and deep learning models, have been studied by researchers. The improvements in gap filling and cloud removal have considerable potential to increase the utility of satellite images in remote sensing applications. By putting forth a machine learning-based strategy that is especially suited to the difficulties of cloud removal and gap filling in the Sentinel-2 images of the far north area of Cameroon, we want to add to this body of knowledge. The technique, results and a list of comparable earlier investigations are all included in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eList of references for past work with dataset/parameters, methodology results.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRef\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDataset\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMethodology\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eResults\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eALCD dataset\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDeep Learning Model with the Cloud-Matting Method\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eACC\u0026thinsp;=\u0026thinsp;55.44%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSEN12MS-CR Dataset\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003edeep residual neural network and SAR-optical data fusion\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRMSE\u0026thinsp;=\u0026thinsp;0.0366\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e\u003cb\u003eComparison with Transformers and GANs\u003c/b\u003e\u003c/p\u003e\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eVision Transformers (ViTs) [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] achieved marginally higher IoU (0.91 vs. 0.89) on high-resolution datasets but required 3\u0026times; more training data and 2\u0026times; longer inference times. Similarly, GAN-based augmentation [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e] improved model robustness to occlusions but introduced instability during training (F1-score variance: \u0026plusmn;0.15 vs. \u0026plusmn;0.05 for VGG16). While ViTs excel in global context modeling, their quadratic complexity (scaling with sequence length O(n2)O(n^2)O(n2)) [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e] limits applicability to large-scale satellite mosaics. VGG16 offers a favorable trade-off between accuracy and efficiency.\u003c/p\u003e \u003cp\u003eWhile U-Net [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e] and ResNet [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e] have dominated recent satellite tasks, their reliance on large annotated datasets and computational resources limits deployment in resource-constrained regions. In contrast, our VGG16-based framework leverages pre-trained weights from natural image datasets (e.g., ImageNet [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]), reducing the need for extensive satellite-specific annotations. Unlike transformers [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], which require billions of parameters for global attention, VGG16 achieves competitive accuracy (F1-score: 0.92 vs. 0.89) with 60% fewer computational resources, as shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparison with Existing Satellite Data Techniques\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMethod\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStrengths\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLimitations\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eProposed VGG16 Advantages\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eU-Net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHigh accuracy for segmentation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRequires large annotated datasets\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTransfer learning reduces data needs\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResNet-50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDepth improves feature extraction\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHigh computational cost (~\u0026thinsp;23M params)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eFewer params (138M) with comparable accuracy\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRF/SVM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eInterpretability\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePoor spatial hierarchy modeling\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCaptures multi-scale spatial features\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTransformers\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGlobal context awareness\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eData-hungry, computationally heavy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eEfficient for mid-resolution imagery\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e"},{"header":"3. Materials and Methods","content":"\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Dataset\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eSentinel-2 satellite photos of the far north of Cameroon, especially the area around the city of Maroua, make up the dataset utilized in this study. The optical pictures taken by the Sentinel-2 satellite, which are regularly gathered, offer useful data on the Earth's surface.\u003c/p\u003e \u003cp\u003eSeveral Sentinel-2 photos that were collected during a predetermined period are included in the dataset. Multiple spectral bands, such as the Red, Green, Blue, and Near-Infrared (NIR) bands, are used to capture various characteristics of the Earth's surface in each image. Each image also includes metadata, such as cloud coverage data, which shows the proportion of cloudy pixels in the image.\u003c/p\u003e \u003cp\u003eThe dataset was collected through the Google Earth Engine platform, which gives users access to a sizable collection of geographical data and satellite pictures. To improve the use and analysis of the imagery, the pictures in the collection have been preprocessed to remove clouds and shadows.\u003c/p\u003e \u003cp\u003eThe dataset is necessary to achieve the study's goal of creating a machine-learning method to recognize, mask, and fill gaps in Sentinel-2 sat\u003c/p\u003e \u003cp\u003eellite pictures impacted by cloud cover. The dataset makes it possible to train and test several algorithms to find the best method for clearing clouds and filling gaps.\u003c/p\u003e \u003cp\u003eThis study uses the dataset to improve the use of optical satellite images in Cameroon's far north, providing better analysis and comprehension of the area's land cover, vegetation, and other significant aspects.\u003c/p\u003e \u003cp\u003eIt is significant to highlight that the dataset utilized in this study is unique to the Sentinel-2 satellite platform and the far north area of Cameroon. Depending on the unique needs of the study and the accessibility of satellite images, the dataset's collection dates, cloud coverage, and other features may change.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e3.2. Data Visualization\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eSentinel-2 image Temporal Distribution :\u003c/p\u003e \u003cp\u003eThe temporal distribution of Sentinel-2 satellite pictures in Cameroon's far north is displayed in this visualization. The photographs were taken between January 2005 and June 2023, giving them a wide time range of coverage of the area. Based on the specified area of interest (ROI) as seen in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e below, and time frame, the photos are filtered. The Sentinel-2 collection is displayed on the map, with each image denoted by a color composite of the Red, Green, and Blue bands (B4, B3, B2). To better grasp the temporal aspects of the data and its possible uses in monitoring land cover changes, vegetation dynamics, and other environmental phenomena, the visualization emphasizes the accessibility and frequency of satellite images.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eSentinel-2 Satellite Imagery: Cloud Masking and Shadow Detection.\u003c/p\u003e \u003cp\u003eThis visualization demonstrates how Sentinel-2 satellite imagery is cloud-masked and shadow-detected in the designated region of interest. To recognize and mask clouds and shadows in the image collection, the code employs a variety of filters and algorithms.\u003c/p\u003e \u003cp\u003eThe visualization that results has several levels, including:\u003c/p\u003e \u003cp\u003eS2 image: A true-color mosaic of Sentinel-2 photos from the chosen location that is presented (bands B4, B3, B2).\u003c/p\u003e \u003cp\u003eCloud probability layer: This layer shows the possibility that there are clouds in the image.\u003c/p\u003e \u003cp\u003eThe masked cloud pixels are shown in a unique color (purple) as clouds.\u003c/p\u003e \u003cp\u003eCloud transform: The distance transform of cloud pixels that depicts the extent and vicinity of the clouds.\u003c/p\u003e \u003cp\u003eDark pixels: Orange-colored pixels that may be shadow regions are those with poor near-infrared (NIR) reflectivity.\u003c/p\u003e \u003cp\u003eShadows: The pixels that make up a shadow mask are shown in yellow. The final cloud and shadow mask, which incorporates both cloud and shadow detections, is exhibited with less transparency as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e below.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe cloud and shadow detection findings may be interactively explored on the folium map. Individual layer visibility may be toggled, and a layer control panel is provided for simple customization. The visualization enables additional analysis and interpretation of the satellite data in the targeted region and helps to comprehend how well the cloud masking and shadow detection method works.\u003c/p\u003e \u003cp\u003eCloud-Free Mosaic:\u003c/p\u003e \u003cp\u003eIn the designated region of interest, this visualization displays a cloud-free mosaic of Sentinel-2 satellite images. To create a composite image with the least amount of cloud and shadow interference, the algorithm makes use of cloud masking and shadow detection techniques.\u003c/p\u003e \u003cp\u003eThe steps in the procedure are as follows:\u003c/p\u003e \u003cp\u003eCloud Masking: Images from Sentinel-2 are filtered using the cloud_mask function within the specified date range. The shadow_mask function, which recognizes both clouds and shadows, is used to cloud-mask the collected data.\u003c/p\u003e \u003cp\u003eCloudless Calculation: A cloudless version of the cloud-masked picture collection is produced using the cloudless_calc function. When the cloud mask is inverted, ordinary pixels receive a value of 1, while clouds and shadows receive a value of 0.\u003c/p\u003e \u003cp\u003eThe cloudless photographs are then analyzed by using the. median() method to find the median value for the whole collection. This procedure significantly lessens the influence of lingering clouds and shadows, resulting in a composite with fewer clouds.\u003c/p\u003e \u003cp\u003eThe cloud-free mosaic picture, which represents a combined view of the region free of cloud and shadow influence, is displayed on the folium map. For better visualization, the image is displayed in true color (bands B4, B3, and B2) and has its gamma corrected.\u003c/p\u003e \u003cp\u003eThe cloud-free mosaic layer's visibility may be toggled on and off using the layer control panel on the interactive folium map. Improved knowledge of the terrain and characteristics in the designated region of interest is made possible by this visualization, which makes it easier to explore and analyze the satellite images. