Deep Learning Based Land Use and Land Cover Classification Using Remote Sensing Imagery

Deep Learning Based Land Use and Land Cover Classification Using Remote Sensing Imagery | 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 Case Report Deep Learning Based Land Use and Land Cover Classification Using Remote Sensing Imagery Aditya Ravi, A. Ashirvad Ashirvad, Sankalpa B. R., Shivaraj Chawan, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8731446/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract In today’s world of urban development, it becomes excessively difficult to observe and monitor land use and land cover changes. Since the growth is relatively more rapid in the current age, this analysis is more crucial than ever. The current approach in place uses many techniques, such as satellite imagery, which are hard to practically apply and require strong technical skills. Common users find it difficult to apply the deep learning model outputs. This work proposes a ResNet-50-based system to acquire satellite imagery, followed by land use and land cover classification with a visualization dashboard, which also features a spatial and temporal comparison of the selected area. This makes it especially useful for common users to interpret the current usage of land and the change in land usage over time. The satellite imagery is obtained from the Sentinel-2 dataset that is acquired through the Google Earth Engine. These acquired images are passed through a custom convolutional neural network based on ResNet-50. The model is initially trained for ten classes of land usage based on the EuroSat dataset. This system also allows comparison of specific land areas over a range of time to understand the change in land usage patterns across different times. The user will be able to view a statistical summary of the selected region, and this report can be used to query an AI chatbot to understand and interpret the statistical results. This increases the usability of a niche system to a regular user who does not have an idea about the land usage patterns. This model achieves an accuracy of 95.11 percent, which makes it a reliable and consistent system that can be utilized in real-world decision making process. This can be used to decide whether a new project in one of these 10 classes can be accommodated or not based on the current land usage pattern in the given area. Image Segmentation Land Use Classification Satellite Imagery Temporal and Spatial change RestNet-50 Deep Learning Unsupervised Learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Background The human activities in a region are reflected by the land usage patterns. On one side of it, we have rapid urbanization and infrastructure development, which also leads to deforestation in some regions. On the other hand, there is a constant attempt to improve the amount of agricultural area. This impacts the environmental sustainability and other ecosystems directly. This created the need for monitoring land usage patterns to decide how efficiently the land needs to be used to ensure environmental sustainability while also ensuring development and urbanization. Satellite imagery lets us acquire data through capturing repeated images for the selected region. It also allows us to acquire data across different time periods to analyze the temporal change in land use patterns. Sentinel-2 gives us multi-spectral imagery that we can use to get more reliable results for land use and land cover patterns. This aligns with the goal of classifying the segmented land to one of the 10 classes of land usage in a region. The classification of land in a region is important to observe how efficiently the land is being used. There are many challenges in classifying land use, and one of the most fundamental challenges is the lack of clear im- ages due to fog, mist, or other natural atmospheric interventions. This could lead to misclassifications due to spectral similarity. When pixels are mixed or diffused, the predictions are generally inaccurate. The existing techniques used fixed or static rules to handle such cases. They are inefficient while classifying lands at different regions or at different times. This is where machine learning techniques can be used to learn the relationship between spectral features of satellite images and the land classes. Deep learning models uncover the hidden spectral features from the images. Unsupervised learning techniques like Convolutional Neural Network (CNN) outperform the regular supervised learning models in such task. An additional layer of transfer learning can make the generic model more suitable for the task in hand. There is no support for common users in such existing work. Geospatial platforms have complex workflows that make it difficult with complex workflows that does not support regular users. This approach supports the common users with a region-based classification along with visualization and statistical summary along with an AI chatbot that allows users to interpret the results of the complex machine learning model. This system contributes by maintaining the reliability and accuracy while making it conceptually simple and straightforward to help users without domain knowledge. The most important beneficiaries of this work include decision making authorities that decide the feasibility of a new project that could either contribute positively or negatively to the environment in the name of urbanization. This is an end-to-end platform that acquires images of a selected region, segments the land based on pixels and classifies each segment into one of the classes thereby determining the overall land usage pattern of the area. 1. Introduction Land Use and Land Cover analysis is a continuously researched area over the last few decades. The initial studies experimented with regular image processing. This was followed by applying supervised machine learning techniques which also did not efficiently do the job. The initial literature work shows the comparison among many such direct techniques and the performance gains in all these techniques. The earlier systems did not have access to such clear satellite imagery too. Previous works used many distorted and unclear images which were a very obvious reason to their low accuracy percentages. There was no existing work that worked precisely for all the regions. Early work using the EuroSat dataset also did not provide the same amount of consistency and accuracy when applied in other regions of the world. Coarse and medium resolution satellite were predominantly used for regional and global analysis. MODIS and other such sensors were used for data acquisition. They were very accurate for large scale vegetations and agricultural land but failed in urban or pre-urban landscapes that were localized due to lot of internal changes. The pixels that were in the intersection of such lands that were mixed or diffused suffered with severe misclassifications. Remote sensing and GIS were used together for classification. Spectral changes were not considered or analyzed to predict the land usage. The resolution of the images, the natural interventions due to seasonality and the model used played key role in determining the results. The amount of preprocessing required for the machine learning model was also unclear and this largely determined the results of the model. Outputs were also not interpretable to be efficiently used for land usage planning. A gradual shift from techniques like support vector machines, random forests, decision trees to neural networks improved the model results. Such models worked better in different regions even in complex pre-urban landscapes. Still the generalization of a model across space and time remained a big challenge. 2. Literature Review Early research on land use and land cover change detection was based on pixel based classic approach and also needed post classification methods on the multi temporal satellite images. Civco et al. showed that these methods could be used to identify the land cover changes but are also bound by spectral confusion, mixed pixels and poor transferability in different time periods and regions. Studies done before show that despite the advancements in technology the main focus should be on the classification accuracy with limited attention to user friendly integrated systems that combine spatial temporal analysis, visualization and accessibility for beginners [ 1 ]. Medium resolution and coarse satellite data helped in large area monitoring but decreased the spatial detail. Zhan et al. showed land cover change detection using MODIS 250 m data, which improved identification of change at regional and global level, with limited capacity to show fine grained changes [ 2 ]. Similar to this Green et al. also showed that multi temporal Landsat images can be used for land use monitoring on long term scale, while also showing that error propagation was present between time steps and we had to manually handle them [ 3 ]. Developments which have happened recently show that efficient deployment and interpretation of complex models are important. Groq’s LPU based API’s can be used for fast execution of deep learning and language models and they also support visual analysis and detailed explanation of results [ 4 ]. Recently machine learning methods have improved classification by using non linear spectral relationships; Kasahun and Legesse showed we can also use random forest, support vector machines and artificial neural networks to perform better change detection and urban mapping compared to traditional methods [ 5 ]. Studies done recently show that land use and land cover change modeling should be applied for analysis of mainly urban areas. Gaur and Singh reviewed machine learning, statistical and hybrid land use and land cover methods, showing the role of data based approaches and also the need to integrate policy and planning methods for urban growth [ 6 ]. With the increasing volume of earth observation data, managing and analyzing large scale datasets has become difficult. Sudmanns et al. introduced the concept of "Big Earth Data" and also showed the need for web based, scalable methods to process multi source satellite datasets in a better [ 7 ]. Modern web frameworks has made it easy to develop and use interactive platforms. Modern full stack frameworks such as Next.js has made the development of interactive, scalable and high speed web applications easy even in data heavy systems [ 8 ]. Deep learning has improved remote sensing image analysis, Pelletier et al. provided a comparative study of deep learning architectures for hyperspectral image classification, identifying the existing challenges with respect to data size, transfer learning and model generalization [ 9 ]. The classical studies on land use and land cover change detection summarized the differences between modern and traditional techniques showing problems such as mixed pixels, lack of standardized land use methods and classification errors [ 10 ]. The ResNet50 based architectures have shown good capacity to handle robust and transferable features in different applications. Kavitha and Karpagam used a ResNet-50 encoder with a hybrid LSTM-GRU decoder which used a beam search for image captioning, achieving better representation and caption quality compared to the traditional encoder decoder model [ 11 ]. Rath et al. used a finetuned ResNet-50 model for brain tumor detection in MRI scans, which showed high classification accuracy and better generalization, while also showing the impact of transfer learning and residual networks in image analysis tasks [ 12 ]. ResNet50 based models have also shown strong results in application-oriented tasks such as disaster recognition, showing stable training behavior and also has improved the classification accuracy in complex environments [ 13 ]. Deep learning based convolutional neural networks pretrained on large datasets have made transfer learning more effective in various domains. The ImageNet dataset has played a very important role in training deep learning visual models, which further improved the generalization and overall performance for a variety of computer vision tasks [ 14 ]. The land use and land cover research reviews have shown the growing change from static land use mapping to impact analysis and factors assessment which are possible by advancements in remote sensing approaches [ 15 ]. Studies done recently have shown modeling land use and land cover changes to understand urban area expansion and its impact on environment. Mehra and Swain integrated artificial neural networks with cellular automata to predict the future land use patterns, while also showing the effectiveness of hybrid approach for forecasting urban growth indicating the influence of socio economic and spatial variables [ 16 ]. Agrawal et al. analyzed the spatiotemporal land use changes in highly populated rural areas of India, showing the change towards urban like patterns which are driven by the population pressure [ 17 ]. Bindajam et al. applied geostatistical methods to study urban expansion in Bengaluru, showing the need for spatial analysis for urban management [ 18 ]. Regional studies have started to make use of advanced learning models for detailed land use and land cover analysis. Mahendra et al. used vision transformers to analyze land use changes in Mysuru district, and showed improved representation of spatial dependencies compared to the traditional methods [ 19 ]. The work by Sharifi et al. used remote sensing and GIS techniques to analyze temporal land use and land cover change in Mysuru city, showing the impact of urbanization on the natural resources [ 20 ]. Some of the deep learning based semantic segmentation methods also improved the classification rate in different land classes, as shown by Deressu et al. through evaluation of UNet and DeepLab based architectures [ 21 ]. Deep learning frameworks have been used widely for land use and land cover classification and change analysis using satellite images. Siddique et al. came up with a ResNet50 