SMR-DeepLabV3+: A Refined Model for Cultivated Land Extraction from Cross-View Imagery Using an Attention Mechanism

SMR-DeepLabV3+: A Refined Model for Cultivated Land Extraction from Cross-View Imagery Using an Attention Mechanism | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article SMR-DeepLabV3+: A Refined Model for Cultivated Land Extraction from Cross-View Imagery Using an Attention Mechanism Zheng Fang, Yuhang Gong, Heli Zhu, Canyang Shi, Zhijia Gong, Lin Tian This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5770339/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 01 Jul, 2025 Read the published version in Discover Sustainability → Version 1 posted 8 You are reading this latest preprint version Abstract The protection and monitoring of cultivated land, as a vital resource, are of utmost importance. Rapid and accurate extraction of cultivated land significantly enhances the quality and efficiency of comprehensive natural resource monitoring. In recent years, advanced technologies for cultivated land extraction, particularly deep learning-based semantic segmentation, have yielded impressive results. The DeepLabV3 + model has demonstrated exceptional performance across multiple datasets. However, conventional approaches primarily rely on high-resolution remote sensing imagery, which may not always be feasible in practical scenarios requiring cross-perspective image analysis (e.g., images captured by handheld devices or drones), such as in cultivated land survey verification. This paper introduces a new algorithm, SMR-DeepLabV3+, which integrates an attention mechanism into the DeepLabV3 + model, along with regularization enhancements. The model was trained using 88,268 meticulously annotated samples and evaluated on 3,466 cultivated land patches from Zhaohua District, Guangyuan City, Sichuan Province. The results showed recognition rates of 80.3% for paddy fields, 81.7% for irrigated land, and 86.1% for dryland. The SMR-DeepLabV3 + model significantly improves the precision of cultivated land extraction and area estimation, offering a practical and effective solution for identifying different types of cultivated land. This study contributes to more refined management of cultivated land resources and provides reliable support for assessing the effectiveness of cultivated land protection policies. Cultivated Land Deep Learning Semantic Segmentation DeepLabV3+ Attention Mechanism Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Figure 13 1. Introduction Cultivated land is a fundamental resource essential for human survival and agricultural production. The protection of cultivated land is intrinsically linked to food security, which is vital for the stable development of economics, politics, and societies[1]. As urbanization accelerates and pressure on land resources increases, timely monitoring of cultivated land dynamics has become crucial, particularly for detecting land loss and illegal land use[2]. Establishing a precise, comprehensive, and dynamic monitoring system for natural resources is key to promoting sustainable development. As cultivated land monitoring systems evolve, traditional paper-based surveys have gradually transitioned to integrated air-space-ground monitoring[3]. This approach leverages high-resolution remote sensing images[4], drone imagery[5], and ground-based mobile measurement equipment[6–7], making cultivated land monitoring more comprehensive and efficient. With the rapid advancements in remote sensing and artificial intelligence (AI), cultivated land extraction methods have evolved from traditional pixel-based and object-oriented classification algorithms to intelligent, deep learning-based technologies[3]. Deep learning models, renown for their robust feature extraction capabilities and adaptability to complex environments, have been widely applied to multi-scale remote sensing image classification tasks[8–9]. These applications include cloud-snow distinction[10–12], building extraction[13–14], crop identification[15–16], dust cover detection[17], and road extraction[18]. In the context of cultivated land monitoring, deep learning has significantly improved the accuracy of land identification and segmentation, providing strong technical support for precise monitoring and management, while advancing the field towards more intelligent and automated solutions[19–20]. Semantic segmentation algorithms based on deep learning offer an innovative approach to cultivated land extraction and monitoring. Convolutional neural networks (CNNs), the core technology behind these algorithms, automatically extract spatial and spectral features from