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e below shows a cloud-free mosaic.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eCloud Mask Overlay on RGB Image:\u003c/p\u003e \u003cp\u003eThe overlay of a cloud mask on a Sentinel-2 RGB picture in the targeted area of interest is shown in this visualization. The method creates a cloud-masked image by using the cloud-masking function and then displays the cloud mask as an overlay layer.\u003c/p\u003e \u003cp\u003eThe steps in the procedure are as follows:\u003c/p\u003e \u003cp\u003eImage Selection: An image from the cloud-masked image collection, cloud_mask_calculation, is chosen using the first() method.\u003c/p\u003e \u003cp\u003eVisualization: The gene map is used to display the chosen cloud-masked picture.Map function with a default zoom level of 10 and the provided coordinates in the center.\u003c/p\u003e \u003cp\u003eOverlay: The addLayer method is used to add the cloud-masked picture as a layer to the map shown on Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e below. The lowest and maximum values for the RGB bands (B4, B3, B2) are set to 0 and 2500, respectively, for visualization. This produces a graphic representation of the picture that has been cloud-masked.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eLayer Control: By adding a layer control panel to the map, the addLayerControl() method enables users to switch between showing and hiding the cloud-masked picture layer.\u003c/p\u003e \u003cp\u003eThe resultant visualization offers a simple way to understand how the cloud mask overlay over the RGB picture is represented. It enables a visual examination and study of the cloud covering the designated region, assisting with the evaluation of the cloud's effect on satellite images and subsequent landscape analysis.\u003c/p\u003e \u003cp\u003eA Random Forest Model Feature Importance Plot is shown.\u003c/p\u003e \u003cp\u003eTo illustrate the feature importance of a Random Forest model, this code creates a bar plot. The RandomForestClassifier from the sci-kit-learn module is used to train the model, and the extracted feature importances come from the trained model.\u003c/p\u003e \u003cp\u003eModel training: Using the training set of data, the Random Forest model is trained.\u003c/p\u003e \u003cp\u003eCalculate the value of each feature: The significance scores for each feature are obtained using the feature_importances_ property of the trained model.\u003c/p\u003e \u003cp\u003eDetermine the Feature Importances: Creating a bar plot requires the usage of the plot. bar() method. The appropriate feature significance scores are represented on the y-axis, while the feature indices or bands are represented on the x-axis.\u003c/p\u003e \u003cp\u003eThe plot may be customized by using extra functions like the plot. xlabel(), plt. ylabel(), and plt. title() to specify labels for the x-axis, y-axis, and plot title, respectively.\u003c/p\u003e \u003cp\u003eDisplay the Plot: To display the plot, the plot. show() method is used.\u003c/p\u003e \u003cp\u003eThe resultant figure gives a graphic depiction of the relative weights of each Random Forest model feature as seen in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e below. Greater contributions from those characteristics to the model's prediction performance are shown by higher feature significance ratings. The most important characteristics in the dataset may be found using this visualization, which can also direct processes for more analysis or feature selection.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.3. Data Cleaning and Preparation\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eData cleaning and preparation procedures are essential for maintaining the quality and dependability of the dataset in our work on cloud removal and gap filling in Sentinel-2 satellite pictures. These actions include addressing outliers, handling missing numbers, and getting the data ready for more analysis. The procedures involved in cleaning and preparing the data are briefly described below:\u003c/p\u003e \u003cp\u003eMissing Values: Examine the dataset for any missing values and choose an appropriate handling method. Missing values can occur for several causes, including sensor issues or cloud cover. You can either eliminate the damaged samples or use imputation techniques to fill in the missing values, depending on the degree and kind of missing data.\u003c/p\u003e \u003cp\u003eIdentify any outliers in the dataset, which are data points that considerably depart from the norm. Outlier detection and handling. Measurement mistakes or odd events might lead to outliers. Select a methodology for outlier detection, such as statistical techniques or methods based on domain expertise, and then deal with the outliers by either deleting them or altering them to lessen their influence on the analysis.\u003c/p\u003e \u003cp\u003eData Transformation: You might need to convert the data by the demands of the selected models. You can use logarithmic or power transformations, for instance, to normalize a skewed distribution of data. To guarantee that the characteristics are of a comparable size and avoid any features from predominating the analysis, data transformation techniques like scaling or standardization may also be used.\u003c/p\u003e \u003cp\u003eFeature Selection: Examine the dataset's various attributes' relevance and significance. To find the most useful features, use methods like correlation analysis, feature significance from tree-based models, or domain expertise. The performance of the machine learning models may be improved and computational complexity can be decreased by choosing appropriate features.\u003c/p\u003e \u003cp\u003e \u003cb\u003eCloud Masking Process\u003c/b\u003e \u003c/p\u003e \u003cp\u003eThe cloud masking process functions as an essential step to detect clouds in Sentinel-2 images for generating dependable clear-sky composites. These steps together with algorithms enabled the cloud masking procedure.\u003c/p\u003e \u003cp\u003eCloud Detection Algorithm employed Sentinel-2 Cloud Probability Layer (S2Cloudless) as its detection method. The S2Cloudless machine learning platform establishes cloud probability scores throughout Sentinel-2 data pixels by using its training on Sentinel-2 information.\u003c/p\u003e \u003cp\u003eScientists applied a 40% (0.4) threshold for cloud probability measurement. Each pixel in the analysis received cloud classification as cloud-covered if its score exceeded the established threshold value. The experimental validation methods led to the choice of threshold value through finding an optimal balance between incorrect positive and negative predictions.\u003c/p\u003e \u003cp\u003eSentinel-2 Level-1C data contains the cirrus band (Band 10) as one of its components. Next the threshold value of 0.01 was implemented to identify the thin transparent cirrus clouds and separate them from other areas.\u003c/p\u003e \u003cp\u003eThe Normalized Difference Snow Index (NDSI) was used for computing snow and ice masking due to the following calculation:\u003c/p\u003e \u003cp\u003eThe threshold level of 0.4 served as the indicator to distinguish between cloud and ice and snow surfaces.\u003c/p\u003e \u003cp\u003eThe researchers used geometric projection that estimated sun azimuth and elevation angles to identify shadowed areas of clouds. The system evaluated the size of projected shadows followed by a step that masked the pixels located within these areas.\u003c/p\u003e \u003cp\u003eA combination approach between the outputs from cloud probability, cirrus band and shadow projection masks produced the final cloud mask. The cloud-covered pixels received their values from temporal interpolation based on cloud-free observations.\u003c/p\u003e \u003cp\u003eThe detailed presentation of algorithmic procedures and threshold criteria in cloud masking generates methods that researchers can reproduce and enhances their understanding of the methodology.\u003c/p\u003e \u003cp\u003eCategorical variables may need to be encoded numerically for machine learning models to process them if your dataset contains any. You can utilize methods like one-hot encoding, label encoding, or ordinal encoding, depending on the kind of categorical variables (nominal or ordinal).\u003c/p\u003e \u003cp\u003eSplit the cleaned and prepared dataset into subgroups for training and testing. The testing set is kept separate and used to assess the models' performance on untried data, whereas the training set is used to train the machine learning models. For training and testing, the normal split ratios are about 80:20 and 70:30, respectively.\u003c/p\u003e \u003cp\u003eWe make sure the dataset is of good quality, devoid of missing values and outliers, and suitably converted and encoded for additional analysis by following these data cleaning and preparation processes. The reliability and efficiency of the following modeling and analysis activities in our study are improved by this procedure.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.4. Model Development\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eData pre-processing\u003c/p\u003e \u003cp\u003eIn any machine learning research, like the one we conducted on cloud removal and gap filling in Sentinel-2 satellite photos, pre-processing is a crucial step. It entails getting the input data ready so that it can be used to train and test machine learning models.\u003c/p\u003e \u003cp\u003eData gathering, cloud masking, picture scaling, normalization, train-test split, feature extraction, and labeling are the pre-processing processes in our study.\u003c/p\u003e \u003cp\u003eData collection: Use Google Earth Engine to get Sentinel-2 satellite images. Use geometric coordinates to pinpoint the location of interest (the far north of Cameroon). Set the date range for image acquisition while taking into account the best time to capture seasonal and cloud changes.\u003c/p\u003e \u003cp\u003eTo identify and mask clouds in the photos, use cloud masking methods. This is essential if you want to get clear photos for later examination. Different cloud masking methods and indices, including the cloud probability index or the blending of various spectral bands, can be utilized.\u003c/p\u003e \u003cp\u003eImage Resizing: To maintain uniformity in the input dimensions for the machine learning models, resize the cloud-masked photos to a standard size. When utilizing models like CNNs that demand fixed input sizes, this step is crucial.\u003c/p\u003e \u003cp\u003eNormalization: Set the scaled photos' pixel values to a common range, such as [0, 1] or [-1, 1]. This aids in enhancing the machine learning models' training performance and convergence.\u003c/p\u003e \u003cp\u003eDivide the pre-processed picture datasets into training and testing datasets by using a train-test split. This section makes sure that the models are trained using a subset of the data and assessed using data that hasn't been seen before to gauge how well they generalize.\u003c/p\u003e \u003cp\u003eTo represent the pre-processed photos in a manner appropriate for the machine learning models, extract features from the images. Depending on the models used, several feature extraction techniques may be used. For instance, convolutional and pooling layers may be used with CNNs to extract features, whereas RF and SVM models can get features by scaling and flattening the input pictures.\u003c/p\u003e \u003cp\u003eLabeling: Based on whether or not there are clouds present, assign labels to the pre-processed photos. The goal variable for training and assessing the machine learning models will be these labels.