based framework for multi period land use and land cover classification using Sentinel-2 data, accomplishing a high accuracy [ 22 ]. Nugroho et al. used patch-based classification using ResNet architecture to find land cover changes in urban areas, which also showed the detection of road expansion and construction activities [ 23 ]. High quality datasets such as Sen-2 LULC have improved research by allowing robust training and analysis of deep learning models for various land classes [ 24 ]. Transfer learning has improved the classification performance on the land use and land cover with limited labeled data. Alem et al. shoed that finetuned ResNet based models perform better than other models for satellite image classification, showing the impact of transfer learning in remote sensing applications [ 25 ]. Satellite image-based land use and land cover analysis is also integrated with environmental and socio-economic modeling as shown by Acuña-Alonso et al., who evaluated the land use changes and their results for the analysis and mitigation of floods using nature-based solutions [ 26 ]. Deep learning used in remote sensing also indicate the huge potential and the challenges of adopting deep learning based models, with respect to interpretability, scalability and adaptation to the domain [ 27 ]. The use of deep learning models for classifying satellite images increased significantly. Convolutional Neural Network (CNN) models were able to read the hidden features from the satellite images that contributed to analyzing the land use and land cover. Popular deep learning architectures like ResNet, U-Net, and other transformer-based models showed great potential for this task. Datasets like EuroSat became the benchmark for the analysis of land use. The challenge of limited data continued to prevail and this was countered by incorporating transfer learning techniques. Limitations appear in much of the literature. Generalization fails across seasons and geographic regions. Models are sensitive to sensor and atmospheric variability. Training and inference demand high computation. Deep learning outputs lack interpretability. Many studies focus on accuracy metrics alone. Research gaps stand out. Usability and accessibility for non-experts get little attention. Integrated platforms for classification, change analysis, and reporting are rare. Few systems turn outputs into easy-to-read information. Practical planning and decision-making need better solutions. This work draws motivation from gaps in integration and usability for planning and monitoring tasks. Systems must balance technical performance and practical use. 3. Related Work A. Conventional Approaches of Remote Sensing Earlier methods of land use and land cover classification used approaches which are statistical and pixel based. Common approaches were maximum likelihood, minimum distance and Bayesian classifiers. These methods took predefined statistical distribution of the spectral values. They performed well in homogeneous regions. For simple cases these methods worked but they didn’t work well with real world use cases. The performance of the methods was less in complex landscapes like urban regions. Spectral similarity confused roads, soil, vegetation and areas. The spatial context was skipped because of pixel wise treatment. It decreased the handling of the heterogeneous urban areas. Issues were also present in areas of changing of lighting and seasons. B. Machine Learning Methods for LULC Non-linear modeling of spectral class relationships was done through machine learning. Some algorithms like support vector machines, random forests, decision trees, and k-nearest neighbors were used. These methods showed more robustness that the usual methods. The handling of heterogeneous and high dimensional data became better. The reliance on strict distribution assumptions decreased. But they still had to use manually engineered features. Feature design was given to domain experts. Generalization in many regions and sensors stayed high. Users had to select and change the features for each new area. This made the process very slow and error prone. C. Deep Learning in Satellite Image Analysis The deep learning methods allowed feature extraction to happen automatically from raw images. Convolutional neural networks learned the hierarchical spatial features. Classification improved in complex landscapes. Architectures such as ResNet, U-Net, and transformer-based models were used. Good results showed up on standard datasets like EuroSAT. Transfer learning was useful especially in region specific tasks. The demand for handcrafted features dropped. The computation cost and complexity of the process of training became more. Even then these models were good with the details like edges and textures. They were also scaled to large size images than before. D. Limitations of Existing Platforms The deployment of the machine learning methods needs knowledge of remote sensing and programming. Scalability is improved with cloud based geospatial services. The users have to deal with workflows and technical methods. There is limited support for visualization and analysis of results. The output was not very good for planners nor the policy makers. There are very few easy ways to track changes over time in most of the systems. Few of the systems have built in tools for maps and reports. 4. Methods 4.1 System Overview The Land Use and Land Cover (LULC) system works as an end-to-end tool right from region selection to summary and analysis as shown in Fig. 1 . It collects raw images of the selected region, segments the region and classifies each segment into one of the ten classes. The main goal of the system is reliability, accuracy and consistency. Once the appropriate region has been selected, basic image processing is used to segment the area. This is followed by using the ML model based on ResNet-50 to efficiently classify and give a statistical summary and also generate a report for temporal analysis. This basic setup is made to handle both single images and time series data. 4.2 Data acquisition The developed system uses the Sentinel-2 satellite imagery as the source which is acquired from Google Earth engine. The user selects a region of interest by creating a polygon on the map interface which is provided in the user dashboard. The user also specifies a time period for the analysis period. The engine pulls data from the available cloud services or public data repositories for processing in the further steps. Then cloud cover filtering is performed to take only images that are not covered with fog or mist. This ensures that clear images are selected for processing. 4.3 Land Use Land Cover Classes The EuroSat dataset is a benchmark dataset, which gives ten different land cover classes derived from the Sentinel-2 multispectral satellite imagery. This dataset covers both urban and rural regions with multiple classes. Each class captures spectral characteristics, allowing the deep learning models discussed later to learn spectral features for classification. The dataset includes vegetation classes such as annual crop, permanent crop, pasture, herbaceous vegetation, and forest. Urban classes such as residential areas, industrial zones, and highways give the heterogeneous regions with complex spatial features and mixed or diffused pixels. Water bodies such as rivers and sea or lake has strong spectral features which can be used. The spatial features differentiate each of these classes and can be observed in the Fig. 2 . These spatial features are too subtle at times and can lead to misclassification. 4.4 Data preprocessing Initially, we perform a set of basic radiometric and atmospheric correction steps to ensure clear images. This step corrects the lighting and blurred effects on the acquired images. We then perform image tiling to generate image segments of the same size. The overlapping between these tiles avoids all sorts of mixed pixel issues. We then normalize the pixel to adjust the value between 0 and 1. Noisy pixels or mixed pixels are masked in this stage. Figure 3 shows the image resized to 224 x 224 pixels size to match the input dimensions required by the ResNet-50 model that we are using in this work. 4.5 Classification model We use a ResNet-50 based CNN model to classify each segment of the image. This model has residual blocks to accommodate deep learning. We then use transfer learning from pre-trained weights. This finetuning makes it suitable for land cover classification thereby making it more domain specific. The output for each tile is one of the ten classes. By counting the number of tiles in each class, the percentage of each land use type can be computed. 4.6 Training procedure The model is trained using the popular benchmark dataset called EuroSat which has captured images from the Europe region. This dataset is split into three parts, namely the training set, the testing set and the validation set. Optimizers like Adam Optimizers are used to update the weights with every epoch. About 40 epochs were used in the training stage of this model. The validation loss is used to identify the ideal number of epochs using a training vs validation loss graph. Figure 4 gives the number of images belonging to each of the ten classes in the EuroSat dataset. Therefore, some amount of data augmentation is required in order to make the classes have the same number of images to add up to the imbalance and this comes with the responsibility to not cause overfitting of weights. Data augmentation techniques such as random rotations and flips in both directions are applied to the input images to remove the imbalance. 4.7 Inference and prediction The region selected usually does not belong to the EuroSat dataset. The selected region is segmented and passed through the trained model. The model then predicts the class of each tile from the ten classes of the training data. Each tile gets a class probability map. Tile predictions aggregate into full spatial maps. Overlapping tiles average predictions smoothly. Post-processing steps apply thresholds or smoothing. This reduces noise in final maps. 4.8 Temporal change analysis The system supports single-period and multi-period analysis. Single-period classifies one image or time window. Multi-period handles several dates. Each time window gets independent classification. Land cover maps from different periods compare side by side. Class-wise changes compute area shifts. Transition statistics track from-to patterns like forest to urban. 4.9 Interpretation module Statistical summaries pull from classification results. Percentages show land cover distribution per class. Numerical outputs convert to readable explanations. Text describes dominant classes and changes. Clarity targets non-expert users like planners. Simple language avoids technical terms. 4.10 Output and visualization Color-coded land cover maps generate for each period. Colors match standard land cover schemes. A dashboard or interface shows maps side by side. Users zoom and pan for inspection. Results export as images for reports. Tables list statistics. Full reports combine maps and summaries. 5. System Architecture 5.1 Purpose of the architecture The architecture supports modular and scalable processing of satellite imagery as shown in Fig. 5 . User interaction stays separate from the backend computation. This setup allows extensibility for new features. Independent component updates avoid full system rebuilds. Scalability handles larger regions or more users. The system is basically designed to handle reliability and accuracy and has the potential to be scaled to accommodate large traffic and select larger areas to be computed. 5.2 Overall architectural design The design follows a client-server architecture. Frontend focuses on visualization and user controls. Backend runs compute intensive tasks like model inference. Lightweight tasks such as map display stay on the frontend. This split keeps the interface responsive. 5.3 Frontend component A web-based interface provides user interaction. Interactive map tools let users draw or select region of interest. Slider controls set temporal range for analysis. Land cover maps display with color coding. Change results show differences over time. Dashboard panels list statistics and summaries. Users can zoom, pan, and toggle layers. 5.4 Backend orchestration layer A central service acts as the orchestration layer. It receives requests from the frontend. The layer sequences preprocessing, inference, and analysis steps. Structured requests log parameters for reproducibility. Error handling routes issues back to the frontend. Workflow status updates keep users informed. 5.5 Satellite data access layer This layer interfaces with Sentinel-2 imagery sources. Cloud-based platforms or APIs fetch the data. Initial filtering removes high cloud cover images. Download queues manage multiple requests. Retrieved imagery stores temporarily. The layer passes files to preprocessing with metadata. 5.6 Preprocessing and data handling module The module handles correction, tiling, and normalization steps. Atmospheric correction adjusts for air and light effects. Tiling breaks large images into model-sized patches. Normalization scales bands consistently. Overlaps on tiles prevent boundary artifacts. Time period consistency checks run across images. Clean data goes to the inference engine. 5.7 Deep learning inference engine The engine loads the trained ResNet50-based model. Tile-level classification processes each patch. Batch processing speeds up large regions. GPU support accelerates predictions where available. Output includes class predictions and probability maps per pixel. Results assemble into full scene maps. 5.8 Temporal analysis module Outputs from multiple time windows feed into this module. Land cover maps align spatially for comparison. Change matrices count area shifts by class. Transition statistics detail from-to movements. Overall change summaries calculate totals. Data formats prepare for visualization and reports. 