images, significantly improving classification accuracy, particularly in multi-scale feature extraction[21]. Popular models such as AlexNet, VGG, and ResNet have been widely applied to land cover classification[22], demonstrating high robustness and generalization ability. Fully Convolutional Networks (FCN) and U-Net architectures are commonly employed in semantic segmentation tasks and have quickly gained traction in remote sensing image classification[23]. DeepLabV3+, which incorporates atrous convolution and Atrous Spatial Pyramid Pooling (ASPP), has been particularly effective in identifying complex land boundaries and improving classification accuracy and refinement[24]. Due to its network structure optimizations, DeepLabV3 + has demonstrated high accuracy across standard datasets[25]. Improvements have been proposed for the DeepLabV3 + model[26–29], such as DAF-DeepLabV3+, MFP-DeepLabV3+,etc. Building upon these advancements, this study introduces the SMR-DeepLabV3 + model to enhance inference speed, reduce memory consumption, and lower data preparation costs. The SMR-DeepLabV3 + model is adopted as the base model for cultivated land extraction and its capabilities are enhanced by integrating the SENet attention mechanism. This addition improves the network's ability to focus on local details and enhances overall feature learning. Furthermore, the model incorporates cross-perspective information, utilizing drone images, remote sensing imagery, and mobile phone photos to build a more diverse and comprehensive sample dataset. This enables a more holistic understanding of cultivated land environments. The model's effectiveness was validated using cultivated land patches, with results showing that it accurately identifies cultivated land areas and types. These findings support the refined management of cultivated land resources. 2. SMR-DeepLabV3 + Model Integration with Attention Mechanism 2.1 Overview of DeepLabV3+ DeepLabV3 + is an enhanced version of the DeepLab series of semantic segmentation algorithms. It builds upon the strengths of DeepLabV3, introducing key innovations and optimizations to overcome common challenges in semantic segmentation, such as the loss of boundary details and the integration of multi-scale features, as shown in Fig. 1 . DeepLabV3 + utilizes an Encoder-Decoder[30] architecture that significantly improves segmentation accuracy by combining high-level semantic features with low-level spatial information, particularly enhancing boundary detail preservation. The decoder employs multi-path feature fusion, integrating features from different stages of the encoder with those generated by the decoder progressively, layer by layer. This approach helps retain fine-resolution details while capturing multi-scale contextual information, leading to more precise object boundary recognition and localization. 2.1.1 Lightweight Backbone Network (Xception) DeepLabV3 + employs Xception as its backbone network. Introduced by Google in 2017[31], it is a lightweight architecture that replaces standard convolution operations with depthwise separable convolutions, which consist of spatial convolutions followed by channel-wise convolutions. This approach significantly reduces the number of parameters and computational complexity, while enhancing the model's capacity for feature representation. Xception (in Fig. 2 ) is composed of 36 convolutional layers, organized into three stages: Entry, Middle, and Exit flows. These stages handle down-sampling, feature learning, and feature aggregation, respectively, providing the foundational structure for feature extraction in DeepLabV3+. The use of residual connections between convolutional layers helps preserve important feature information, allowing the network to maintain a balance between computational efficiency and high performance. This makes Xception particularly well-suited for high-precision semantic segmentation tasks, where accurate feature extraction is critical. 2.1.2 Combined Stochastic Gradient Descent and Adam Optimizers In deep learning, optimizers adjust model parameters to minimize the loss function, directly impacting model performance. DeepLabV3 + adopts a combined strategy that leverages the strengths of both Stochastic Gradient Descent (SGD) and Adam optimizers. SGD excels in handling non-convex optimization problems and has good generalization ability, while Adam achieves faster convergence through adaptive learning rates[32]. SGD, one of the fundamental optimization algorithms in deep learning, updating weights based on gradients computed from each training batch, making it suitable for achieving robust generalization. In contrast, Adam accelerates convergence by combining momentum-based gradient descent with adaptive learning rate adjustments, allowing it to perform well in scenarios involving sparse data or noisy gradients[33]. By combining these optimizers, DeepLabV3 + benefits from the rapid convergence of Adam and the fine-tuning capabilities of SGD, which often yields better final results. This SGD + Adam strategy alternates or stages the two optimizers, using Adam for early-stage optimization and then switching to SGD for refined adjustments. This approach improves both training efficiency and the model's generalization ability, leading to better overall performance. 