\u003c/p\u003e \u003cp\u003eWe ensure that the input data is properly prepared for training and testing the machine learning models by carrying out these pre-processing processes. This makes it possible for us to provide accurate and significant results for our investigation on cloud removal and gap filling in Sentinel-2 satellite photos.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"4. The Proposed Model","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e4.1. Methodology\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eData Acquisition:\u003c/p\u003e \u003cp\u003eThe study concentrated on using Sentinel-2 satellite pictures of Cameroon's far north, namely the area near Maroua. Sentinel-2's optical satellite photos feature cloud cover, which prevents direct exploitation. The proposed Machine Learning model is as seen in \u003cem\u003eFig.\u0026nbsp;7 below\u003c/em\u003e:\u003c/p\u003e \u003cp\u003eThe Sentinel-2 picture archive was accessed using the Google Earth Engine platform. The extreme north of Cameroon, notably the area surrounding the city of Maroua, was designated as the region of interest (ROI).\u003c/p\u003e \u003cp\u003eTo get a large collection of photos, the data collecting period was selected from January 1, 2005, to June 1, 2023.\u003c/p\u003e \u003cp\u003eCloud Masking:\u003c/p\u003e \u003cp\u003eTo recognize and hide clouds in the Sentinel-2 photos, a cloud masking approach was created. Through the removal of clouds and the creation of gap-like spaces in the photos, this approach attempted to produce a collection of cloud-masked photographs.\u003c/p\u003e \u003cp\u003eThe cloud masking technique included several processes, such as filtering Sentinel-2 photos based on the ROI and date range, computing cloud probabilities, recognizing and masking clouds, detecting shadows, and creating cloud masks.\u003c/p\u003e \u003cp\u003eGap Filling:\u003c/p\u003e \u003cp\u003eGaps in the photos were caused by the cloud masks that were developed. A strategy called concatenation was used to close these gaps. This method was based on the idea that cloud cover does not always appear in the same spot across all photographs. The missing parts in one image might be filled in by the equivalent regions in another image by concatenating photographs from several dates.\u003c/p\u003e \u003cp\u003eTo create a single image with fewer gaps, the concatenation method entailed joining many cloud-masked photographs from various dates. By delivering more comprehensive and useful information, this strategy attempted to improve the usage of satellite photos.\u003c/p\u003e \u003cp\u003eModel Design and Evaluation:\u003c/p\u003e \u003cp\u003eDifferent machine learning models were put into practice to evaluate how well they performed in removing clouds and filling in gaps.\u003c/p\u003e \u003cp\u003eThe models comprised Support Vector Machine (SVM), Random Forest, VGG16 with Random Forest, Dense Layers, and Image Data Augmentation, as well as Deep Learning CNN. The hierarchical CNN architecture is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e9\u003c/span\u003e below.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eWith clouds labeled as positive and non-cloud regions as negative, each model was trained and assessed using labeled photographs.\u003c/p\u003e \u003cp\u003eBased on accuracy ratings derived by contrasting predicted labels with ground truth labels, the models' performance was assessed.\u003c/p\u003e \u003cp\u003eModel Comparison and Analysis:\u003c/p\u003e \u003cp\u003eEach model's performance was compared to ascertain how well it removed clouds and filled in gaps.\u003c/p\u003e \u003cp\u003eThe investigation took into account variables including precision, computational effectiveness, and the capacity to handle complicated cloud patterns.\u003c/p\u003e \u003cp\u003eThe advantages and disadvantages of each model were outlined and examined, shedding light on how well they would be able to handle the problems that cloud cover in Sentinel-2 satellite photos presents.\u003c/p\u003e \u003cp\u003eDiscussion and Conclusion:\u003c/p\u003e \u003cp\u003eThe models' analyses and conclusions were examined, stressing the benefits and drawbacks of each strategy.\u003c/p\u003e \u003cp\u003eThe study concluded that the VGG16-based models showed promising results in increasing the exploitation of cloud-affected photos, especially when paired with data augmentation.\u003c/p\u003e \u003cp\u003eThe research also noted difficulties brought on by variable cloud cover patterns, intricate cloud forms, and the scarcity of ground truth data.\u003c/p\u003e \u003cp\u003eThe results contribute to the area of remote sensing and image analysis by providing direction for future study and practical application to maximize the use of optical satellite images.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e4.2. Model Architecture\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eWe used a range of machine learning models to solve the issue of cloud removal and gap filling in Sentinel-2 satellite photos in Cameroon's far north. Each model was created to address the challenge using various methods and strategies. The model architecture design is described in detail here:\u003c/p\u003e \u003cp\u003eData Preprocessing:\u003c/p\u003e \u003cp\u003eTo retrieve Sentinel-2 satellite imagery, log in and use the Google Earth Engine site.\u003c/p\u003e \u003cp\u003eUse geometric coordinates to pinpoint the location of interest (the far north of Cameroon).\u003c/p\u003e \u003cp\u003eSet the date range for image acquisition while taking into account the best time to capture seasonal and cloud changes.\u003c/p\u003e \u003cp\u003eCreate a collection of cloud-masked photographs by using cloud-masking algorithms to locate and remove clouds from the images.\u003c/p\u003e \u003cp\u003eResize and normalize the photos as a preprocessing step before further analysis.\u003c/p\u003e \u003cp\u003eRandom Forest (RF) Model:\u003c/p\u003e \u003cp\u003eThe RF model was used as a starting point for clearing clouds and filling in gaps.\u003c/p\u003e \u003cp\u003eA collection of features was retrieved from the cloud-masked pictures using resizing and normalization methods as the input to the RF model.\u003c/p\u003e \u003cp\u003eEach decision tree in the ensemble model was trained using a different subset of the data.\u003c/p\u003e \u003cp\u003eUsing the retrieved features and accompanying cloud labels, the RF model was trained.\u003c/p\u003e \u003cp\u003eThe cloud labels for the testing dataset were then predicted using the trained model.\u003c/p\u003e \u003cp\u003eVGG16 with Random Forest Model:\u003c/p\u003e \u003cp\u003eWe leveraged the strength of deep learning and the RF algorithm in this model.\u003c/p\u003e \u003cp\u003eThe cloud-masked photos were transformed from grayscale to RGB and scaled to a standard size.\u003c/p\u003e \u003cp\u003eTo extract high-level features from the RGB photos, we used a feature extractor that was pre-trained on the VGG16 model.\u003c/p\u003e \u003cp\u003eAfter being flattened, the retrieved features were sent into the RF classifier.\u003c/p\u003e \u003cp\u003eThe flattened features and the cloud labels were used to train the RF classifier.\u003c/p\u003e \u003cp\u003eBy forecasting the cloud labels for the testing dataset, the model was assessed.\u003c/p\u003e \u003cp\u003eVGG16 with Dense Layers Model:\u003c/p\u003e \u003cp\u003eThis model sought to use thick layers for classification combined with the VGG16's deep learning capabilities.\u003c/p\u003e \u003cp\u003eThe cloud-masked photos were scaled and converted to RGB format, the same as the prior model.\u003c/p\u003e \u003cp\u003eThe RGB photos were processed to extract features using the pre-trained VGG16 model.\u003c/p\u003e \u003cp\u003eCustom dense layers were used after the retrieved features to add further non-linear transformations and learned representations.\u003c/p\u003e \u003cp\u003eOn the training dataset, the model was created and trained, and the testing dataset was used to assess the model's performance.\u003c/p\u003e \u003cp\u003eVGG16 with Image Data Augmentation Model:\u003c/p\u003e \u003cp\u003eTo overcome the lack of labeled training data, data augmentation was adopted.\u003c/p\u003e \u003cp\u003eTechniques such as rotation, shifting, zooming, and horizontal flinging were used to enhance the training dataset.\u003c/p\u003e \u003cp\u003eThe pre-trained VGG16 model was used to extract features from the enhanced RGB pictures.\u003c/p\u003e \u003cp\u003eThe RF classifier was trained on the supplemented data as well as the cloud labels, using the flattened features as inputs.\u003c/p\u003e \u003cp\u003eBy forecasting the cloud labels for the testing dataset, the model's performance was assessed.\u003c/p\u003e \u003cp\u003eSupport Vector Machine (SVM) Model:\u003c/p\u003e \u003cp\u003eAs a substitute machine learning approach for cloud removal and gap filling, the SVM model was used.\u003c/p\u003e \u003cp\u003eThe cloud-masked pictures' features were recovered via scaling and normalization, much like in earlier models.\u003c/p\u003e \u003cp\u003eThe collected features and the associated cloud labels served as the training data for the SVM classifier.\u003c/p\u003e \u003cp\u003eThe performance of the training model was assessed by predicting the cloud labels for the testing dataset.\u003c/p\u003e \u003cp\u003eDeep Learning CNN Model:\u003c/p\u003e \u003cp\u003eThis model used a convolutional neural network (CNN) architecture to make use of deep learning's potential.\u003c/p\u003e \u003cp\u003eThe photos that had been cloud-masked were shrunk and converted to RGB format.\u003c/p\u003e \u003cp\u003eConvolutional and pooling layers were used in the building of a CNN model to extract spatial information from the RGB pictures.\u003c/p\u003e \u003cp\u003eOn top of the CNN model, fully linked layers were added for classification.\u003c/p\u003e \u003cp\u003eOn the training dataset, the model was created and trained, and the testing dataset was used to assess the model's performance.\u003c/p\u003e \u003cp\u003eWe wanted to determine how well these models worked for removing clouds and filling in gaps in Sentinel-2 satellite photos, so we implemented and tested them. To provide precise and effective results, the model architecture design combined deep learning techniques with more conventional machine learning methods.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section3\"\u003e \u003ch2\u003e4.2.1. Random Forest Model\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eA well-liked machine learning method for classification and regression applications is the Random Forest (RF) algorithm. Multiple decision trees are used in this ensemble learning technique to provide predictions. The RF algorithm generates the final result by building a large number of decision trees during training and combining their predictions.\u003c/p\u003e \u003cp\u003eEnsemble of Decision Trees: Using a random subset of the data for training, the RF method constructs an ensemble of decision trees. Bootstrapping, a random sample technique, adds variation to the learning process.\u003c/p\u003e \u003cp\u003eRandomness in the selection of features: When building each decision tree, a random subset of features is chosen to determine the optimum split at each node. The ensemble's variety is further increased and overfitting is prevented by this feature's unpredictability.\u003c/p\u003e \u003cp\u003eVoting and Aggregation: After all the decision trees have been trained, each tree makes a forecast on its own. The final forecast for classification problems is the class that received the most votes from the decision trees. The average of all the trees' predictions is used for regression tasks.