5.9 Interpretation and reporting module This module generates statistical summaries from raw outputs. Percentages and areas describe land cover distribution. Numerical results convert to plain text explanations. Examples note major changes like urban growth. Content fits dashboards and export formats. 5.10 Data flow and communication Modules use structured data exchange. REST APIs or message queues connect components. JSON formats define inputs and outputs clearly. Authentication secures backend calls. Logging tracks data movement end-to-end. 5.11 Link to system diagram Figure 5 shows the full system architecture. Arrows mark data flow from frontend through modules to outputs. The orchestration layer sits central. The diagram summarizes the workflow from imagery access to final reports. 6. Results and Discussion 6.1 Overview of experimental setup Experiments tested the platform using Sentinel-2 imagery from different regions. The trained ResNet50 model was used to classify land cover. Evaluation covered multiple time periods across several recent years. Selected areas included both urban zones and rural landscapes. The goal was to examine classification quality and change detection behavior. Tests ran on typical desktop hardware. 6.2 Land cover classification results Generated land cover maps covered full regions clearly. Dominant classes included urban built-up areas and forest vegetation. Crop fields appeared in agricultural regions. Maps remained visually consistent across test sites. Urban clusters concentrated near roads and cities, while vegetation zones followed rivers and hilly areas. Some bare soil appeared during dry seasons. Overall patterns matched expected land use as shown in Fig. 3 and Fig. 6 . The land cover map in Fig. 3 shows the distribution of major land use and land cover classes across Mysuru region, Karnataka, which are derived from Sentinel-2 imagery. Every grid cell shows a classified tile given by the ResNet-50 based deep learning model, with color coded classes showing the major land types such as water bodies, vegetation, built up areas, bare and open land. The vegetation regions cover the major area of Mysuru, while the built-up clusters are majorly present near the urban core, showing patterns of urbanization and infrastructures. The water bodies are shown distinctly in the northwestern part of the region, showing the models ability to capture classes even at medium spatial resolution. The grid-based output has both spatial maps and localized outputs between land classes, and it will make it easy for anyone to analyze the land use patterns for urban planning and environmental monitoring. 6.3 Interpretation of change maps The generated change maps use a basic set of colors to highlight the difference and this is used to analyze in the further steps as shown in Fig. 6 . Red marked areas converting to urban use. Blue indicated water gain or loss. Green represented unchanged vegetation. Visual patterns revealed growth directions clearly. Comparing maps side by side was quick and intuitive. Color usage remained consistent across time periods. 6.4 Visualization and output analysis Color-coded maps made land cover distribution easy to interpret. Spatial representation captured small details such as field boundaries. Dashboards allowed users to switch between time periods quickly. Summary tables listed area percentages for each class as shown in Fig. 7 . Export options supported saving maps for reports. Overall, the visual tools sped up result review. 6.5 Interpretation module behavior Generated summaries reported class areas and percentages of land cover changes are as shown in Fig. 8 . Text explanations described changes in simple terms, such as urban growth across the region. Non-expert users could understand insights without remote sensing knowledge. Examples highlighted major changes as well as stable areas. Figure 9 shows a comparison of the land cover maps for the place Mysuru belonging to Karnataka state, between 2022–2023 and 2023–2025, showing the differences in urban and vegetated areas. The side-by-side comparison enables easy analysis of land use changes and also allows for regional planning and monitoring activities. Figure 10 shows an AI based chatbot integrated with the land use and land cover analysis system which allows users to ask questions regarding the distribution of land area across different classes and time periods. Users can interact with the chatbot and ask natural language questions related class wise area division, the major land cover pattern, and the impact of urbanization between the timelines. It converts statistical analysis into simple human understandable explanations and so the system improves the interpretability for non-experts hence allowing better decision making in land management. 6.6 Comparative analysis of ResNet50 with EfficientNet-B0 The model ResNet-50 was selected as the most appropriate model for this project after a comparative analysis with another popular benchmark model called EfficietNet-B0. This was done using the same EuroSat dataset. The models were initialized with ImageNet pretrained weights and subsequently updated with EuroSat dataset during the training process. The input resolution for both the models is 224 x 224, but the number of parameters is much higher in ResNet. ResNet-50 has 25.6 million parameters, whereas EfficientNet-B0 has only 5.3 million parameters. On the other hand, EfficientNet-B0 has 237 layers, but ResNet-50 has only 50 layers, as the name suggests. EfficientNet has very few trainable parameters due to its compound scaling of width and resolution in an image. This makes it very suitable for lightweight applications in general but this limitation makes it struggle in medium resolution images with complex hidden spatial features which are not obvious. This leads to a high number of misclassifications in regions with multiple land use types like water body, vegetation, industrial area etc. ResNet-50 fundamentally uses residual learning on the other hand, which helps it capture complex spatial features. This basic architectural difference allows it to learn hidden feature representations thereby making it more suitable for classifying land use types. Table 1 gives the classification accuracy for all the ten classes of the EuroSat dataset for the comparison between ResNet-50 and EfficientNet models. ResNet-50 can be observed to perform better for most classes in this dataset. The difference in performance is more evident in vegetation and urban classes such as residential areas, industrial regions, and highways, due to the misclassification rate of EfficientNet-B0 caused by diffused pixels. Water bodies have evident spectral characteristics and hence give high accuracy in both models. Table 1 Comparison of Accuracy for each class between EfficientNet-B0 and ResNet-50 Class ResNet-50 (%) EfficientNet-B0 (%) Annual Crop 94.62 87.31 Forest 97.18 96.46 Herbaceous Vegetation 95.41 88.72 Highway 93.89 86.05 Industrial Area 95.76 85.14 Pasture 94.35 83.68 Permanent Crop 93.94 84.92 Residential Area 96.02 89.14 River 97.41 92.67 Sea / Lake 98.06 97.88 As displayed in Table 2 , ResNet-50 achieves a very high accuracy relative to the contemporary EfficientNet-50 model with an accuracy of 95.11 percentage compared to 88.42 percentage obtained by the EfficientNet model. The improvements can also be observed in precision, recall, and F1-score thereby proving the superiority of the ResNet model. Based on these observations made, ResNet-50 was selected as the primary classification model for the land use land cover classification system as it can identify all the hidden spatial features of the Sentinel-2 multi spectral imagery. Table 2 Comparative Performance Analysis of ResNet-50 and EfficienetNet-B0 Metric ResNet-50 EfficientNet Overall Accuracy (%) 95.11 88.42 Macro Precision (%) 94.87 87.96 Macro Recall (%) 94.53 87.21 Macro F1-Score (%) 94.70 87.58 Mis-classification Rate in Diffused Pixels (%) 4.8 11.6 The usability of the system was an important part of the platform. Maps which were coded with colors and dashboards made results easy to interpret while the readable summaries helped users such as planners and managers to understand key changes. This made the outputs to be used directly and improved the decision making of the user. The system can be used for urban planning, land management and also for long term planning. There were some limitations of the system. The results depend on the quality of the input images. Clouds covering the areas can restrict the image quality, and the environmental conditions also mattered. Seasonal changes also affect the areas which can change the classification in different time periods. The errors in classification were very common in similar classes such as soil, roads, and surfaces. The features which were very close were not easy to classify. These were common in medium resolution satellite images and is a challenge for classification. The workflow requirements also increased with the study area size. The regions which are comparatively larger had to be processed for longer and took more memory. The model also needs training data which is labeled. The model cannot be easily adapted to new regions and needs retraining. The results part provided easy to understand explanations for better understanding. The answers were not able to give the technical details. It can be improved further by including the use of higher resolution images, exploration of different model architectures, and better methods to improve outputs and scalability. The platform provides practical workflow for land use and land cover analysis which improves analysis and decision making. 6.7 Discussion of observations Strengths were observed in handling mixed urban–rural regions. Spatial detail resolution supported identification of small features. Since the images acquired are filtered for cloud and are preprocessed both in training and inferencing stages, the model is robust and works in good lighting conditions. The results are critical in determining sustainability of the development and the Urbanization process. This is crucial also for making decisions while approving a new project that belongs to one of the ten classes. 6.8 Limitations observed Soil and Bare field were misclassified in several instances due to the similarity of these land use types. The season determined the nature of vegetation which also potentially increased the misclassification opportunities. This system is very dependent on cloud filtering to work efficiently. Presence of cloud throughout the time period will severely affect the results of the model. The computation time highly depends on the size of the area selected. Large areas take excessively long time to determine the predictions for each tile in the large area and even lead to time out or excessive API calls in some cases. 7. Conclusion and limitations This project was developed for a single system for the complete land use and land cover analysis that uses satellite data, deep learning-based classification and visualization. The platform uses Sentinel-2 imagery using a ResNet50 based model and produces spatial maps, summaries, and outputs. The overall goal was to decrease the workflow while keeping the analysis useful and practical. The system showed good land cover classification in both urban and rural areas. It supports single period mapping and comparison across multiple time periods. Change detection showed major changes such as urban expansion into vegetated areas. The in-depth analysis allowed land cover results to be shown over time, along with the visual maps showing shifts. These support continuous monitoring of land areas. Declarations Data Availability Statement The dataset used in this study is publicly available. The EuroSAT dataset is available on Kaggle at: https://www.kaggle. com/datasets/apollo2506/eurosat-dataset Ethical Approval Not applicable. Clinical Trial Number Not applicable. Consent to Participate Not applicable. Consent for Publication Not applicable. Competing interests The authors declare that they have no competing interests. Funding No funding was received for this research. Author’s contributions Aditya Ravi, A. Ashirvad, Sankalpa B. R., and Shivaraj Chawan were involved in conceptualization, dataset collection and organization, methodology preparation, manuscript writing and writing the original draft. Rajashree Shettar was involved in the conceptualization, validation, supervision, along with reviewing and editing. Acknowledgments The authors thank RV College of Engineering, Bengaluru, for providing academic support and computational resources that facilitated this work. References Civco DL, Hurd JD, Wilson EH, Arnold MP, Prisloe S. A comparison of land use and land cover change detection methods. In Proceedings of the ASPRS–ACSM Annual Conference , vol. 21 (2002). Zhan X, et al. Detection of land cover changes using MODIS 250 m data. Remote Sens Environ. 2002;83:336–50. https://doi.org/10.1016/S0034-4257(02)00024-2 . Green K, Kempka D, Lackey L. Using remote sensing to detect and monitor land-cover and land-use change. Photogrammetric Eng Remote Sens. 1994;60:331–7. GROQ Inc. Groq LPU inference APIs and LLaMA 3 integration. Developer Doc (2025). https://docs.groq.com/ Kasahun M, Legesse A. Machine learning for urban land use/cover mapping: comparison of artificial neural network, random forest and support vector machine. Heliyon. 2024;10:e39146. https://doi.org/10.1016/j.heliyon.2024.e39146 . Gaur S, Singh R. A comprehensive review on land use and land cover (LULC) change modeling for urban development: current status and future prospects. Sustainability. 2023;15:903. https://doi.org/10.3390/su15020903 . Sudmanns M, Tiede D, Lang S, Augustin S. Big Earth data: disruptive changes in Earth observation data management and analysis? Int J Digit Earth. 2020;13:832–50. https://doi.org/10.1080/17538947.2019.1585976 . Vercel. Next.js documentation: the React framework for production (2025). https://nextjs.org/docs Pelletier C, Ienco G, Ban Y, Begué A. Deep learning for the classification of hyperspectral images: a comparative review. IEEE Geoscience Remote Sens Magazine. 