2.1.3 Loss Function (Cross-Entropy Loss) The loss function is a fundamental component in machine learning and deep learning, used to quantify the error between the model’s predictions and the actual values. Selecting an appropriate loss function is crucial for optimizing model performance. Cross-entropy loss is commonly used in classification tasks, as it measures the divergence between the predicted probability distribution and the true labels[34]. A lower cross-entropy value signifies better model performance by indicating a closer match between the predicted and true distributions. Cross-entropy loss is often used alongside the Softmax function, which normalizes the output into a probability distribution, allowing cross-entropy to compute the error more effectively by comparing the true labels to the predicted probabilities. This combination ensures that the model can optimize its parameters more efficiently during training. 2.2 Optimized SMR-DeepLabV3 + Model This paper introduces the SMR-DeepLabV3 + model, which builds upon and optimizes the DeepLabV3 + architecture at multiple levels. The architecture of the SMR-DeepLabV3 + model is illustrated in Fig. 3 . 2.2.1 Attention Mechanism Integration (Squeeze-and-Excitation Networks) To improve feature representation and enhance semantic segmentation performance, this paper integrates the Squeeze-and-Excitation Networks (SENet) module (as shown in Fig. 4 ), an attention mechanism, into the DeepLabV3 + model. This introduces an efficient and novel approach to feature enhancement. The core idea behind SENet is to assign different weights to each feature channel, allowing the model to highlight critical features while suppressing irrelevant background noise [35] . SENet consists of two steps: "Squeeze" and "Excitation." First, global average pooling is applied to compress spatial information. Then, fully connected layers and a Sigmoid activation function generate weight values for each channel, optimizing feature extraction by emphasizing essential channels and downplaying less relevant ones. The SENet module evaluates the significance of feature channels at a global scale, enabling the model to prioritize important features by assigning adaptive weights. This attention mechanism complements DeepLabV3+'s multi-scale feature fusion strategy, significantly improving the efficiency of blending features across multiple layers. By incorporating SENet, the model effectively integrates shallow spatial details with deeper semantic information, addressing the previous limitations of insufficient deep feature blending. This enhancement leads to better preservation of fine-grained details and a more accurate understanding of global semantic structures, ultimately improving quality and accuracy of image segmentation.In the extraction of cultivated land, SENet's attention mechanism enhances the model's ability to capture subtle differences between various types of cultivated land, thereby improving segmentation accuracy. 2.2.2 Multimodal Information Fusion Multimodal fusion refers to the integration of two or more information modalities, each representing different sources or forms of information (as shown in Fig. 5 ) [36] . These modalities can exhibit redundancy or complementarity, and effectively combining them results in richer and more comprehensive feature representations. In multimodal research, modality fusion enhances the understanding and expression of complex scenes by leveraging the strengths of different modalities, enabling the capture of diverse multidimensional information. This approach has become an essential method for improving model performance in various tasks. In the DeepLabV3 + model, multimodal fusion incorporates images from different perspectives, including mobile phone photos, drone images, and remote sensing data[37]. Each modality provides distinct advantages: mobile phone photos provide rich details, drone images offer a broader field of view, and remote sensing images capture macro-environmental context. By exploiting the complementarity of these modalities, the model significantly improves segmentation performance. To achieve this, data from different modalities are normalized and spatially registered, ensuring they are processed within a unified representation space. High-level semantic features from each modality are then fused using mechanisms such as the ASPP module and attention mechanisms. This multimodal fusion strategy improves the accuracy of segmentation boundaries and enhances the model's robustness, allowing it to adapt more effectively to complex and diverse scenes. 