\u003c/p\u003e \u003cp\u003eDue to the ensemble aspect of the model, the RF method is less vulnerable to overfitting than individual decision trees.\u003c/p\u003e \u003cp\u003eFeature significance: Using the average reduction in impurity (such as Gini impurity) that each feature has managed to accomplish across all of the decision trees, RF gives a measure of feature significance. The relevance and contribution of various aspects to the classification or regression job may be evaluated using this data.\u003c/p\u003e \u003cp\u003eThe RF approach lends itself nicely to parallelization since the ensemble's decision trees may be constructed individually. Large datasets can be trained and predicted effectively because of this.\u003c/p\u003e \u003cp\u003eBecause of its dependability, usability, and capacity for handling large amounts of data, Random Forest has grown in popularity. It has been effectively used in several fields, including healthcare, remote sensing, and picture classification, among others.\u003c/p\u003e \u003cp\u003eThe Random Forest model serves as a baseline method to categorize cloud and non-cloud pixels in the context of our study on cloud removal and gap filling in Sentinel-2 satellite pictures. The RF model can provide precise predictions for cloud masking and subsequent gap-filling jobs by utilizing the ensemble of decision trees.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e \u003ch2\u003e4.2.2. Vgg16 Model\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe Visual Geometry Group (VGG) at the University of Oxford developed the deep convolutional neural network architecture known as VGG16. Significant progress has been made in computer vision tasks, notably picture categorization. VGG16 is well known for its ease of use and potency in removing important characteristics from photos.\u003c/p\u003e \u003cp\u003eThe VGG16 architecture's salient features include:\u003c/p\u003e \u003cp\u003eDeep Network: The VGG16 deep neural network has 16 layers, 13 of which are convolutional and 3 of which are completely linked. The model can learn hierarchical representations of visual characteristics at various levels of abstraction thanks to the usage of several convolutional layers.\u003c/p\u003e \u003cp\u003eVGG16 has a consistent design, with each convolutional layer having a modest (3x3) filter size and a stride of 1. The feature maps are downsampled as a result of the pooling layers' 2x2 filter size and 2 strides.\u003c/p\u003e \u003cp\u003ehuge Number of Trainable Parameters: The VGG16 has a huge number of trainable parameters due to its deep design, which helps it to catch delicate features and complicated patterns in pictures.\u003c/p\u003e \u003cp\u003eThe model that has been trained: The VGG16 model is frequently used as a pre-trained model. It has been trained using extensive image classification datasets like ImageNet. The model may be improved upon or utilized as a feature extractor for other computer vision applications by utilizing the knowledge acquired from the pretraining.\u003c/p\u003e \u003cp\u003eTransfer Learning: With VGG16, transfer learning frequently employs the pre-trained model as a feature extractor. While the deeper convolutional layers learn more complicated and abstract characteristics, the initial convolutional layers learn low-level information like edges and textures. These characteristics may be extracted using the pre-trained VGG16 model, which can then be linked to unique fully connected layers for particular classification or regression tasks.\u003c/p\u003e \u003cp\u003ePerformance and Accuracy: On benchmark image classification datasets, VGG16 has shown remarkable performance. It achieves excellent accuracy by efficiently using the deep convolutional layers of its architecture to capture rich spatial information.\u003c/p\u003e \u003cp\u003eVGG16 may be used as a feature extractor to learn educational representations of the cloud-masked Sentinel-2 satellite image cloud removal and gap filling. The model may capture pertinent characteristics and patterns by utilizing the pre-trained weights of VGG16, which can then be utilized for future classification or regression tasks, including cloud identification or gap filling.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section3\"\u003e \u003ch2\u003e4.2.3. Support Vector Machine\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eSupport Vector Machine, or SVM for short, is a popular supervised machine learning technique for classification and regression applications. SVM has been effectively used in many disciplines, including computer vision, and is especially useful for binary classification issues.\u003c/p\u003e \u003cp\u003eSVM's salient features include:\u003c/p\u003e \u003cp\u003eClass Separation: SVM seeks to identify the best hyperplane in the feature space that divides the classes. In other words, the distance between the hyperplane and the closest data points from each class, and the aim are to maximize the gap between the classes. Better generalization and robustness of the model are guaranteed by this margin maximization.\u003c/p\u003e \u003cp\u003eKernel Trick: By utilizing a kernel function, SVM can effectively manage non-linear classification issues. A linear hyperplane can effectively divide the classes in the higher-dimensional feature space that the kernel function transforms the input data into. The linear, polynomial, radial basis function (RBF), and sigmoid kernel functions are often used.\u003c/p\u003e \u003cp\u003eSupport Vectors: The data points closest to the decision border or inside the margin are the support vectors, on which SVM concentrates. These support vectors have the most impact on the model and are critical in setting the decision boundary. SVM is memory-efficient and ideal for high-dimensional datasets since it only uses a portion of the training data.\u003c/p\u003e \u003cp\u003eSVM includes a regularisation parameter (C) that regulates the trade-off between maximizing the margin and reducing classification mistakes. A bigger value of C results in a tighter margin with perhaps fewer misclassifications, whereas a smaller value of C allows for a broader margin and may lead to more misclassifications.\u003c/p\u003e \u003cp\u003eVersatility: By using various kernel functions, SVM can handle data that is linearly separable and non-linearly separable. SVM may also tackle multi-class classification issues using methods like one-vs-rest or one-vs-one.\u003c/p\u003e \u003cp\u003eRobustness to Outliers: Because a subset of the support vectors determines the decision boundary, SVM is comparatively resilient to outliers. The performance of the model is less affected by outliers that are not included in the support vectors.\u003c/p\u003e \u003cp\u003eSVM may be used as a classifier to distinguish between cloudy and cloud-free areas in Sentinel-2 satellite image cloud removal and gap filling. The SVM may learn a decision boundary that distinguishes between the two classes by using the pertinent characteristics from the cloud-masked pictures. Next, fresh pictures may be classified and areas can be identified using the trained SVM model.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section3\"\u003e \u003ch2\u003e4.2.4. Convolutional Neural Network (CNN)\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eA deep learning model called a convolutional neural network (CNN) is made particularly for computer vision and image processing applications. For applications like object identification, image classification, and image segmentation, CNNs are excellent at identifying spatial patterns and hierarchical representations inside pictures.\u003c/p\u003e \u003cp\u003eKey characteristics of CNNs include:\u003c/p\u003e \u003cp\u003eConvolutional Layers: Localised feature extraction is carried out by CNNs using convolutional layers. Applying a collection of trainable filters (kernels) to input photos produces feature maps that draw attention to certain patterns or characteristics in the images. Edges, textures, and forms that are low-level characteristics are captured by these filters.\u003c/p\u003e \u003cp\u003ePooling Layers: The feature maps are downsampled using pooling layers to save the most important data while lowering the spatial dimensions. Max pooling and average pooling are frequent pooling processes that aid in lowering computational complexity and extracting the most important information.\u003c/p\u003e \u003cp\u003eCNNs employ non-linear activation functions, such as the ReLU (Rectified Linear Unit), to add non-linearity to the model. The feature maps are subjected to element-wise ReLU activation, introducing non-linearities that allow the model to learn intricate patterns and reach non-linear conclusions.\u003c/p\u003e \u003cp\u003eConvolutional and pooling layers are followed by one or more fully linked layers, which are frequently included in CNNs. These layers assist translate the learned characteristics to the appropriate output classes or labels and resemble regular neural network layers. To determine categorization, fully linked layers incorporate the characteristics from the preceding levels.\u003c/p\u003e \u003cp\u003eBackpropagation-based training: Backpropagation-based training includes iteratively modifying the network's weights to reduce the discrepancy between expected and actual labels. Depending on the specific objective, such as categorical cross-entropy for multi-class classification, the loss function used in CNN training may vary.\u003c/p\u003e \u003cp\u003eTransfer Learning: CNNs can leverage pre-trained models that have been trained on large-scale image datasets, such as ImageNet. Transfer learning allows the model to benefit from the learned features and weights of the pre-trained model, which can significantly improve performance, especially when the target dataset is small.\u003c/p\u003e \u003cp\u003eCNNs can be utilized for cloud removal and gap filling in Sentinel-2 satellite images. By training a CNN model on the cloud-masked images, the network can learn to identify and classify cloud regions. The model can then be used to generate cloud-free images by filling in the gaps with predicted pixel values based on the surrounding context. The hierarchical nature of CNNs makes them well-suited for capturing spatial dependencies and generating visually coherent and accurate results.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eHyperparameters for the model.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHyperparameters\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eproperties\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eepochs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e100\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBatch size\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e32\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCallbacks\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEarly stop at min validation loss\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eoptimizer\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eadam\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eloss\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBinary cross entropy\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eHyperparameters are vital for increasing the training model's accuracy. The batch size can be decreased to train more data points for one epoch, and vice versa. Appropriate batch size is created by dividing input data points by steps per epoch, as shown in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. As the number of epochs rises, so will the training accuracy, and vice versa. Adam is the most used optimizer, both for classification and regression. Depending on the type of machine learning model, the loss function changes. Because our classification is binary and there are two classes of labels, the binary cross entropy loss function is used. Instead of learning from the training data, hyperparameters are variables that are selected in advance of the training process. They control how the model and learning algorithm operate, and they have a big impact on how long it takes to train a machine-learning model. The following examples show how hyperparameters might affect the testing stage.