2017;5:29–56. https://doi.org/10.1109/MGRS.2017.2762345 . Singh A, Kumar V. Change detection in land use and land cover using remote sensing data: a review. Int J Comput Appl. 2017;168:8–15. https://doi.org/10.5120/ijca2017914563 . Kavitha PV, Karpagam V. Image captioning deep learning model using ResNet50 encoder and hybrid LSTM–GRU decoder optimized with beam search. J Indian Inst Sci. 2025;105:394–410. https://doi.org/10.1080/00051144.2025.2485695 . Rath A, Mishra BSP, Bagal DK. ResNet50-based deep learning model for accurate brain tumor detection in MRI scans. Neurosci Res Notes. 2024;6:100104. https://doi.org/10.1016/j.nexres.2024.100104 . Wen L, Xiao Z, Xu X, Liu B. Disaster recognition and classification based on improved ResNet-50 neural network. Appl Sci. 2025;15:5143. https://doi.org/10.3390/app15095143 . Deng J et al. ImageNet: a large-scale hierarchical image database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 248–255 (2009). https://doi.org/10.1109/CVPR.2009.5206848 Chang Y, Hou K, Li X, Zhang Y, Chen P. Review of land use and land cover change research progress. IOP Conference Series: Earth and Environmental Science 113, 012087 (2018). https://doi.org/10.1088/1755-1315/113/1/012087 Mehra N, Swain JB. Assessment of land use land cover change and its effects using artificial neural network-based cellular automata. J Eng Appl Sci. 2024;71:70. Agrawal S, Welegedara N, Parida D. Spatiotemporal land use changes in remote rural regions of In- dia between 2000 and 2020. Environ Plann B: Urban Analytics City Sci. 2024;33:1–18. https://doi.org/10.1177/10185291241236307 . Bindajam AA, Mallick J, Alshayeb MJ, Poddar S. Understanding urbanization and its expansion process in Bengaluru through geostatistical models for landscape management. Adv Space Res. 2025;75:11–8. https://doi.org/10.1016/j.asr.2025.11.018 . Mahendra HN, Pushpalatha V, Mallikarjunaswamy S, Rama Subramoniam S, Rao AS, Shekar NC. Analyzing land use land cover changes in Mysuru taluk, Karnataka state, India using vision transformers. Adv Comput Appl Geospatial Stud. 2025;3:100308. https://doi.org/10.1016/j.acags.2025.100308 . Sharifi V, Shivanna S, Manjunatha MC. Study of land use and land cover changes of Mysuru city, Karnataka, India using remote sensing and GIS techniques. Int J Geomatics Geosci. 2016;6:1452–64. Deressu TF, Bojer AK, Debelee TG, Negera WG, Nadarajah S, Gebissa KW. Enhancing land use and land cover classification with deep learning-based satellite imagery segmentation. Int J Appl Earth Obs Geoinf. 2025;6:1–15. Siddique J. Deep learning framework for analysing land use and land cover Proc. SPIE Remote Sensing 6vol. 13558, 2025, pp. 1–10. Nugroho AR, Hidayat R, Wibowo S. Patch-based classification using ResNet for land cover changes detection of Batu. CityInt J Remote Sens Earth Sci. 2023;6(2):85–96. Sawant S, Joshi A, Patil R. Sen-2 LULC: A land use and land cover dataset from Sentinel-2 satellite imagery for deep learning Sci. Data 6vol. 10, 1, pp. 1–12, (2023). Alem A, Kassahun ST, Gashaw TK. Transfer learning-based deep learning models for land cover classification using satellite imagery. J Appl Remote Sens. 2022;16(2):1–14. Acuña-Alonso C, Novo A, Rodríguez JL, Varandas S, Álvarez X. Modelling and evaluation of land use changes through satellite images in a multifunctional catchment: social, economic and environmental implications. Ecol Inf. 2022;74:101777. https://doi.org/10.1016/j.ecoinf.2022.101777 . Zhu G, Hu J, Zhang L. Deep learning in remote sensing: A review IEEE Geosci . Remote Sens Mag 69, 4. pp. 8–36, Dec. (2021). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted 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-8731446","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Case Report","associatedPublications":[],"authors":[{"id":589626414,"identity":"8b309edf-abac-43ca-864a-11662487c1f9","order_by":0,"name":"Aditya Ravi","email":"","orcid":"","institution":"RV College of Engineering","correspondingAuthor":false,"prefix":"","firstName":"Aditya","middleName":"","lastName":"Ravi","suffix":""},{"id":589626417,"identity":"dddba641-d4f4-400a-91a1-e2c482ae59c1","order_by":1,"name":"A. Ashirvad Ashirvad","email":"","orcid":"","institution":"RV College of Engineering","correspondingAuthor":false,"prefix":"","firstName":"A.","middleName":"Ashirvad","lastName":"Ashirvad","suffix":""},{"id":589626421,"identity":"8845494a-0092-4630-af86-0c53b163c2f2","order_by":2,"name":"Sankalpa B. R.","email":"","orcid":"","institution":"RV College of Engineering","correspondingAuthor":false,"prefix":"","firstName":"Sankalpa","middleName":"B.","lastName":"R.","suffix":""},{"id":589626422,"identity":"13a81a80-00e7-41db-a986-e2bf02b8e3e3","order_by":3,"name":"Shivaraj Chawan","email":"","orcid":"","institution":"RV College of Engineering","correspondingAuthor":false,"prefix":"","firstName":"Shivaraj","middleName":"","lastName":"Chawan","suffix":""},{"id":589626423,"identity":"b4e19f0f-49f1-49ed-94b3-c4e2510c6496","order_by":4,"name":"Rajashree Shettar","email":"data:image/png;base64,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","orcid":"","institution":"RV College of Engineering","correspondingAuthor":true,"prefix":"","firstName":"Rajashree","middleName":"","lastName":"Shettar","suffix":""}],"badges":[],"createdAt":"2026-01-29 12:23:27","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8731446/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8731446/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":102837804,"identity":"9b7eb991-c0d5-4e6a-8653-fd2aa954ddf3","added_by":"auto","created_at":"2026-02-17 11:26:04","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":40367,"visible":true,"origin":"","legend":"\u003cp\u003eProcessing pipeline of the LULC system\u003c/p\u003e","description":"","filename":"image1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8731446/v1/b29730601fb7d3e753a5e252.jpeg"},{"id":102837806,"identity":"a53710a3-e3b8-42cc-8087-8566e859a76d","added_by":"auto","created_at":"2026-02-17 11:26:06","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":73472,"visible":true,"origin":"","legend":"\u003cp\u003eSpatial features of all ten classes of EuroSat dataset\u003c/p\u003e","description":"","filename":"image2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8731446/v1/5a292417e71e7840959d79ea.jpeg"},{"id":102837836,"identity":"e5c3a70c-8bdc-4771-955f-ad6d9d82ec2c","added_by":"auto","created_at":"2026-02-17 11:26:17","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":263725,"visible":true,"origin":"","legend":"\u003cp\u003eLand cover map of the place Mysuru, Karnataka\u003c/p\u003e","description":"","filename":"image3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8731446/v1/6c0ed21beee6b5ce367af251.jpeg"},{"id":102837755,"identity":"ea9037b8-7ced-481c-aa35-c7d63246559c","added_by":"auto","created_at":"2026-02-17 11:25:49","extension":"jpeg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":42509,"visible":true,"origin":"","legend":"\u003cp\u003eCount of Images belonging to each class in the Eurosat dataset\u003c/p\u003e","description":"","filename":"image4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8731446/v1/44ed7fa5547d1b23b08d9063.jpeg"},{"id":102837775,"identity":"4d3e2969-e6e5-4d9b-8028-c7defbea1a16","added_by":"auto","created_at":"2026-02-17 11:25:59","extension":"jpeg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":158031,"visible":true,"origin":"","legend":"\u003cp\u003eThe system architecture of land use land cover classification system\u003c/p\u003e","description":"","filename":"image5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8731446/v1/deda785d790d3c0263f706e8.jpeg"},{"id":102837796,"identity":"248e1e39-302c-406e-a8c3-c14c57bfedd1","added_by":"auto","created_at":"2026-02-17 11:26:03","extension":"jpeg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":198663,"visible":true,"origin":"","legend":"\u003cp\u003eA region selected along with its statistical summary in the dashboard\u003c/p\u003e","description":"","filename":"image6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8731446/v1/079ab5c3618a214d917f58ff.jpeg"},{"id":102837842,"identity":"4725d43d-d502-4bc2-bdd6-adc5a60de6fe","added_by":"auto","created_at":"2026-02-17 11:26:19","extension":"jpeg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":95415,"visible":true,"origin":"","legend":"\u003cp\u003eDashboard of the LULC system displaying the statistical summary with charts\u003c/p\u003e","description":"","filename":"image7.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8731446/v1/9511c886b88d5624c9178c3d.jpeg"},{"id":102837778,"identity":"9b8596cd-0039-4452-8ba7-03f590856aac","added_by":"auto","created_at":"2026-02-17 11:26:01","extension":"jpeg","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":88602,"visible":true,"origin":"","legend":"\u003cp\u003eDashboard of the LULC system displaying the statistical summaryof land cover changes for 2022-23 and 2023-2025\u003c/p\u003e","description":"","filename":"image8.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8731446/v1/af22e1681bc809921562a579.jpeg"},{"id":102837757,"identity":"14155094-27b9-4f81-ac5c-61d6a0e553f4","added_by":"auto","created_at":"2026-02-17 11:25:49","extension":"jpeg","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":78694,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of land use pattern for the Mysore region across 2022-2023 and 2023-2025\u003c/p\u003e","description":"","filename":"image9.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8731446/v1/02f01e857df4f63aba4b1c3b.jpeg"},{"id":102837777,"identity":"4dfd8576-dfe1-4056-be31-ddec9d27bf83","added_by":"auto","created_at":"2026-02-17 11:26:00","extension":"jpeg","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":66590,"visible":true,"origin":"","legend":"\u003cp\u003eAI based chatbot integrated to the system to query about the changes in the land use pattern from one period to another\u003c/p\u003e","description":"","filename":"image10.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8731446/v1/c332a17d10f73001f6ea579d.jpeg"},{"id":105365320,"identity":"831a0ece-9c7c-4fb7-85d7-c7d65ac2b00a","added_by":"auto","created_at":"2026-03-25 08:29:26","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2188148,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8731446/v1/5cc7f9e7-1279-4140-b697-7b76289d4f6a.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Deep Learning Based Land Use and Land Cover Classification Using Remote Sensing Imagery","fulltext":[{"header":"Background","content":"\u003cp\u003eThe human activities in a region are reflected by the land usage patterns. On one side of it, we have rapid urbanization and infrastructure development, which also leads to deforestation in some regions. On the other hand, there is a constant attempt to improve the amount of agricultural area. This impacts the environmental sustainability and other ecosystems directly. This created the need for monitoring land usage patterns to decide how efficiently the land needs to be used to ensure environmental sustainability while also ensuring development and urbanization. Satellite imagery lets us acquire data through capturing repeated images for the selected region. It also allows us to acquire data across different time periods to analyze the temporal change in land use patterns. Sentinel-2 gives us multi-spectral imagery that we can use to get more reliable results for land use and land cover patterns. This aligns with the goal of classifying the segmented land to one of the 10 classes of land usage in a region. The classification of land in a region is important to observe how efficiently the land is being used.\u003c/p\u003e\u003cp\u003eThere are many challenges in classifying land use, and one of the most fundamental challenges is the lack of clear im- ages due to fog, mist, or other natural atmospheric interventions. This could lead to misclassifications due to spectral similarity. When pixels are mixed or diffused, the predictions are generally inaccurate. The existing techniques used fixed or static rules to handle such cases. They are inefficient while classifying lands at different regions or at different times. This is where machine learning techniques can be used to learn the relationship between spectral features of satellite images and the land classes. Deep learning models uncover the hidden spectral features from the images. Unsupervised learning techniques like Convolutional Neural Network (CNN) outperform the regular supervised learning models in such task. An additional layer of transfer learning can make the generic model more suitable for the task in hand. There is no support for common users in such existing work. Geospatial platforms have complex workflows that make it difficult with complex workflows that does not support regular users.\u003c/p\u003e\u003cp\u003eThis approach supports the common users with a region-based classification along with visualization and statistical summary along with an AI chatbot that allows users to interpret the results of the complex machine learning model. This system contributes by maintaining the reliability and accuracy while making it conceptually simple and straightforward to help users without domain knowledge. The most important beneficiaries of this work include decision making authorities that decide the feasibility of a new project that could either contribute positively or negatively to the environment in the name of urbanization. This is an end-to-end platform that acquires images of a selected region, segments the land based on pixels and classifies each segment into one of the classes thereby determining the overall land usage pattern of the area.