2.2.3 Regularization Term to Optimize Loss Function The application of loss functions and regularization is crucial in cultivated land information extraction tasks. This paper introduces regularization techniques to optimize the loss function by imposing constraints on model parameters, thus preventing overfitting, and improving the model’s generalization capability. L1 regularization encourages sparsity in the model by effectively selecting important features, while L2 regularization enhances stability and robustness by reducing the magnitude of parameter values. The core concept of regularization is to add a penalty term to the loss function to avoid overfitting the training data, and to ensure better performance on unseen data. 3. Data Acquisition and Processing 3.1 Study Area and Experimental Data The study area for this research is Zhaohua District, Guangyuan City, located in the northern region of the Sichuan Basin, along the upper reaches of the Jialing River. Cultivated land in this region is mainly distributed across the river valley plains and hilly areas. Due to the challenging terrain, the cultivated land is fragmented, characteristic of mountainous agricultural landscapes. Zhaohua District’s cultivated land in primarily consists of paddy fields and dryland, with major crops including rice, corn, and wheat. To ensure food security, the Zhaohua District enforces stringent protection policies, focusing on maintaining high-standard cultivated land and regulating land transfer systems to prevent illegal land occupation. Zhaohua District was selected as the study area to investigate land information extraction. The experimental data include photos captured by handheld devices and drones. 3.1.1 Photos from Handheld Devices Handheld devices, such as smartphones and tablets, are now indispensable tools in everyday life. The high-resolution cameras enable the easy capture of high-quality images, providing a valuable data source for various applications. In the context of cultivated land survey and verification, photos taken with handheld devices can quickly gather on-site information, such as land use patterns and crop growth conditions, making them a critical source of data for this study. However, handheld device photos may suffer from issues such as noise or blur due to hand movement, lens contamination, or suboptimal shooting conditions. To ensure accurate extraction and identification of cultivated land information, it is essential to mitigate these issues while preserving important features. 3.1.2 Drone Photos Drone imagery provides a new perspective for identifying cultivated land resources. Equipped with high-resolution cameras, drones can capture detailed real-time images over large areas. Unlike traditional satellite imagery, drone photos offer finer spatial detail, enhancing data collection efficiency and providing a more comprehensive dataset for deep learning models, ultimately improving model accuracy and stability. Drones also offer flexibility in adjusting flight altitude and angles, enabling targeted shots and better adaptation to complex cultivated land environments. However, variations in crop growth can cause blurred boundaries in the imagery, making fine segmentation challenging for deep-learning models. 3.2 Sample Annotation Sample annotation is a crucial step in deep learning-based semantic segmentation, as the quality of annotations directly influences model training and prediction outcomes. In this study, a total of 88,268 samples were annotated, including 53,767 images from handheld devices and 34,501 images from drones, details in Table 1 . These annotated datasets are essential for optimizing the model's training process and improving prediction accuracy. Table 1 Sample Annotation Data Quantity Type Category Annotated Samples (Count) Handheld Device Mixed 3056 Building 2478 Construction 8179 Roads 7872 Trees 10333 Water Bodies 5671 cultivated Land 8340 Indoor 8340 Drone Photos Building 16664 Greenhouse 1869 Cultivated land 4136 Photovoltaic Fields 1796 Roads 6855 Water Bodies 