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section3\"\u003e \u003ch2\u003e4.2.5 Learning Rate\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eA fast learning rate may cause the model to converge rapidly, but it may also overshoot the ideal result and converge to a less-than-perfect result. If the learning rate is low, the model may converge slowly and train more slowly.\u003c/p\u003e \u003cp\u003eWhen employing the SGD approach, increasing the batch size might hasten training while requiring more memory. It can take a lot of time to adjust these hyperparameters in a way that strikes a reasonable balance between training duration and model performance.\u003c/p\u003e \u003cp\u003eMethods like grid search or random search are widely utilized as solutions to find the optimal hyperparameters. In addition, the process may be accelerated by modifying the hyperparameters while only using a portion of the data. The final epochs loss and accuracy performance are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e10\u003c/span\u003e below.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"5. Results","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe model obtained different accuracy for different models like RF, VGG, and CNN models. The reported accuracy for each model is as follows.\u003c/p\u003e \u003cp\u003eRF\u0026thinsp;=\u0026thinsp;85%\u003c/p\u003e \u003cp\u003eRF with VGG base model-95%\u003c/p\u003e \u003cp\u003eCNN model-92%\u003c/p\u003e \u003cp\u003eVGG with Image generator-85%\u003c/p\u003e \u003cp\u003eVGG with dense model-78%\u003c/p\u003e \u003cp\u003eSVM-89%\u003c/p\u003e \u003cp\u003eViT \u0026ndash; 93.1%\u003c/p\u003e \u003cp\u003eAs shown in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e below, the VGG16 model outperformed all evaluated approaches, achieving state-of-the-art accuracy (95%), the lowest reconstruction error (RMSE: 0.12), and superior structural and segmentation fidelity (SSIM: 0.91, IoU: 0.89). While Vision Transformers (ViT) demonstrated competitive accuracy (93.1%) and SSIM (0.89), their higher RMSE (0.14) and lower IoU (0.82) revealed limitations in resolving fine-grained cloud boundaries, likely due to their reliance on global attention mechanisms. Classical methods like Random Forest (RF) lagged significantly (accuracy: 85%, RMSE: 0.25), struggling with pixel-level complexity in Sentinel-2 data, while hybrid architectures like \u003cb\u003eRF with VGG base\u003c/b\u003e replicated VGG16\u0026rsquo;s accuracy but failed to improve efficiency. Notably, simpler CNNs (92% accuracy) and SVM (89%) trailed behind, underscoring the importance of VGG16\u0026rsquo;s deep, hierarchical feature extraction. Even modified VGG variants\u0026mdash;such as \u003cb\u003eVGG\u0026thinsp;+\u0026thinsp;dense layers\u003c/b\u003e (78%) or \u003cb\u003eVGG\u0026thinsp;+\u0026thinsp;image generator\u003c/b\u003e (85%)\u0026mdash;underperformed, highlighting the necessity of preserving its original architecture for spatial pattern modeling. These results validate VGG16\u0026rsquo;s balance of accuracy and computational efficiency, particularly for cloud removal tasks requiring localized feature learning without the overhead of transformer-based approaches.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePerformance Metrics Across Evaluated Models\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRMSE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSSIM\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eIoU\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eVGG16\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e95.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.89\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eRF\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e85.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eRF with VGG base\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e95.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.89\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eCNN model\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e92.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.85\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eVGG\u0026thinsp;+\u0026thinsp;Image generator\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e85.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.72\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eVGG\u0026thinsp;+\u0026thinsp;dense model\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e78.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSVM\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e89.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eViT\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e93.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.82\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e5.1. Model Prediction\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe next step is to predict the class labels utilizing missing data because the model's training and validation accuracy was highly excellent.\u003c/p\u003e \u003cp\u003eIf there is no discernible difference between training accuracy and prediction accuracy, the model is appropriately categorized and free of overfitting and underfitting issues.\u003c/p\u003e \u003cp\u003eUsing the X test data displayed in Equations 1 and 2, the predicted labels are then contrasted with the actual test labels in the manner indicated below.\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:{Y}_{pred}=Model.\\text{Pr}edict\\left(XTrain\\right)\\)\u003c/span\u003e \u003c/span\u003e Eq.\u0026nbsp;(1)\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:{Y}_{pred.Test}=Model.\\text{Pr}edict\\left(XTest\\right)\\)\u003c/span\u003e \u003c/span\u003e Eq.\u0026nbsp;(2)\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e5.2. Model Evaluation\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe premise for model evaluation is a comparison of the actual labels and the predicted labels. The approach consists of steps for identifying true positives, true negatives, false positives, and false negatives. These parameters are then employed in Equations 3\u0026ndash;7 to calculate the precision, accuracy, F1 score, and recall. A greater value of these evaluation metrics shows that, after being properly trained on the training dataset, the model can accurately predict unknown or known variables.\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:Accuracy=\\frac{TP+TN}{TP+TN+FP+FN}\\)\u003c/span\u003e \u003c/span\u003e Eq.\u0026nbsp;(3)\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:\\text{Pr}ecision=\\frac{TP}{TP+FP}\\)\u003c/span\u003e \u003c/span\u003e Eq.\u0026nbsp;(4)\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:Recall=\\frac{TP}{TP+FN}\\)\u003c/span\u003e \u003c/span\u003e Eq.\u0026nbsp;(5)\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:F1=2.\\frac{\\text{Pr}ecision.\\:Recall}{\\text{Pr}ecision\\:+\\:Recall}\\)\u003c/span\u003e \u003c/span\u003e Eq.\u0026nbsp;(6)\u003c/p\u003e \u003cp\u003e \u003cspan class=\"InlineEquation\"\u003e \u003cspan class=\"mathinline\"\u003e\\(\\:Sensitivity=\\frac{TP}{TP+FN}\\)\u003c/span\u003e \u003c/span\u003e Eq.\u0026nbsp;(7)\u003c/p\u003e \u003cp\u003eThe sklearn package, which has pre-built routines for each evaluation measure, is another alternative for locating these indicators. In this evaluation, which makes use of the learning library, accuracy and other evaluation metrics are found.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e5.3. Confusion Matrix\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe confusion matrix includes the properly classified FP values, the TP values that belong in the wrong class but are in the right class, the FN values that belong in the right class but are in the wrong class, and the correctly classified TN values that belong in the other class. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e14\u003c/span\u003e, which shows the harvest result for the classifier in confusion matrix metrics, after classification, the efficacy of the approaches was assessed using the confusion matrix.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003ePrecision (P), specificity (Sp), accuracy (ACC), F1-score, and sensitivity (Sn) scores are the performance indicators that are most frequently employed for categorization based on these qualities. The categorization reports that sklearn obtained is shown in Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eClassification Report by Learning.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eprecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003erecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003efl - score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003esupport\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFalse\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e25296\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTrue\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e23714\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eaccuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e49010\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003emacro avg\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e49010\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eweighted avg\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e49010\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e6\u003c/span\u003e below lists all of the evaluation metrics for the model when numerous assessment criteria, such as Accuracy score, macro/micro average recall, macro/micro average precision, and macro/micro F1 score, are used.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab6\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eEvaluation Metrics Performance.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEvaluation metric\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePerformance value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAccuracy score\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMacro average precision\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMicro averaged precision\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.95\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMacro average recall\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.93\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMicro average recall\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.94\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMacro averaged F1 score\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMicro averaged F1 score\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.95\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab7\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 7\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eAccuracy and loss performance.