\u003c/p\u003e"},{"header":"1. Introduction","content":"\u003cp\u003eLand Use and Land Cover analysis is a continuously researched area over the last few decades. The initial studies experimented with regular image processing. This was followed by applying supervised machine learning techniques which also did not efficiently do the job. The initial literature work shows the comparison among many such direct techniques and the performance gains in all these techniques. The earlier systems did not have access to such clear satellite imagery too. Previous works used many distorted and unclear images which were a very obvious reason to their low accuracy percentages. There was no existing work that worked precisely for all the regions. Early work using the EuroSat dataset also did not provide the same amount of consistency and accuracy when applied in other regions of the world. Coarse and medium resolution satellite were predominantly used for regional and global analysis. MODIS and other such sensors were used for data acquisition. They were very accurate for large scale vegetations and agricultural land but failed in urban or pre-urban landscapes that were localized due to lot of internal changes. The pixels that were in the intersection of such lands that were mixed or diffused suffered with severe misclassifications. Remote sensing and GIS were used together for classification. Spectral changes were not considered or analyzed to predict the land usage. The resolution of the images, the natural interventions due to seasonality and the model used played key role in determining the results. The amount of preprocessing required for the machine learning model was also unclear and this largely determined the results of the model. Outputs were also not interpretable to be efficiently used for land usage planning. A gradual shift from techniques like support vector machines, random forests, decision trees to neural networks improved the model results. Such models worked better in different regions even in complex pre-urban landscapes. Still the generalization of a model across space and time remained a big challenge.\u003c/p\u003e"},{"header":"2. Literature Review","content":"\u003cp\u003eEarly research on land use and land cover change detection was based on pixel based classic approach and also needed post classification methods on the multi temporal satellite images. Civco et al. showed that these methods could be used to identify the land cover changes but are also bound by spectral confusion, mixed pixels and poor transferability in different time periods and regions. Studies done before show that despite the advancements in technology the main focus should be on the classification accuracy with limited attention to user friendly integrated systems that combine spatial temporal analysis, visualization and accessibility for beginners [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eMedium resolution and coarse satellite data helped in large area monitoring but decreased the spatial detail. Zhan et al. showed land cover change detection using MODIS 250 m data, which improved identification of change at regional and global level, with limited capacity to show fine grained changes [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Similar to this Green et al. also showed that multi temporal Landsat images can be used for land use monitoring on long term scale, while also showing that error propagation was present between time steps and we had to manually handle them [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eDevelopments which have happened recently show that efficient deployment and interpretation of complex models are important. Groq\u0026rsquo;s LPU based API\u0026rsquo;s can be used for fast execution of deep learning and language models and they also support visual analysis and detailed explanation of results [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Recently machine learning methods have improved classification by using non linear spectral relationships; Kasahun and Legesse showed we can also use random forest, support vector machines and artificial neural networks to perform better change detection and urban mapping compared to traditional methods [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eStudies done recently show that land use and land cover change modeling should be applied for analysis of mainly urban areas. Gaur and Singh reviewed machine learning, statistical and hybrid land use and land cover methods, showing the role of data based approaches and also the need to integrate policy and planning methods for urban growth [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. With the increasing volume of earth observation data, managing and analyzing large scale datasets has become difficult. Sudmanns et al. introduced the concept of \"Big Earth Data\" and also showed the need for web based, scalable methods to process multi source satellite datasets in a better [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eModern web frameworks has made it easy to develop and use interactive platforms. Modern full stack frameworks such as Next.js has made the development of interactive, scalable and high speed web applications easy even in data heavy systems [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Deep learning has improved remote sensing image analysis, Pelletier et al. provided a comparative study of deep learning architectures for hyperspectral image classification, identifying the existing challenges with respect to data size, transfer learning and model generalization [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. The classical studies on land use and land cover change detection summarized the\u003c/p\u003e \u003cp\u003edifferences between modern and traditional techniques showing problems such as mixed pixels, lack of standardized land use methods and classification errors [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe ResNet50 based architectures have shown good capacity to handle robust and transferable features in different applications. Kavitha and Karpagam used a ResNet-50 encoder with a hybrid LSTM-GRU decoder which used a beam search for image captioning, achieving better representation and caption quality compared to the traditional encoder decoder model [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Rath et al. used a finetuned ResNet-50 model for brain tumor detection in MRI scans, which showed high classification accuracy and better generalization, while also showing the impact of transfer learning and residual networks in image analysis tasks [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eResNet50 based models have also shown strong results in application-oriented tasks such as disaster recognition, showing stable training behavior and also has improved the classification accuracy in complex environments [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Deep learning based convolutional neural networks pretrained on large datasets have made transfer learning more effective in various domains. The ImageNet dataset has played a very important role in training deep learning visual models, which further improved the generalization and overall performance for a variety of computer vision tasks [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. The land use and land cover research reviews have shown the growing change from static land use mapping to impact analysis and factors assessment which are possible by advancements in remote sensing approaches [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eStudies done recently have shown modeling land use and land cover changes to understand urban area expansion and its impact on environment. Mehra and Swain integrated artificial neural networks with cellular automata to predict the future land use patterns, while also showing the effectiveness of hybrid approach for forecasting urban growth indicating the influence of socio economic and spatial variables [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Agrawal et al. analyzed the spatiotemporal land use changes in highly populated rural areas of India, showing the change towards urban like patterns which are driven by the population pressure [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Bindajam et al. applied geostatistical methods to study urban expansion in Bengaluru, showing the need for spatial analysis for urban management [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eRegional studies have started to make use of advanced learning models for detailed land use and land cover analysis. Mahendra et al. used vision transformers to analyze land use changes in Mysuru district, and showed improved representation of spatial dependencies compared to the traditional methods [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. The work by Sharifi et al. used remote sensing and GIS techniques to analyze temporal land use and land cover change in Mysuru city, showing the impact of urbanization on the natural resources [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. Some of the deep learning based semantic segmentation methods also improved the classification rate in different land classes, as shown by Deressu et al. through evaluation of UNet and DeepLab based architectures [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eDeep learning frameworks have been used widely for land use and land cover classification and change analysis using satellite images. Siddique et al. came up with a ResNet50 based framework for multi period land use and land cover classification using Sentinel-2 data, accomplishing a high accuracy [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Nugroho et al. used patch-based classification using ResNet architecture to find land cover changes in urban areas, which also showed the detection of road expansion and construction activities [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. High quality datasets such as Sen-2 LULC have improved research by allowing robust training and analysis of deep learning models for various land classes [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eTransfer learning has improved the classification performance on the land use and land cover with limited labeled data. Alem et al. shoed that finetuned ResNet based models perform better than other models for satellite image classification, showing the impact of transfer learning in remote sensing applications [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. Satellite image-based land use and land cover analysis is also integrated with environmental and socio-economic modeling as shown by Acu\u0026ntilde;a-Alonso et al., who evaluated the land use changes and their results for the analysis and mitigation of floods using nature-based solutions [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. Deep learning used in remote sensing also indicate the huge potential and the challenges of adopting deep learning based models, with respect to interpretability, scalability and adaptation to the domain [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe use of deep learning models for classifying satellite images increased significantly. Convolutional Neural Network (CNN) models were able to read the hidden features from the satellite images that contributed to analyzing the land use and land cover. Popular deep learning architectures like ResNet, U-Net, and other transformer-based models showed great potential for this task. Datasets like EuroSat became the benchmark for the analysis of land use. The challenge of limited data continued to prevail and this was countered by incorporating transfer learning techniques. Limitations appear in much of the literature. Generalization fails across seasons and geographic regions. Models are sensitive to sensor and atmospheric variability. Training and inference demand high computation. Deep learning outputs lack interpretability. Many studies focus on accuracy metrics alone.\u003c/p\u003e \u003cp\u003eResearch gaps stand out. Usability and accessibility for non-experts get little attention. Integrated platforms for classification, change analysis, and reporting are rare. Few systems turn outputs into easy-to-read information. Practical planning and decision-making need better solutions. This work draws motivation from gaps in integration and usability for planning and monitoring tasks. Systems must balance technical performance and practical use.\u003c/p\u003e"},{"header":"3. Related Work","content":"\u003ch2\u003eA.\u0026nbsp;Conventional\u0026nbsp;Approaches\u0026nbsp;of\u0026nbsp;Remote\u0026nbsp;Sensing\u003c/h2\u003e\n\u003cp\u003eEarlier methods of land use and land cover classification used approaches which are statistical and pixel based. Common approaches were maximum likelihood, minimum distance and Bayesian classifiers. These methods took predefined statistical distribution of the spectral values. They performed well in homogeneous regions. For simple cases these methods worked but they didn\u0026rsquo;t work well with real world use cases.\u003c/p\u003e\n\u003cp\u003eThe performance of the methods was less in complex landscapes like urban regions.\u0026nbsp;Spectral similarity confused roads,\u0026nbsp;soil, vegetation and areas. The spatial context was skipped because of pixel wise treatment. It decreased the handling of the heterogeneous urban areas. Issues were also present in areas of changing of lighting and seasons.\u003c/p\u003e\n\u003ch2\u003eB.\u0026nbsp;Machine\u0026nbsp;Learning\u0026nbsp;Methods\u0026nbsp;for\u0026nbsp;LULC\u003c/h2\u003e\n\u003cp\u003eNon-linear\u0026nbsp;modeling\u0026nbsp;of\u0026nbsp;spectral\u0026nbsp;class\u0026nbsp;relationships\u0026nbsp;was\u0026nbsp;done\u0026nbsp;through\u0026nbsp;machine\u0026nbsp;learning. Some\u0026nbsp;algorithms\u0026nbsp;like\u0026nbsp;support\u0026nbsp;vector machines,\u0026nbsp;random\u0026nbsp;forests,\u0026nbsp;decision\u0026nbsp;trees,\u0026nbsp;and\u0026nbsp;k-nearest\u0026nbsp;neighbors\u0026nbsp;were\u0026nbsp;used. These\u0026nbsp;methods\u0026nbsp;showed\u0026nbsp;more\u0026nbsp;robustness\u0026nbsp;that\u0026nbsp;the usual methods. The handling of heterogeneous and high dimensional data became better. The reliance on strict distribution assumptions decreased.\u003c/p\u003e\n\u003cp\u003eBut they still had to use manually engineered features.