3183 In the intelligent verification method for cultivated land surveys based on deep learning semantic segmentation, precise sample annotation is a key element in constructing and training the model. The annotated categories are primarily as follows: Cultivated Land: This category includes crop areas such as rice fields, wheat fields, and cornfields, as well as idle cultivated land and agricultural production facilities like greenhouses (shown in Fig. 6 ). Other categories: These include buildings, water bodies, trees, construction sites, and roads, as shown in Fig. 7 . To ensure the accuracy of the sample annotations, a multi-level review mechanism was established. Each sample was annotated independently by at least two individuals, followed by cross-checks and expert reviews to ensure consistency and precision. Additionally, by observing land-use changes over multiple time periods, more accurate annotations were made for different land categories. A quality threshold was applied to the annotated samples, accompanied by quantitative evaluations and rigorous quality control procedures. During model training, feedback was incorporated to continuously optimize the annotations, ensuring that adjustments were made where necessary. Finally, a sufficient and balanced distribution of samples across all categories was maintained to mitigate the risk of overfitting or underfitting caused by data imbalance. 3.3 Data Preprocessing Data preprocessing is a critical step in developing an efficient and accurate model. By utilizing open-source computer vision libraries like OpenCV, along with the Python programming language, raw data samples were meticulously processed and optimized to extract their potential and enhance model performance. Strict Quality Control of Data: The initial stage of data preprocessing involved removing duplicate, erroneous, or incomplete samples. Following the preliminary annotation, a thorough quality check was conducted on segmentation masks to ensure pixel-level accuracy corresponding to the correct land type, as shown in Figs. 8 and 9 . These checks focused on boundary precision, category accuracy, and the completeness of annotated areas. Any discrepancies were promptly addressed and corrected to ensure high-quality data for training. Label Processing: Annotation tools such as Labelme and COCO were employed to label the images, producing JSON or XML format files that were later converted into segmentation masks. Rigorous quality assurance ensured that each annotation had well-defined boundaries and precise category labels, avoiding mislabeling, omissions, or incomplete areas. The consistency of image and mask sizes was verified, and discrepancies were corrected through cropping, scaling, or other adjustments to maintain alignment and data integrity. Data Augmentation: Data augmentation techniques were applied to diversify the training samples, simulating complex real-world scenarios and thereby expanding the dataset. This process is essential in cultivated land image extraction, where images are often affected by variables such as lighting, seasonal changes, and varying perspectives. The study employed geometric transformations (e.g., flipping, rotation, scaling, cropping) and color space transformations (e.g., brightness, contrast, and saturation adjustments) to generate more diverse samples. These augmentations addressed the issue of limited sample diversity in the cultivated land dataset, enriching the dataset and improving the model's generalization ability. Dataset Splitting: A high-quality dataset is essential for constructing and fine-tuning deep learning models. The dataset was split into training, validation, and test sets; the training set for model learning, the validation set for tuning model parameters, and the test set for evaluating model performance. Given the diversity of cultivated land types and the complexity of environmental conditions, careful dataset splitting was crucial. By ensuring consistent data distribution and preventing overfitting, the model's ability to generalize to unseen data was significantly improved, contributing to the advancement of intelligent algorithms for cultivated land extraction. 3.4 3D Spatial Land Area Calculation The optimized SMR-DeepLabV3 + model must not only achieve accurate image segmentation but also address the challenge of converting image pixel coordinates into actual physical areas. This process involves transforming between multiple coordinate systems, including pixel, physical, camera, and world coordinate systems. The actual area of the target region can be calculated using camera calibration techniques and the pinhole camera model. First, the segmented cultivated land regions in the pixel coordinate system retain boundary information. Then, using the camera's intrinsic and extrinsic parameters, a three-step transformation matrix is applied to derive the boundary points in the world coordinate system. 