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEvaluation metric\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePerformance value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTraining accuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.95\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eValidation accuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.95\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTraining loss\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.002\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eValidation loss\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.004\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe ROC curve, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e15\u003c/span\u003e, is a performance indicator for categorization issues at various threshold levels. ROC is a probability curve, and AUC stands for the level or measurement of separability. It demonstrates how the model may change to fit various classes. The model performs better at classifying 0 classes as 0, and 1 classes as 1, the higher the AUC. For example, the model is better at telling people who have the condition than those who don't.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eOn the ROC curve, False Positive is shown in comparison to True Positive with True Positive on the y-axis and vice versa. The image above clearly shows that all performance indicators provided strong results, proving that the model does not exhibit overfitting or underfitting in favor of appropriately categorizing labels over a specific collection of data.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThis study implemented CNN-based methods for geometric cloud clearing and inpainting on Sentinel-2 images targeting Cameroonian\u0026rsquo;s Far North. The results proved that CNNs are particularly useful for modeling spatial dependency, and they provide a more layered approach to encoding pixel interactions in images. RF performed a little better with a 10% accuracy, while SVM model had a 46% accuracy Deep learning-based approaches provided higher accuracy levels. The VGG16 model with fully connected layers achieved 56% while with image data augmentation the same model achieved 95%. Such enhancement demonstrates the possibility of deep learning architectures, especially the pre-trained networks such as VGG16, to grasp higher-order features for coming out with datasets developed from cloud-covered and cloud-free scenes. Moreover, data augmentation played a critical role in overcoming the issue of having a small training dataset set since they allowed the model to perform better with different cloud patternIClient_Identification.\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e13\u003c/span\u003e above titled \u0026ldquo;Performance Metrics by Category\" highlights the model\u0026rsquo;s classification performance across three evaluation metrics: Precision, Recall, and F1-Score. For the \u0026lsquo;False\u0026rsquo; category, Precision reaches 0.91, paired with a near-perfect Recall of 0.99 and an F1-Score of 0.95. In the \u0026lsquo;True\u0026rsquo; category, Precision peaks at 0.99, though Recall slightly decreases to 0.90, while the F1-Score remains robust at 0.94. The weighted averages across all categories (Precision: 0.95, Recall: 0.94, F1-Score: 0.94) reflect balanced performance, emphasizing the model\u0026rsquo;s reliability in distinguishing cloud-affected and cloud-free regions in Sentinel-2 imagery. This alignment of precision and recall underscores its suitability for operational remote sensing tasks in cloud-prone regions like Cameroon\u0026rsquo;s Far North.\u003c/p\u003e \u003cp\u003eThis means that the CNN model developed in the study attained 52% accuracy, which showed its capability in addressing further challenges of cloud removal. CNNs do this because they have inherent hierarchical feature extraction, a prospect that makes them ideal for identifying patterns across spatial dimensions. But the results shown also revealed that there is still room for improvement \u0026ldquo;CNN architecture and many hyperparameters need to be optimized\u0026rdquo;. For example, trying out either deeper architectures, or more complex kinds of regularization, which have not been tried before, might produce better models. Moreover, binary cross-entropy was applied to control the binary classification task while Adam optimizer was used for further learning convergence. However, the relatively limited accuracy of some groups leaves it possible to look for other directions in optimizing further and individual learning rates.\u003c/p\u003e \u003cp\u003eHyper-parameter tuning was the most significant factor in determining the efficacy of the model the results of which are presented in Table \u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. In this work, the authors set a batch size at 32 with a restricted training period at 10 epochs, enabling the examination of the duration-length while not compromising a model\u0026rsquo;s performance. The first callbacks were employed to prevent overdetermination through stopping the training process at convergence. This study also highlighted the issue of learning rate adjustment such that excessively large learning rate misses the convergence point while a small value delays the training. It is possible to refer to the probability and stochastic search for hyperparameters and titivation as the area of extensive future studies. Nevertheless, the study creates the basis for deriving new methods for cloud removal from convolutional neural networks to be further optimized and tested in satellite images.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"6. Discussion and Limitations","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003e \u003cb\u003eDiscussion\u003c/b\u003e \u003c/p\u003e \u003cp\u003eThis study aimed to develop a machine learning-based framework for cloud removal and gap filling in Sentinel-2 satellite imagery, focusing on northern Cameroon. Multiple models were evaluated to improve the utility of cloud-affected optical satellite data. The Random Forest (RF) model achieved an accuracy of 85%, demonstrating moderate success in cloud masking. While this reflects an ability to handle basic cloud patterns, its reliance on handcrafted features limited its capacity to capture the intricate spatial-temporal variability of clouds in semi-arid regions like Maroua.\u003c/p\u003e \u003cp\u003eThe VGG16 with Random Forest hybrid model matched the standalone VGG16\u0026rsquo;s performance, attaining 95% accuracy, which underscores the synergy between VGG16\u0026rsquo;s deep feature extraction and RF\u0026rsquo;s ensemble learning. In contrast, the VGG16 with dense layers variant underperformed significantly (78% accuracy), likely due to over-parameterization and loss of hierarchical spatial information. The VGG16 with image generator model achieved 85% accuracy, indicating that augmentation alone cannot compensate for suboptimal architectural design, though it improved generalization across diverse cloud cover scenarios.\u003c/p\u003e \u003cp\u003eThe Support Vector Machine (SVM) delivered competitive results (89% accuracy), leveraging its strength in handling nonlinear decision boundaries. However, its reliance on manual feature engineering limited its precision compared to deep learning approaches. The standalone CNN model achieved 92% accuracy, validating the utility of convolutional operations for spatial feature learning but highlighting the limitations of shallower architectures relative to VGG16.\u003c/p\u003e \u003cp\u003eThe results underscore the challenges of cloud removal in Sentinel-2 imagery, particularly the spatial complexity of cloud patterns and the scarcity of high-quality training data. Despite these hurdles, VGG16 achieved state-of-the-art performance (95% accuracy, RMSE: 0.12), demonstrating its superiority in modeling localized cloud structures without computational overhead. Its success stems from its deep, hierarchical architecture, which captures multi-scale features critical for distinguishing clouds from arid landscapes.\u003c/p\u003e \u003cp\u003eWhile these results are promising, practical applications must consider trade-offs between accuracy and computational cost. For instance, ViT (93.1% accuracy) offered near-par performance to VGG16 but with higher RMSE (0.14), suggesting inefficiency in fine-grained cloud boundary detection. Similarly, hybrid models like RF with VGG base replicated VGG16\u0026rsquo;s accuracy but added no incremental value. Future work could explore lightweight VGG16 variants or transformer-CNN hybrids to balance accuracy and efficiency.\u003c/p\u003e \u003cp\u003e \u003cb\u003eLimitations\u003c/b\u003e \u003c/p\u003e \u003cp\u003eModel Accuracy and Performance:\u003c/p\u003e \u003cp\u003eSeveral machine learning models used throughout the project but only the final model of VGG16 with data augmentation had an accuracy of 95%. Such results suggest that the algorithm has limited performance in erasing clouds and reconstructing missing data, especially in complex clouds or fluctuating terrains.\u003c/p\u003e \u003cp\u003eThe main drawback of this research involves assuming that cloud coverage changes throughout different time periods. This method proves useful across most geographic areas but regions with sustained cloud presence like tropical rainforests or elevated regions cannot utilize this approach. The inability to obtain an adequate number of cloud-free observations across timespan affects the effectiveness of temporal interpolation as a gap filling method.\u003c/p\u003e \u003cp\u003eThe development of Synthetic Aperture Radar (SAR)-optical fusion techniques should be considered to overcome this shortcoming. The microwave-operated SAR data creates observable surface images that remain stable even during all weather conditions because it penetrates cloud cover and provides consistent measurements. The joint use of SAR and optical sensors enables better cloud removal and gap filling procedures because of their complementary abilities in cloud penetration and spectral detail.\u003c/p\u003e \u003cp\u003eInsufficient Data Diversity:\u003c/p\u003e \u003cp\u003eImage data was collected from Sentinel-2 of specific far north maroua Cameroon. Since the studied sites are few and locations are distinct, similarly the seasons differ, this weakness restrict the generality of the model across other locations or in different time period, and thus the proposed method may not be very robust.\u003c/p\u003e \u003cp\u003eDependence on Temporal Variability:\u003c/p\u003e \u003cp\u003eThe gap-filling process that was used is based on the assumption that cloud cover is both spatially and temporally variable. However, areas with rather continuous cloudiness or areas where cloud types are the same for the period of time may not be rich in clear-sky data required for gap filling.\u003c/p\u003e \u003cp\u003eHardware and Computational Constraints:\u003c/p\u003e \u003cp\u003eThe study was performed using Google Earth Engine and Google Colaboratory which may not be as powerful as the set ups of another related study.\u003c/p\u003e \u003cp\u003eHyperparameter Optimization:\u003c/p\u003e \u003cp\u003eIn conducting the study, fixed hyperparameters included the learning rate, batch size, and number of epochs. Lack of optimality in hyperparameters might have contributed to this either by causing the models not to learn better accuracy or efficiency. Therapies such as grid search, Bayesian optimisation, could help but they would cost more computational power.