\u0026nbsp;Feature design was given to domain experts.\u0026nbsp;Generalization in\u0026nbsp;many\u0026nbsp;regions\u0026nbsp;and\u0026nbsp;sensors stayed\u0026nbsp;high. Users\u0026nbsp;had to\u0026nbsp;select\u0026nbsp;and change\u0026nbsp;the\u0026nbsp;features for\u0026nbsp;each new\u0026nbsp;area. This\u0026nbsp;made the\u0026nbsp;process very slow and error prone.\u003c/p\u003e\n\u003ch2\u003eC.\u0026nbsp;Deep\u0026nbsp;Learning\u0026nbsp;in\u0026nbsp;Satellite\u0026nbsp;Image\u0026nbsp;Analysis\u003c/h2\u003e\n\u003cp\u003eThe deep learning methods allowed feature extraction to happen automatically from raw images. Convolutional neural networks learned the hierarchical spatial features. Classification improved in complex landscapes. Architectures such as ResNet, U-Net, and transformer-based models were used. Good results showed up on standard datasets like EuroSAT.\u003c/p\u003e\n\u003cp\u003eTransfer learning was useful especially in region specific tasks. The demand for handcrafted features dropped. The computation cost and complexity of the process of training became more. Even then these models were good with the details like edges and textures. They were also scaled to large size images than before.\u003c/p\u003e\n\u003ch2\u003eD.\u0026nbsp;Limitations\u0026nbsp;of\u0026nbsp;Existing\u0026nbsp;Platforms\u003c/h2\u003e\n\u003cp\u003eThe deployment of the machine learning methods needs knowledge of remote sensing and programming. Scalability is improved with cloud based geospatial services. The users have to deal with workflows and technical methods. There is limited support for visualization and analysis of results.\u003c/p\u003e\n\u003cp\u003eThe output was not very good for planners nor the policy makers. There are very few easy ways to track changes over time in most of the systems. Few of the systems have built in tools for maps and reports.\u003c/p\u003e"},{"header":"4. Methods","content":"\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e4.1 System Overview\u003c/h2\u003e \u003cp\u003eThe Land Use and Land Cover (LULC) system works as an end-to-end tool right from region selection to summary and analysis as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. It collects raw images of the selected region, segments the region and classifies each segment into one of the ten classes. The main goal of the system is reliability, accuracy and consistency. Once the appropriate region has been selected, basic image processing is used to segment the area. This is followed by using the ML model based on ResNet-50 to efficiently classify and give a statistical summary and also generate a report for temporal analysis. This basic setup is made to handle both single images and time series data.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Data acquisition\u003c/h2\u003e \u003cp\u003eThe developed system uses the Sentinel-2 satellite imagery as the source which is acquired from Google Earth engine. The user selects a region of interest by creating a polygon on the map interface which is provided in the user dashboard. The user also specifies a time period for the analysis period. The engine pulls data from the available cloud services or public data repositories for processing in the further steps. Then cloud cover filtering is performed to take only images that are not covered with fog or mist. This ensures that clear images are selected for processing.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Land Use Land Cover Classes\u003c/h2\u003e \u003cp\u003eThe EuroSat dataset is a benchmark dataset, which gives ten different land cover classes derived from the Sentinel-2 multispectral satellite imagery. This dataset covers both urban and rural regions with multiple classes. Each class captures spectral characteristics, allowing the deep learning models discussed later to learn spectral features for classification.\u003c/p\u003e \u003cp\u003eThe dataset includes vegetation classes such as annual crop, permanent crop, pasture, herbaceous vegetation, and forest. Urban classes such as residential areas, industrial zones, and highways give the heterogeneous regions with complex spatial features and mixed or diffused pixels. Water bodies such as rivers and sea or lake has strong spectral features which can be used. The spatial features differentiate each of these classes and can be observed in the Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. These spatial features are too subtle at times and can lead to misclassification.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Data preprocessing\u003c/h2\u003e \u003cp\u003eInitially, we perform a set of basic radiometric and atmospheric correction steps to ensure clear images. This step corrects the lighting and blurred effects on the acquired images. We then perform image tiling to generate image segments of the same size. The overlapping between these tiles avoids all sorts of mixed pixel issues. We then normalize the pixel to adjust the value between 0 and 1. Noisy pixels or mixed pixels are masked in this stage.\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows the image resized to 224 x 224 pixels size to match the input dimensions required by the ResNet-50 model that we are using in this work.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e4.5 Classification model\u003c/h2\u003e \u003cp\u003eWe use a ResNet-50 based CNN model to classify each segment of the image. This model has residual blocks to accommodate deep learning. We then use transfer learning from pre-trained weights. This finetuning makes it suitable for land cover classification thereby making it more domain specific. The output for each tile is one of the ten classes. By counting the number of tiles in each class, the percentage of each land use type can be computed.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e4.6 Training procedure\u003c/h2\u003e \u003cp\u003eThe model is trained using the popular benchmark dataset called EuroSat which has captured images from the Europe region. This dataset is split into three parts, namely the training set, the testing set and the validation set. Optimizers like Adam Optimizers are used to update the weights with every epoch. About 40 epochs were used in the training stage of this model. The validation loss is used to identify the ideal number of epochs using a training vs validation loss graph.\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e gives the number of images belonging to each of the ten classes in the EuroSat dataset. Therefore, some amount of data augmentation is required in order to make the classes have the same number of images to add up to the imbalance and this comes with the responsibility to not cause overfitting of weights. Data augmentation techniques such as random rotations and flips in both directions are applied to the input images to remove the imbalance.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e4.7 Inference and prediction\u003c/h2\u003e \u003cp\u003eThe region selected usually does not belong to the EuroSat dataset. The selected region is segmented and passed through the trained model. The model then predicts the class of each tile from the ten classes of the training data. Each tile gets a class probability map. Tile predictions aggregate into full spatial maps. Overlapping tiles average predictions smoothly. Post-processing steps apply thresholds or smoothing. This reduces noise in final maps.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e4.8 Temporal change analysis\u003c/h2\u003e \u003cp\u003eThe system supports single-period and multi-period analysis. Single-period classifies one image or time window. Multi-period handles several dates. Each time window gets independent classification. Land cover maps from different periods compare side by side. Class-wise changes compute area shifts. Transition statistics track from-to patterns like forest to urban.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e4.9 Interpretation module\u003c/h2\u003e \u003cp\u003eStatistical summaries pull from classification results. Percentages show land cover distribution per class. Numerical outputs convert to readable explanations. Text describes dominant classes and changes. Clarity targets non-expert users like planners. Simple language avoids technical terms.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e4.10 Output and visualization\u003c/h2\u003e \u003cp\u003eColor-coded land cover maps generate for each period. Colors match standard land cover schemes. A dashboard or interface shows maps side by side. Users zoom and pan for inspection. Results export as images for reports. Tables list statistics. Full reports combine maps and summaries.\u003c/p\u003e \u003c/div\u003e"},{"header":"5. System Architecture","content":"\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e5.1 Purpose of the architecture\u003c/h2\u003e \u003cp\u003eThe architecture supports modular and scalable processing of satellite imagery as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e. User interaction stays separate from the backend computation. This setup allows extensibility for new features. Independent component updates avoid full system rebuilds. Scalability handles larger regions or more users. The system is basically designed to handle reliability and accuracy and has the potential to be scaled to accommodate large traffic and select larger areas to be computed.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e5.2 Overall architectural design\u003c/h2\u003e \u003cp\u003eThe design follows a client-server architecture. Frontend focuses on visualization and user controls. Backend runs compute intensive tasks like model inference. Lightweight tasks such as map display stay on the frontend. This split keeps the interface responsive.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e5.3 Frontend component\u003c/h2\u003e \u003cp\u003eA web-based interface provides user interaction. Interactive map tools let users draw or select region of interest. Slider controls set temporal range for analysis. Land cover maps display with color coding. Change results show differences over time. Dashboard panels list statistics and summaries. Users can zoom, pan, and toggle layers.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e5.4 Backend orchestration layer\u003c/h2\u003e \u003cp\u003eA central service acts as the orchestration layer. It receives requests from the frontend. The layer sequences preprocessing, inference, and analysis steps. Structured requests log parameters for reproducibility. Error handling routes issues back to the frontend. Workflow status updates keep users informed.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e5.5 Satellite data access layer\u003c/h2\u003e \u003cp\u003eThis layer interfaces with Sentinel-2 imagery sources. Cloud-based platforms or APIs fetch the data. Initial filtering removes high cloud cover images. Download queues manage multiple requests. Retrieved imagery stores temporarily. The layer passes files to preprocessing with metadata.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003e5.6 Preprocessing and data handling module\u003c/h2\u003e \u003cp\u003eThe module handles correction, tiling, and normalization steps. Atmospheric correction adjusts for air and light effects. Tiling breaks large images into model-sized patches. Normalization scales bands consistently. Overlaps on tiles prevent boundary artifacts. Time period consistency checks run across images. Clean data goes to the inference engine.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003e5.7 Deep learning inference engine\u003c/h2\u003e \u003cp\u003eThe engine loads the trained ResNet50-based model. Tile-level classification processes each patch. Batch processing speeds up large regions. GPU support accelerates predictions where available. Output includes class predictions and probability maps per pixel. Results assemble into full scene maps.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003e5.8 Temporal analysis module\u003c/h2\u003e \u003cp\u003eOutputs from multiple time windows feed into this module. Land cover maps align spatially for comparison. Change matrices count area shifts by class. Transition statistics detail from-to movements. Overall change summaries calculate totals. Data formats prepare for visualization and reports.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec23\" class=\"Section2\"\u003e \u003ch2\u003e5.9 Interpretation and reporting module\u003c/h2\u003e \u003cp\u003eThis module generates statistical summaries from raw outputs. Percentages and areas describe land cover distribution. Numerical results convert to plain text explanations. Examples note major changes like urban growth. Content fits dashboards and export formats.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section2\"\u003e \u003ch2\u003e5.10 Data flow and communication\u003c/h2\u003e \u003cp\u003eModules use structured data exchange. REST APIs or message queues connect components. JSON formats define inputs and outputs clearly. Authentication secures backend calls. Logging tracks data movement end-to-end.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec25\" class=\"Section2\"\u003e \u003ch2\u003e5.11 Link to system diagram\u003c/h2\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e shows the full system architecture. Arrows mark data flow from frontend through modules to outputs. The orchestration layer sits central. The diagram summarizes the workflow from imagery access to final reports.\u003c/p\u003e \u003c/div\u003e"},{"header":"6. Results and Discussion","content":"\u003cdiv id=\"Sec27\" class=\"Section2\"\u003e \u003ch2\u003e6.1 Overview of experimental setup\u003c/h2\u003e \u003cp\u003eExperiments tested the platform using Sentinel-2 imagery from different regions. The trained ResNet50 model was used to classify land cover. Evaluation covered multiple time periods across several recent years. Selected areas included both urban zones and rural landscapes. The goal was to examine classification quality and change detection behavior. Tests ran on typical desktop hardware.