4. Results and Evaluation 4.1 Model Evaluation Metrics To assess the performance of the enhanced SMR-DeepLabV3 + model in segmenting cultivated land images, two widely used and complementary metrics were employed: the confusion matrix and pixel classification accuracy (PA), as well as Mean Intersection over Union (mIoU) and the Dice coefficient. These metrics provide a comprehensive assessment of the model’s segmentation accuracy and overall performance. 4.1.1 Confusion Matrix The confusion matrix is a valuable tool in classification and semantic segmentation tasks, especially for evaluating the model’s performance across different categories. In a binary classification task, a basic confusion matrix is represented as shown in Table 2 , where True Positive (TP) indicates the number of correctly predicted positive cases, False Negative (FN) indicates the number of actual positive cases incorrectly predicted as negative, False Positive (FP) is the number of actual negative cases incorrectly predicted as positive, and True Negative (TN) represents the number of correctly predicted negative cases. Table 2 Confusion matrix Confusion Matrix Predicted Class 1 Predicted Class 0 Actual Class 1 TP FN Actual Class 0 FP TN 4.1.2 Pixel Classification Accuracy (PA) Pixel classification accuracy (PA) is a straightforward evaluation metric that measures the accuracy of the model’s pixel-level classification. It calculates the proportion of correctly classified pixels for each category. Specifically, PA is computed as the total number of correctly classified pixels divided by the total number of pixels: $$\:\text{P}\text{A}=\left(\frac{\text{T}\text{P}}{\text{T}\text{P}+\text{F}\text{N}}\right).\left(1\right)$$ 4.1.3 Mean Intersection over Union (mIoU) Mean Intersection over Union (mIoU) is a critical metric for image segmentation tasks, particularly for evaluating how well the model delineates segmentation boundaries for each category. It is based on the IoU, which is the ratio of the intersection area between the predicted segmentation and the actual labelled area to their union: $$\:\text{m}\text{I}\text{O}\text{U}=\left(\frac{\text{T}\text{P}}{\text{T}\text{P}+\text{F}\text{P}+\text{F}\text{N}}\right)$$ 2 . mIoU is the average of IoU values across all categories, providing a more comprehensive evaluation than PA, particularly in complex scenes. 4.1.4 Dice Coefficient The Dice coefficient is another important metric for assessing the overlap between the model's predicted segmentation and the actual ground truth labels. It is particularly useful in semantic segmentation tasks and serves as a variant of the Jaccard similarity coefficient. The Dice coefficient is more sensitive to small objects, as it measures the proportion of overlapping areas between predicted and actual regions. The Dice coefficient is computed as: $$\:\text{D}\text{i}\text{c}\text{e}=\left(\frac{2\text{T}\text{P}}{2\text{T}\text{P}+\text{F}\text{P}+\text{F}\text{N}}\right).\:\left(3\right)$$ 4.2 Classification Accuracy of the SMR-DeepLabV3 + Model The training dataset used in this study consists of six key categories typically encountered in cultivated land surveys, including water bodies, hydraulic structures, roads, cultivated land, trees, construction sites, and buildings. This diversity ensures the model is trained on a wide range of features. The improved SMR-DeepLabV3 + model was fine-tuned for cultivated land survey tasks, using pre-trained parameters to initialize the Xception backbone network. Critical hyperparameters, such as learning rate, batch size, and optimizer settings, were carefully monitored and adjusted to achieve optimal performance within a limited training cycle. An independent test set, comprising 10,000 handheld device images and 10,000 drone images, was used to evaluate the model’s performance. Three core evaluation metrics—mIoU, PA, and Dice coefficient—were used to quantify segmentation accuracy for each land type. Table 3 presents the detailed performance of the model across the various land categories. Table 3 Model Performance Across Various Land Types Land Type Subcategory TP FN FP TN mIOU PA Dice Water Bodies Water Surface 1000 6 9 8985 98.54% 99.42% 99.27% Hydraulic Structures 1000 27 21 8952 95.42% 97.4% 97.66% Roads Cement Roads 1000 9 11 8980 98.03% 99.14% 99% Cement Ground 1000 19 26 8955 95.67% 98.13% 97.79% Dirt Roads 