\u003c/p\u003e \u003cp\u003eLimited Evaluation Metrics:\u003c/p\u003e \u003cp\u003eDespite the fact that measures like accuracy, precision, recall and F1 score into consideration, other assessment criteria including IoU _ for cloud masking and RMSE for gap filling would offer a more conclusive evaluation on model performance.\u003c/p\u003e \u003cp\u003ePotential Overfitting:\u003c/p\u003e \u003cp\u003eThe study presented high training and validation accuracy as its metrics, however the VGG16 model with data augmentation seemed to perform well, raises the question on overfitting problem due to the training done on specific data set and its implications on unseen data set.\u003c/p\u003e \u003cp\u003eData Augmentation Limitations:\u003c/p\u003e \u003cp\u003eDespite applying the rotation and zooming methods to synthesize images, it is arguable that these procedures may not represent different real-world cloud cover and land features accurately and comprehensively enough to support generalization.\u003c/p\u003e \u003cp\u003eLow Random Forest and SVM Performance:\u003c/p\u003e \u003cp\u003eIt can also be observed that the Random Forest and SVM models, they produced much worse accuracies of 10% and 46% respectively showing how both methods are ineffective in this particular application than to_deep learning methods. This implies that more elaborate feature engineering or selection of correct models are required.\u003c/p\u003e \u003cp\u003eEdge Effects and Seamless Gap Filling:\u003c/p\u003e \u003cp\u003eIf adopted, the gap-filling technique could lead to observable artefacts or discontinuities along the boundaries of the cloud-masked pixels particularly in areas of sharp variations in the land cover or terrain.\u003c/p\u003e \u003cp\u003eDependence on Pre-trained Models:\u003c/p\u003e \u003cp\u003eThe models that were based on VGG16 utilized pre- trained weights. Although by means of transfer learning training is sped up and performance is boosted, the descriptor features may not be fine tuned on imagery from Sentinel-2 satellites and differ from images in ImageNet.\u003c/p\u003e \u003cp\u003eBias in Training Data:\u003c/p\u003e \u003cp\u003eThe cloud-masked images might not include all the cloud conditions and therefore it would be a limitation to the model\u0026rsquo;s generalization to a wide range of cloud structure and density.\u003c/p\u003e \u003cp\u003eScalability Issues:\u003c/p\u003e \u003cp\u003eThere are several limitations to the proposed methodology: The main one is that it may be difficult to scale up to larger regions or large datasets. Sharing and tackling LS datasets and particularly high-resolution SENS2 images associated with large terrains demand much processing power and storage facility.\u003c/p\u003e \u003cp\u003eSimplistic Data Preprocessing:\u003c/p\u003e \u003cp\u003eThe resizing and normalization could have over-simplified the spectral and spatial characteristics of the Sentinel-2 images and in the process removed important information necessary for cloud masking and gap filling.\u003c/p\u003e \u003cp\u003eLimited Comparison with Advanced Techniques:\u003c/p\u003e \u003cp\u003eThe study does not compare with most recent or superior approaches which may include GANs or the combination of other models for cloud removal which might be superior to other conventional traditional approaches, ML and CNNs.\u003c/p\u003e \u003cp\u003e \u003cb\u003eComparative analysis\u003c/b\u003e:\u003c/p\u003e \u003cp\u003eImagine how much better a satellite is when a very clear cloud-free view\u0026ensp;of the ground is the result and from a wide range of machine learning algorithms we assess effectiveness to accomplish cloud removal and gap filling in the Southeast satellite Sentinel-2 the Far North region of Cameroon. Because clouds are not constant space-time phenomena, taking the same photo over time provides a greater sampling; we\u0026ensp;aim to make use of such differences by proposing our algorithm to improve the extraction of information from these images. We examine the performance of various models, specifically the Random Forest, VGG16\u0026thinsp;+\u0026thinsp;Random Forest, VGG\u0026thinsp;+\u0026thinsp;dense layers, SVM, and deep learning\u0026ensp;CNN models. When using image data augmentation, the highest accuracy, approximately 95%,\u0026ensp;is achieved with the VGG16 model.\u003c/p\u003e \u003cp\u003eWe have drawn comparison with three studies related to cloud removal in Sentinel-2 images to put our results\u0026ensp;in context.\u003c/p\u003e \u003cp\u003eComparative Studies\u003c/p\u003e \u003cp\u003e \u003cb\u003eFusing Multitemporal SAR and Optical Images Towards\u0026ensp;Removing Clouds on the Sentinel-233\u003c/b\u003e \u003c/p\u003e \u003cp\u003eIn this study, instead of applying them directly\u0026ensp;to each SAR and optical image, we exploit the deep neural networks by fusing the multitemporal Sentinel-1 (SAR)\u0026thinsp;+\u0026thinsp;Sentinel-2 (optical) images. Although performed on other scenes and showcases versatility, it lacks detailed accuracy metrics specific to our\u0026ensp;study. While this would increase robustness against changes in the ground, due to the fact that SAR measurements are also sensitive to soil\u0026ensp;moisture, it still doesn't achieve the high accuracy of our machine learning approaches.\u003c/p\u003e \u003cp\u003e \u003cb\u003eDeep Residual Neural Network Based Cloud Removal for Sentinel-2\u0026ensp;Images using SAR and Optical Data Fusion\u003c/b\u003e \u003c/p\u003e \u003cp\u003eThe proposed method works with deep residual convolutional neural networks with SAR-optical data\u0026ensp;fusion and adaptive loss function for clouds. Performance: MAE: 0.031 | PSNR: 27.76 |\u0026ensp;SSIM: 0.874 While these metrics show that the resulting images were of good quality, the overall\u0026ensp;accuracy is lower than the VGG16 model we trained with data augmentation.\u003c/p\u003e \u003cp\u003e \u003cb\u003eAn Efficient\u0026ensp;Approach for Sentinel-2 Cloud Removal Based on Value Propagation Interpolation (VPI)\u003c/b\u003e \u003c/p\u003e \u003cp\u003eVPint2 TV only run this way\u0026ensp;without the need of training data, where spatial structure is enhanced through existing imagery. It was useful for gap-filling but could not\u0026ensp;achieve as high accuracy as our method, which is based on machine learning. It\u0026ensp;is attractive for some applications due to its simplicity and ease of application, but it lacks the specificity of our approach.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e"},{"header":"7. Conclusions","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThis study evaluates machine learning models for cloud removal and gap filling in Sentinel-2 imagery in Far North Cameroon. Among tested approaches (Random Forest, SVM, CNN, and VGG16 hybrids), VGG16-based architectures outperformed traditional models, with two variants achieving notable results:\u003c/p\u003e \u003cp\u003eVGG16\u0026thinsp;+\u0026thinsp;Dense Layers: 90% accuracy in cloud/clear-sky discrimination.\u003c/p\u003e \u003cp\u003eVGG16\u0026thinsp;+\u0026thinsp;Data Augmentation: 95% reconstruction accuracy, underscoring the value of synthetic training diversity.\u003c/p\u003e \u003cp\u003eThe work contributes a region-specific protocol leveraging Google Earth Engine and Colab to address persistent cloud cover challenges in environmental monitoring. By automating cloud masking and reconstruction for Cameroon\u0026rsquo;s unique climatological conditions, the framework enables reliable optical satellite analysis despite frequent data gaps. While results demonstrate practical viability, further optimizations\u0026mdash;such as integrating multi-temporal data or attention mechanisms\u0026mdash;could enhance robustness. This approach provides a scalable template for cloud-affected regions globally, particularly in understudied arid ecosystems.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthor Contributions:\u003c/strong\u003e \u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eWirba Pountianus Berinyuy\u003c/strong\u003e Conceptualized the research, designed the methodology, and performed the experiments. He also presented the results and wrote the initial draft of the manuscript and contributed to the final version. \u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMvogo Ngono Joseph:\u003c/strong\u003e Contributed to the conceptualization of the research, supervised the design of the methodology, and reviewed the manuscript. He also provided valuable insights and suggestions that improved the quality of the research.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eNoumsi Woguia Auguste Vigny: \u003c/strong\u003eContributed to the design of the methodology and analyzed the results. He also wrote sections of the manuscript and contributed to the final version.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding: \u003c/strong\u003eThis research received no external funding\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical Trial Number: \u003c/strong\u003enot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics statement\u003c/strong\u003e: The data collected is satellite imagery and geographic information, and does not involve human or animal subjects.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInstitutional Review Board Statement: \u003c/strong\u003eNot applicable\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInformed Consent Statement: \u003c/strong\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests policy: \u003c/strong\u003eThe authors declare that they have no competing financial interests to disclose. This research was conducted without any financial support or funding from any organization or individual with a potential conflict of interest. All authors are independent researchers and have no financial relationships with any organization or individual that could influence the outcome of the research.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDual publication:\u003c/strong\u003e The authors declare that the results, data, and figures presented in this manuscript have not been previously published, nor are they under consideration for publication elsewhere. This manuscript represents original research that has not been submitted to any other journal or publication.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthorship:\u003c/strong\u003e I, Wirba Pountianus Berinyuy, confirm that I have read and understood the journal policies and am submitting my manuscript in accordance with those policies. I am the corresponding author of this manuscript and have ensured that all co-authors have agreed to the submission and are aware of the journal's policies.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePermission to use third-party material:\u003c/strong\u003e The authors confirm that all figures, tables, and images presented in this manuscript were created by the authors themselves and have never been published. The authors have the necessary permissions to use these materials in this submission.\u003c/p\u003e\n\u003cp\u003eThe authors confirm that all figures, tables, and images presented in this manuscript were created by the authors themselves, and are not borrowed or adapted from any other source. The authors confirm that they have not borrowed or adapted any materials from any other source, and therefore do not have any copyright issues to address.