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec28\" class=\"Section2\"\u003e \u003ch2\u003e6.2 Land cover classification results\u003c/h2\u003e \u003cp\u003eGenerated land cover maps covered full regions clearly. Dominant classes included urban built-up areas and forest vegetation. Crop fields appeared in agricultural regions. Maps remained visually consistent across test sites. Urban clusters concentrated near roads and cities, while vegetation zones followed rivers and hilly areas. Some bare soil appeared during dry seasons. Overall patterns matched expected land use as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eThe land cover map in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows the distribution of major land use and land cover classes across Mysuru region, Karnataka, which are derived from Sentinel-2 imagery. Every grid cell shows a classified tile given by the ResNet-50 based deep learning model, with color coded classes showing the major land types such as water bodies, vegetation, built up areas, bare and open land. The vegetation regions cover the major area of Mysuru, while the built-up clusters are majorly present near the urban core, showing patterns of urbanization and infrastructures. The water bodies are shown distinctly in the northwestern part of the region, showing the models ability to capture classes even at medium spatial resolution. The grid-based output has both spatial maps and localized outputs between land classes, and it will make it easy for anyone to analyze the land use patterns for urban planning and environmental monitoring.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec29\" class=\"Section2\"\u003e \u003ch2\u003e6.3 Interpretation of change maps\u003c/h2\u003e \u003cp\u003eThe generated change maps use a basic set of colors to highlight the difference and this is used to analyze in the further steps as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e. Red marked areas converting to urban use. Blue indicated water gain or loss. Green represented unchanged vegetation. Visual patterns revealed growth directions clearly. Comparing maps side by side was quick and intuitive. Color usage remained consistent across time periods.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec30\" class=\"Section2\"\u003e \u003ch2\u003e6.4 Visualization and output analysis\u003c/h2\u003e \u003cp\u003eColor-coded maps made land cover distribution easy to interpret. Spatial representation captured small details such as field boundaries. Dashboards allowed users to switch between time periods quickly. Summary tables listed area percentages for each class as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e. Export options supported saving maps for reports. Overall, the visual tools sped up result review.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec31\" class=\"Section2\"\u003e \u003ch2\u003e6.5 Interpretation module behavior\u003c/h2\u003e \u003cp\u003eGenerated summaries reported class areas and percentages of land cover changes are as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e. Text explanations described changes in simple terms, such as urban growth across the region. Non-expert users could understand insights without remote sensing knowledge. Examples highlighted major changes as well as stable areas.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e shows a comparison of the land cover maps for the place Mysuru belonging to Karnataka state, between 2022\u0026ndash;2023 and 2023\u0026ndash;2025, showing the differences in urban and vegetated areas. The side-by-side comparison enables easy analysis of land use changes and also allows for regional planning and monitoring activities.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e shows an AI based chatbot integrated with the land use and land cover analysis system which allows users to ask questions regarding the distribution of land area across different classes and time periods. Users can interact with the chatbot and ask natural language questions related class wise area division, the major land cover pattern, and the impact of urbanization between the timelines. It converts statistical analysis into simple human understandable explanations and so the system improves the interpretability for non-experts hence allowing better decision making in land management.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec32\" class=\"Section2\"\u003e \u003ch2\u003e6.6 Comparative analysis of ResNet50 with EfficientNet-B0\u003c/h2\u003e \u003cp\u003eThe model ResNet-50 was selected as the most appropriate model for this project after a comparative analysis with another popular benchmark model called EfficietNet-B0. This was done using the same EuroSat dataset. The models were initialized with ImageNet pretrained weights and subsequently updated with EuroSat dataset during the training process. The input resolution for both the models is 224 x 224, but the number of parameters is much higher in ResNet. ResNet-50 has 25.6\u0026nbsp;million parameters, whereas EfficientNet-B0 has only 5.3\u0026nbsp;million parameters. On the other hand, EfficientNet-B0 has 237 layers, but ResNet-50 has only 50 layers, as the name suggests. EfficientNet has very few trainable parameters due to its compound scaling of width and resolution in an image. This makes it very suitable for lightweight applications in general but this limitation makes it struggle in medium resolution images with complex hidden spatial features which are not obvious. This leads to a high number of misclassifications in regions with multiple land use types like water body, vegetation, industrial area etc. ResNet-50 fundamentally uses residual learning on the other hand, which helps it capture complex spatial features. This basic architectural difference allows it to learn hidden feature representations thereby making it more suitable for classifying land use types.\u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e gives the classification accuracy for all the ten classes of the EuroSat dataset for the comparison between ResNet-50 and EfficientNet models. ResNet-50 can be observed to perform better for most classes in this dataset. The difference in performance is more evident in vegetation and urban classes such as residential areas, industrial regions, and highways, due to the misclassification rate of EfficientNet-B0 caused by diffused pixels. Water bodies have evident spectral characteristics and hence give high accuracy in both models.\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\u003eComparison of Accuracy for each class between EfficientNet-B0 and ResNet-50\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\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 \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eClass\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eResNet-50 (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEfficientNet-B0 (%)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAnnual Crop\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e94.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e87.31\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eForest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e97.18\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e96.46\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHerbaceous Vegetation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e95.41\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e88.72\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHighway\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e93.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e86.05\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIndustrial Area\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e95.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e85.14\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePasture\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e94.35\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e83.68\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePermanent Crop\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e93.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e84.92\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResidential Area\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e96.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e89.14\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRiver\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e97.41\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e92.67\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSea / Lake\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e98.06\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e97.88\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\u003eAs displayed in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, ResNet-50 achieves a very high accuracy relative to the contemporary EfficientNet-50 model with an accuracy of 95.11 percentage compared to 88.42 percentage obtained by the EfficientNet model. The improvements can also be observed in precision, recall, and F1-score thereby proving the superiority of the ResNet model.\u003c/p\u003e \u003cp\u003eBased on these observations made, ResNet-50 was selected as the primary classification model for the land use land cover classification system as it can identify all the hidden spatial features of the Sentinel-2 multi spectral imagery.\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\u003eComparative Performance Analysis of ResNet-50 and EfficienetNet-B0\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\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 \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMetric\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eResNet-50\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEfficientNet\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOverall Accuracy (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e95.11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e88.42\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMacro Precision (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e94.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e87.96\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMacro Recall (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e94.53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e87.21\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMacro F1-Score (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e94.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e87.58\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMis-classification Rate in Diffused Pixels (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e4.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e11.6\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\u003eThe usability of the system was an important part of the platform. Maps which were coded with colors and dashboards made results easy to interpret while the readable summaries helped users such as planners and managers to understand key changes. This made the outputs to be used directly and improved the decision making of the user. The system can be used for urban planning, land management and also for long term planning. There were some limitations of the system. The results depend on the quality of the input images. Clouds covering the areas can restrict the image quality, and the environmental conditions also mattered. Seasonal changes also affect the areas which can change the classification in different time periods. The errors in classification were very common in similar classes such as soil, roads, and surfaces. The features which were very close were not easy to classify. These were common in medium resolution satellite images and is a challenge for classification.\u003c/p\u003e \u003cp\u003eThe workflow requirements also increased with the study area size. The regions which are comparatively larger had to be processed for longer and took more memory. The model also needs training data which is labeled. The model cannot be easily adapted to new regions and needs retraining. The results part provided easy to understand explanations for better understanding. The answers were not able to give the technical details. It can be improved further by including the use of higher resolution images, exploration of different model architectures, and better methods to improve outputs and scalability. The platform provides practical workflow for land use and land cover analysis which improves analysis and decision making.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec33\" class=\"Section2\"\u003e \u003ch2\u003e6.7 Discussion of observations\u003c/h2\u003e \u003cp\u003eStrengths were observed in handling mixed urban\u0026ndash;rural regions. Spatial detail resolution supported identification of small features. Since the images acquired are filtered for cloud and are preprocessed both in training and inferencing stages, the model is robust and works in good lighting conditions. The results are critical in determining sustainability of the development and the Urbanization process. This is crucial also for making decisions while approving a new project that belongs to one of the ten classes.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec34\" class=\"Section2\"\u003e \u003ch2\u003e6.8 Limitations observed\u003c/h2\u003e \u003cp\u003eSoil and Bare field were misclassified in several instances due to the similarity of these land use types. The season determined the nature of vegetation which also potentially increased the misclassification opportunities. This system is very dependent on cloud filtering to work efficiently. Presence of cloud throughout the time period will severely affect the results of the model. The computation time highly depends on the size of the area selected. Large areas take excessively long time to determine the predictions for each tile in the large area and even lead to time out or excessive API calls in some cases.