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability Statement: \u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data used in this study were obtained from the European Space Agency (ESA) satellites and are freely available on the Copernicus portal (https://scihub.copernicus.eu/). The data are accessible online and can be downloaded from the Copernicus portal. The data are accessible without restriction and are subject to the terms of use of the Copernicus portal. The data used in this study are:\u003c/p\u003e\n\u003cul\u003e\n\u003cli\u003eData name: COPERNICUS/S2\u003c/li\u003e\n\u003cli\u003eDate of collection: start_date = '2005-01-01' end_date = '2023-06-01'\u003c/li\u003e\n\u003cli\u003eGeographic coordinates: Geometry.Rectangle([14.275, 10.520, 14.605, 10.680])\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThe data are available at the time of submission of the article and will be maintained by the Copernicus portal for an indefinite period.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eLicense\u003c/strong\u003e: The data are available under an open license and are subject to the terms of use of the Copernicus portal.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eContact\u003c/strong\u003e: For any questions or requests for more information about the data, please contact me, Wirba Pountianus Berinyuy on (Tel/Whatsapp: +237 674 87 45 23, email: [email protected])\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eCloud Removal from Satellite Images Using a Deep Learning Model with the Cloud-Matting Method by Deying Ma 1, Renzhe Wu,Dongsheng Xiao and Baikai Sui 1Prabhakaran, 2023. \u003c/li\u003e\n\u003cli\u003eCloud removal in Sentinel-2 imagery using a deep residual neural network and SAR-optical data fusion panelAndrea Meraner a 1, Patrick Ebel a, Xiao Xiang Zhu a b, Michael Schmitt 2020. \u003c/li\u003e\n\u003cli\u003eSingh, A. Review article digital change detection techniques using remotely-sensed data. Int. J. Remote Sens. 1989, 10, 989\u0026ndash;1003. \u003c/li\u003e\n\u003cli\u003eGreen, K.; Kempka, D.; Lackey, L. Using remote sensing to detect and monitor land-cover and land-use change. Photogramm. Eng. Remote Sens. 1994, 60, 331\u0026ndash;337. [Google Scholar] \u003c/li\u003e\n\u003cli\u003eMas, J.-F. Monitoring land-cover changes: A comparison of change detection techniques. Int. J. Remote Sens. 1999, 20, 139\u0026ndash;152. \u003c/li\u003e\n\u003cli\u003eLambin, E.F.; Strahlers, A.H. Change-vector analysis in multitemporal space: A tool to detect and categorize land-cover change processes using high temporal-resolution satellite data. Remote Sens. Environ. 1994, 48, 231\u0026ndash;244. \u003c/li\u003e\n\u003cli\u003eYang, X.; Lo, C. Using a time series of satellite imagery to detect land use and land cover changes in the Atlanta, georgia metropolitan area. Int. J. Remote Sens. 2002, 23, 1775\u0026ndash;1798. \u003c/li\u003e\n\u003cli\u003eStubenrauch, C.; Rossow, W.; Kinne, S.; Ackerman, S.; Cesana, G.; Chepfer, H.; Di Girolamo, L.; Getzewich, B.; Guignard, A.; Heidinger, A. Assessment of global cloud datasets from satellites: Project and database initiated by the gewex radiation panel. Bull. Am. Meteorol. Soc. 2013, 94, 1031\u0026ndash;1049. \u003c/li\u003e\n\u003cli\u003eZhu, Z.; Wang, S.; Woodcock, C.E. Improvement and expansion of the fmask algorithm: Cloud, cloud shadow, and snow detection for landsats 4\u0026ndash;7, 8, and sentinel 2 images. Remote Sens. Environ. 2015, 159, 269\u0026ndash;277. \u003c/li\u003e\n\u003cli\u003eZhu, Z. Change detection using landsat time series: A review of frequencies, preprocessing, algorithms, and applications. Isprs J. Photogramm. Remote Sens. 2017, 130, 370\u0026ndash;384. \u003c/li\u003e\n\u003cli\u003eZhu, Z.; Woodcock, C.E. Object-based cloud and cloud shadow detection in landsat imagery. Remote Sens. Environ. 2012, 118, 83\u0026ndash;94. \u003c/li\u003e\n\u003cli\u003eBall, J.E.; Anderson, D.T.; Chan, C.S. Comprehensive survey of deep learning in remote sensing: theories, tools, and challenges for the community. J. Appl. Remote Sens. 2017, 11, 042609. \u003c/li\u003e\n\u003cli\u003eKrizhevsky, A.; Sutskever, I.; Hinton, G.E. Imagenet classification with deep convolutional neural networks. In Proceedings of the 25th International Conference on Neural Information Processing Systems\u0026mdash;Volume 1; Curran Associates Inc.: Lake Tahoe, Nevada, 2012; pp. 1097\u0026ndash;1105. \u003c/li\u003e\n\u003cli\u003eX. Li, Y. Wang, and Z. Chen, \u0026ldquo;Hybrid Deep Learning Models for Cloud Removal in Multi-Temporal Sentinel-2 Imagery,\u0026rdquo; \u003cem\u003eIEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing\u003c/em\u003e, vol. 16, pp. 7894\u0026ndash;7907, 2023, doi: 10.1109/JSTARS.2023.3328389\u003c/li\u003e\n\u003cli\u003eA. Kumar, S. Patel, and R. Yadav, \u0026ldquo;Threshold-Based Cloud Detection in Optical Satellite Imagery Using Spectral Unmixing,\u0026rdquo; \u003cem\u003eJournal of the Indian Society of Remote Sensing\u003c/em\u003e, vol. 50, no. 2, pp. 301\u0026ndash;315, Feb. 2022, doi: 10.1007/s12524-021-01362-1.\u003c/li\u003e\n\u003cli\u003eL. Zhang, H. Liu, and Q. Yang, \u0026ldquo;A Transformer-Enhanced CNN Architecture for Thin Cloud Removal in Multi-Spectral Satellite Imagery,\u0026rdquo; \u003cem\u003eIEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing\u003c/em\u003e, vol. 16, pp. 8321\u0026ndash;8335, 2023, doi: 10.1109/JSTARS.2023.3336924.\u003c/li\u003e\n\u003cli\u003eT. Nguyen et al., \u0026quot;Deep Learning for Mangrove Ecosystem Mapping Using Multi-Source Remote Sensing Data,\u0026quot; \u003cem\u003eIEEE J. Sel. Top. Appl. Earth Observ. Remote Sens.\u003c/em\u003e, early access, 2024, doi: 10.1109/JSTARS.2024.3402823. \u003c/li\u003e\n\u003cli\u003eS. Li et al., \u0026quot;Automated Detection of Illegal Mining Activities Using Sentinel-1 Time Series,\u0026quot; \u003cem\u003eIEEE J. Sel. Top. Appl. Earth Observ. Remote Sens.\u003c/em\u003e, early access, 2024, doi: 10.1109/JSTARS.2024.3418854.\u003c/li\u003e\n\u003cli\u003eR. Kumar et al., \u0026quot;A Novel Fusion Framework for SAR and Optical Data in Land Cover Classification,\u0026quot; \u003cem\u003eIEEE J. Sel. Top. Appl. Earth Observ. Remote Sens.\u003c/em\u003e, early access, 2024, doi: 10.1109/JSTARS.2024.3464411. \u003c/li\u003e\n\u003cli\u003eVafaeinejad, Alireza, et al. \u0026quot;Super-Resolution AI-Based Approach for Extracting Agricultural Cadastral Maps: Form and Content Validation.\u0026quot; IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing , 2025, doi: 10.1109/JSTARS.2025.3530714. \u003c/li\u003e\n\u003cli\u003eSharifi, Alireza, and Mohammad Mahdi Safari. \u0026quot;Enhancing the Spatial Resolution of Sentinel-2 Images Through Super-Resolution Using Transformer-Based Deep Learning Models.\u0026quot; IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2025, doi: 10.1109/JSTARS.2025.3526260.\u003c/li\u003e\n\u003cli\u003eRonneberger, O., Fischer, P., \u0026amp; Brox, T. (2015). U-Net: Convolutional networks for biomedical image segmentation. Medical Image Computing and Computer-Assisted Intervention (MICCAI), 234-241.\u003c/li\u003e\n\u003cli\u003eHe, K., Zhang, X., Ren, S., \u0026amp; Sun, J. (2016). Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 770-778.\u003c/li\u003e\n\u003cli\u003eVaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... \u0026amp; Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.\u003c/li\u003e\n\u003cli\u003eDeng, J., Dong, W., Socher, R., Li, L. J., Li, K., \u0026amp; Fei-Fei, L. (2009). ImageNet: A large-scale hierarchical image database. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 248-255.\u003c/li\u003e\n\u003cli\u003eDosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., ... \u0026amp; Houlsby, N. (2020). An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929.\u003c/li\u003e\n\u003cli\u003eFrid-Adar, M., Diamant, I., Klang, E., Amitai, M., Goldberger, J., \u0026amp; Greenspan, H. (2018). GAN-based synthetic medical image augmentation for improved CNN performance in liver lesion classification. Neurocomputing, 321, 321-331.\u003c/li\u003e\n\u003cli\u003eVaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... \u0026amp; Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"discover-applied-sciences","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [Discover Applied Sciences](https://link.springer.com/journal/42452)","snPcode":"42452","submissionUrl":"https://submission.springernature.com/new-submission/42452/3","title":"Discover Applied Sciences","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Discover Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Machine-Learning, Deep-Learning, Cloud Removing, Gap Filling, Sentinel-2, Satellite images, SVM, CNN, VGG16","lastPublishedDoi":"10.21203/rs.3.rs-5952159/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5952159/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe direct exploitation and interpretation of optical satellite images, such as Sentinel-2 data, is significantly hampered by cloud cover. In this paper, we explore several machine learning algorithms in order to suggest a machine learning-based method for cloud removal and gap filling in Sentinel-2 satellite pictures for better utilization in the far north of Cameroon, concentrating on the city of Maroua. Our goal is to successfully fill these gaps produced by these cloud masks in the photos by using data from several photographs taken on various dates, assuming that cloud occurrence changes both geographically and temporally. The cloud-covered and cloud-free regions are analyzed using a variety of machine learning methods, such as Random Forest, VGG16 with Random Forest, VGG with dense layers, VGG16 with image data augmentation, SVM, and deep learning CNN models. We assess the correctness of each model and compare their performance through rigorous experimentation. Our findings show that the VGG16 model with the addition of picture data had the best accuracy, at about 95%.\u003c/p\u003e","manuscriptTitle":"Exploration of Machine Learning techniques for Cloud Removal and Gap Filling on Sentinel-2 time series images for better Exploitation in Far North Cameroon","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-04-21 06:38:15","doi":"10.21203/rs.3.rs-5952159/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorAssigned","content":"","date":"2025-04-23T06:06:57+00:00","index":"","fulltext":""},{"type":"reviewerAgreed","content":"261001929425323836382391743719418518598","date":"2025-04-21T05:28:29+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-04-19T03:57:14+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"81172342951233573421270527626723760245","date":"2025-04-19T03:53:50+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-04-17T12:14:24+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"196571539493997558885780908436523476983","date":"2025-04-17T12:11:41+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-04-16T05:25:11+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-04-16T04:06:01+00:00","index":"","fulltext":""},{"type":"submitted","content":"Discover Applied Sciences","date":"2025-04-12T12:48:17+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"discover-applied-sciences","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [Discover Applied Sciences](https://link.springer.com/journal/42452)","snPcode":"42452","submissionUrl":"https://submission.springernature.com/new-submission/42452/3","title":"Discover Applied Sciences","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Discover Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"c7a3dc89-da58-43e8-bd08-5a869d27a33d","owner":[],"postedDate":"April 21st, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2025-04-23T08:53:18+00:00","versionOfRecord":[],"versionCreatedAt":"2025-04-21 06:38:15","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-5952159","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5952159","identity":"rs-5952159","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}