\u003c/p\u003e \u003c/div\u003e"},{"header":"7. Conclusion and limitations","content":"\u003cp\u003eThis project was developed for a single system for the complete land use and land cover analysis that uses satellite data, deep learning-based classification and visualization. The platform uses Sentinel-2 imagery using a ResNet50 based model and produces spatial maps, summaries, and outputs. The overall goal was to decrease the workflow while keeping the analysis useful and practical. The system showed good land cover classification in both urban and rural areas. It supports single period mapping and comparison across multiple time periods. Change detection showed major changes such as urban expansion into vegetated areas. The in-depth analysis allowed land cover results to be shown over time, along with the visual maps showing shifts. These support continuous monitoring of land areas.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eData\u0026nbsp;Availability\u0026nbsp;Statement\u003c/h2\u003e\n\u003cp\u003eThe dataset used in this study is publicly available. The EuroSAT dataset is available on Kaggle at: https://www.kaggle. com/datasets/apollo2506/eurosat-dataset\u003c/p\u003e\n\u003ch2\u003eEthical Approval\u003c/h2\u003e\n\u003cp\u003eNot\u0026nbsp;applicable.\u003c/p\u003e\n\u003ch2\u003eClinical\u0026nbsp;Trial\u0026nbsp;Number\u003c/h2\u003e\n\u003cp\u003eNot\u0026nbsp;applicable.\u003c/p\u003e\n\u003ch2\u003eConsent\u0026nbsp;to Participate\u003c/h2\u003e\n\u003cp\u003eNot\u0026nbsp;applicable.\u003c/p\u003e\n\u003ch2\u003eConsent\u0026nbsp;for Publication\u003c/h2\u003e\n\u003cp\u003eNot\u0026nbsp;applicable.\u003c/p\u003e\n\u003ch2\u003eCompeting interests\u003c/h2\u003e\n\u003cp\u003eThe\u0026nbsp;authors\u0026nbsp;declare\u0026nbsp;that\u0026nbsp;they\u0026nbsp;have\u0026nbsp;no\u0026nbsp;competing\u0026nbsp;interests.\u003c/p\u003e\n\u003ch2\u003eFunding\u003c/h2\u003e\n\u003cp\u003eNo\u0026nbsp;funding\u0026nbsp;was\u0026nbsp;received\u0026nbsp;for\u0026nbsp;this\u0026nbsp;research.\u003c/p\u003e\n\u003ch2\u003eAuthor\u0026rsquo;s\u0026nbsp;contributions\u003c/h2\u003e\n\u003cp\u003eAditya Ravi, A. Ashirvad, Sankalpa B. R., and Shivaraj Chawan were involved in conceptualization, dataset collection and organization, methodology preparation, manuscript writing and writing the original draft. Rajashree Shettar was involved in the conceptualization, validation, supervision, along with reviewing and editing.\u003c/p\u003e\n\u003ch2\u003eAcknowledgments\u003c/h2\u003e\n\u003cp\u003eThe authors thank RV College of Engineering, Bengaluru, for providing academic support and computational resources that facilitated this work.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eCivco DL, Hurd JD, Wilson EH, Arnold MP, Prisloe S. A comparison of land use and land cover change detection methods. In \u003cem\u003eProceedings of the ASPRS\u0026ndash;ACSM Annual Conference\u003c/em\u003e, vol. 21 (2002).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhan X, et al. Detection of land cover changes using MODIS 250 m data. Remote Sens Environ. 2002;83:336\u0026ndash;50. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/S0034-4257(02)00024-2\u003c/span\u003e\u003cspan address=\"10.1016/S0034-4257(02)00024-2\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGreen K, Kempka D, Lackey L. Using remote sensing to detect and monitor land-cover and land-use change. Photogrammetric Eng Remote Sens. 1994;60:331\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGROQ Inc. Groq LPU inference APIs and LLaMA 3 integration. Developer Doc (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://docs.groq.com/\u003c/span\u003e\u003cspan address=\"https://docs.groq.com/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKasahun M, Legesse A. Machine learning for urban land use/cover mapping: comparison of artificial neural network, random forest and support vector machine. Heliyon. 2024;10:e39146. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.heliyon.2024.e39146\u003c/span\u003e\u003cspan address=\"10.1016/j.heliyon.2024.e39146\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGaur S, Singh R. A comprehensive review on land use and land cover (LULC) change modeling for urban development: current status and future prospects. Sustainability. 2023;15:903. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/su15020903\u003c/span\u003e\u003cspan address=\"10.3390/su15020903\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSudmanns M, Tiede D, Lang S, Augustin S. Big Earth data: disruptive changes in Earth observation data management and analysis? Int J Digit Earth. 2020;13:832\u0026ndash;50. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/17538947.2019.1585976\u003c/span\u003e\u003cspan address=\"10.1080/17538947.2019.1585976\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVercel. Next.js documentation: the React framework for production (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://nextjs.org/docs\u003c/span\u003e\u003cspan address=\"https://nextjs.org/docs\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePelletier C, Ienco G, Ban Y, Begu\u0026eacute; A. Deep learning for the classification of hyperspectral images: a comparative review. IEEE Geoscience Remote Sens Magazine. 2017;5:29\u0026ndash;56. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/MGRS.2017.2762345\u003c/span\u003e\u003cspan address=\"10.1109/MGRS.2017.2762345\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSingh A, Kumar V. Change detection in land use and land cover using remote sensing data: a review. Int J Comput Appl. 2017;168:8\u0026ndash;15. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.5120/ijca2017914563\u003c/span\u003e\u003cspan address=\"10.5120/ijca2017914563\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKavitha PV, Karpagam V. Image captioning deep learning model using ResNet50 encoder and hybrid LSTM\u0026ndash;GRU decoder optimized with beam search. J Indian Inst Sci. 2025;105:394\u0026ndash;410. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/00051144.2025.2485695\u003c/span\u003e\u003cspan address=\"10.1080/00051144.2025.2485695\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRath A, Mishra BSP, Bagal DK. ResNet50-based deep learning model for accurate brain tumor detection in MRI scans. Neurosci Res Notes. 2024;6:100104. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.nexres.2024.100104\u003c/span\u003e\u003cspan address=\"10.1016/j.nexres.2024.100104\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWen L, Xiao Z, Xu X, Liu B. Disaster recognition and classification based on improved ResNet-50 neural network. Appl Sci. 2025;15:5143. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/app15095143\u003c/span\u003e\u003cspan address=\"10.3390/app15095143\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDeng J et al. ImageNet: a large-scale hierarchical image database. In \u003cem\u003eProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)\u003c/em\u003e, 248\u0026ndash;255 (2009). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/CVPR.2009.5206848\u003c/span\u003e\u003cspan address=\"10.1109/CVPR.2009.5206848\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChang Y, Hou K, Li X, Zhang Y, Chen P. Review of land use and land cover change research progress. \u003cem\u003eIOP Conference Series: Earth and Environmental Science\u003c/em\u003e 113, 012087 (2018). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1088/1755-1315/113/1/012087\u003c/span\u003e\u003cspan address=\"10.1088/1755-1315/113/1/012087\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMehra N, Swain JB. Assessment of land use land cover change and its effects using artificial neural network-based cellular automata. J Eng Appl Sci. 2024;71:70.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAgrawal S, Welegedara N, Parida D. Spatiotemporal land use changes in remote rural regions of In- dia between 2000 and 2020. Environ Plann B: Urban Analytics City Sci. 2024;33:1\u0026ndash;18. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1177/10185291241236307\u003c/span\u003e\u003cspan address=\"10.1177/10185291241236307\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBindajam AA, Mallick J, Alshayeb MJ, Poddar S. Understanding urbanization and its expansion process in Bengaluru through geostatistical models for landscape management. Adv Space Res. 2025;75:11\u0026ndash;8. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.asr.2025.11.018\u003c/span\u003e\u003cspan address=\"10.1016/j.asr.2025.11.018\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMahendra HN, Pushpalatha V, Mallikarjunaswamy S, Rama Subramoniam S, Rao AS, Shekar NC. Analyzing land use land cover changes in Mysuru taluk, Karnataka state, India using vision transformers. Adv Comput Appl Geospatial Stud. 2025;3:100308. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.acags.2025.100308\u003c/span\u003e\u003cspan address=\"10.1016/j.acags.2025.100308\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSharifi V, Shivanna S, Manjunatha MC. Study of land use and land cover changes of Mysuru city, Karnataka, India using remote sensing and GIS techniques. Int J Geomatics Geosci. 2016;6:1452\u0026ndash;64.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDeressu TF, Bojer AK, Debelee TG, Negera WG, Nadarajah S, Gebissa KW. Enhancing land use and land cover classification with deep learning-based satellite imagery segmentation. Int J Appl Earth Obs Geoinf. 2025;6:1\u0026ndash;15.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSiddique J. Deep learning framework for analysing land use and land cover\u003cem\u003eProc. SPIE Remote Sensing\u003c/em\u003e 6vol. 13558, 2025, pp. 1\u0026ndash;10.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNugroho AR, Hidayat R, Wibowo S. Patch-based classification using ResNet for land cover changes detection of Batu. CityInt J Remote Sens Earth Sci. 2023;6(2):85\u0026ndash;96.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSawant S, Joshi A, Patil R. Sen-2 LULC: A land use and land cover dataset from Sentinel-2 satellite imagery for deep learning\u003cem\u003eSci. Data\u003c/em\u003e 6vol. 10, 1, pp. 1\u0026ndash;12, (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlem A, Kassahun ST, Gashaw TK. Transfer learning-based deep learning models for land cover classification using satellite imagery. J Appl Remote Sens. 2022;16(2):1\u0026ndash;14.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAcu\u0026ntilde;a-Alonso C, Novo A, Rodr\u0026iacute;guez JL, Varandas S, \u0026Aacute;lvarez X. Modelling and evaluation of land use changes through satellite images in a multifunctional catchment: social, economic and environmental implications. Ecol Inf. 2022;74:101777. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ecoinf.2022.101777\u003c/span\u003e\u003cspan address=\"10.1016/j.ecoinf.2022.101777\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhu G, Hu J, Zhang L. Deep learning in remote sensing: A review\u003cem\u003eIEEE Geosci\u003c/em\u003e. Remote Sens Mag 69, 4.\u003c/span\u003e \u003cspan\u003epp. 8\u0026ndash;36, Dec. (2021).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Image Segmentation, Land Use Classification, Satellite Imagery, Temporal and Spatial change, RestNet-50, Deep Learning, Unsupervised Learning","lastPublishedDoi":"10.21203/rs.3.rs-8731446/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8731446/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eIn today\u0026rsquo;s world of urban development, it becomes excessively difficult to observe and monitor land use and land cover changes. Since the growth is relatively more rapid in the current age, this analysis is more crucial than ever. The current approach in place uses many techniques, such as satellite imagery, which are hard to practically apply and require strong technical skills. Common users find it difficult to apply the deep learning model outputs. This work proposes a ResNet-50-based system to acquire satellite imagery, followed by land use and land cover classification with a visualization dashboard, which also features a spatial and temporal comparison of the selected area. This makes it especially useful for common users to interpret the current usage of land and the change in land usage over time. The satellite imagery is obtained from the Sentinel-2 dataset that is acquired through the Google Earth Engine. These acquired images are passed through a custom convolutional neural network based on ResNet-50. The model is initially trained for ten classes of land usage based on the EuroSat dataset. This system also allows comparison of specific land areas over a range of time to understand the change in land usage patterns across different times. The user will be able to view a statistical summary of the selected region, and this report can be used to query an AI chatbot to understand and interpret the statistical results. This increases the usability of a niche system to a regular user who does not have an idea about the land usage patterns. This model achieves an accuracy of 95.11 percent, which makes it a reliable and consistent system that can be utilized in real-world decision making process. This can be used to decide whether a new project in one of these 10 classes can be accommodated or not based on the current land usage pattern in the given area.\u003c/p\u003e","manuscriptTitle":"Deep Learning Based Land Use and Land Cover Classification Using Remote Sensing Imagery","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-17 11:24:15","doi":"10.21203/rs.3.rs-8731446/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"495693f8-5a03-4536-9842-69e7ef7354b0","owner":[],"postedDate":"February 17th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-03-25T08:28:21+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-17 11:24:15","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8731446","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8731446","identity":"rs-8731446","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}