1000 71 16 8913 91.95% 93.37% 95.81% Buildings Buildings 1000 13 14 8973 97.36% 98.77% 98.66% Construction Sites Construction Sites 1000 11 11 8978 97.87% 98.95% 98.92% Cultivated Land Crops 1000 29 27 8944 94.69% 97.16% 97.27% Idle Cultivated Land 1000 26 28 8946 94.88% 97.49% 97.37% Greenhouses 1000 22 21 8957 95.89% 97.82% 97.90% Trees Forestland 1000 49 39 8912 91.92% 95.37% 95.79% Orchards 1000 41 50 8909 91.6% 96.09% 95.62% Figure 10 illustrates segmentation results for various categories. Table 4 Model Performance (Mean ± SD) Land Type Subcategory mIOU (Mean ± SD) PA(Mean ± SD) Dice(Mean ± SD) note Water Bodies Water Surface 98.54% ±0.18% 99.42% ±0.12% 99.27% ±0.10% high-consistency categories effectively suppress noise in multimodal data Hydraulic Structures 95.42% ±0.45% 97.40% ±0.35% 97.66% ±0.30% elongated structures cause localized prediction fluctuations. Roads Cement Roads 98.03% ±0.25% 99.14% ±0.15% 99.00% ±0.18% homogeneous regions demonstrate high stability with manageable shadow interference. Cement Ground 95.67% ±0.60% 98.13% ±0.40% 97.79% ±0.45% cracks and stains contribute to localized false detection rate variations. Dirt Roads 91.95% ±0.85% 93.37% ±1.20% 95.81% ±0.75% boundary ambiguity and small target omissions result in greater variance. Buildings Buildings 97.36% ±0.30% 98.77% ±0.25% 98.66% ±0.22% regular shapes exhibit strong robustness, with minor fluctuations in reflective areas. Construction Sites Construction Sites 97.87% ±0.35% 98.95% ±0.28% 98.92% ±0.30% dynamic objects (e.g., vehicles) cause localized instability. Cultivated Land Crops 94.69% ±0.70% 97.16% ±0.55% 97.27% ±0.60% differences in growth stages lead to intra-class feature variations. Idle Cultivated Land 94.88% ±0.65% 97.49% ±0.50% 97.37% ±0.55% misclassification rates fluctuate near grassland boundaries. Greenhouses 95.89% ±0.50% 97.82% ±0.40% 97.90% ±0.45% film reflections are manageable, but support structure shadows affect localized stability. Trees Forestland 91.92% ±1.10% 95.37% ±0.90% 95.79% ±0.85% canopy gaps and lighting changes cause significant fluctuations. Orchards 91.60% ±1.20% 96.09% ±0.95% 95.62% ±1.00% variations in planting density and species diversity result in segmentation inconsistencies. Table 4 illustrate the standard deviation values to measure consistency. The node sections are error analysis. The experimental results demonstrate that the improved SMR-DeepLabV3 + model performs exceptionally well in cultivated land survey image segmentation tasks. It achieves high overall classification accuracy and demonstrates precise segmentation across various categories. These quantitative findings offer valuable insights into the model's capabilities, laying a solid foundation for further optimization of cultivated land image segmentation techniques. 4.3 Comparative Analysis with other models To quantify the improvements of SMR-DeepLabV3+, this study compares it with other models such as U-Net, SegNet, and PSPNet. Table 4 provides a detailed comparison of each model's performance in terms of mIoU, PA, and Dice. Table 5 comparative analysis with other models Land Type Subcategory model mIOU PA Dice Water Bodies Water Surface U-Net 96.80% 98.50% 98.30% SegNet 94.20% 96.90% 96.50% PSPNet 97.10% 98.80% 98.50% SMR-DeepLabV3+ 98.54% 99.42% 99.27% Hydraulic Structures U-Net 92.50% 95.80% 95.20% SegNet 89.30% 93.10% 92.70% PSPNet 93.80% 96.20% 95.90% SMR-DeepLabV3+ 95.42% 97.4% 97.66% Roads Cement Roads U-Net 96.20% 98.00% 97.80% SegNet 92.70% 95.40% 95.10% PSPNet 96.90% 98.50% 98.20% SMR-DeepLabV3+ 98.03% 99.14% 99% Cement Ground U-Net 93.10% 96.20% 95.60% SegNet 88.50% 92.70% 91.90% PSPNet 94.30% 96.80% 96.20% SMR-DeepLabV3+ 95.67% 98.13% 97.79% Dirt Roads U-Net 88.40% 90.10% 92.30% SegNet 83.60% 86.90% 88.50% PSPNet 89.70% 91.50% 93.80% SMR-DeepLabV3+ 91.95% 93.37% 95.81% Buildings Buildings U-Net 95.30% 97.20% 97.00% SegNet 91.80% 94.60% 94.20% PSPNet 96.10% 97.90% 97.60% SMR-DeepLabV3+ 97.36% 98.77% 98.66% Construction Sites Construction Sites U-Net 94.70% 96.80% 96.50% SegNet 90.10% 93.20% 92.90% PSPNet 95.20% 97.10% 96.90% SMR-DeepLabV3+ 97.87% 98.95% 98.92% Cultivated Land Crops U-Net 91.50% 94.30% 94.00% SegNet 87.20% 90.80% 90.50% PSPNet 92.80% 95.10% 94.90% SMR-DeepLabV3+ 94.69% 97.16% 97.27% Idle Cultivated Land U-Net 91.80% 94.70% 94.30% SegNet 87.60% 91.20% 90.90% PSPNet 93.10% 95.40% 95.10% SMR-DeepLabV3+ 94.88% 97.49% 97.37% Greenhouses U-Net 93.20% 95.60% 95.30% SegNet 89.40% 92.80% 92.10% PSPNet 94.50% 96.30% 96.00% SMR-DeepLabV3+ 95.89% 97.82% 97.90% Trees Forestland U-Net 88.70% 92.50% 92.10% SegNet 84.10% 88.30% 87.90% PSPNet 89.90% 93.10% 92.80% SMR-DeepLabV3+ 91.92% 95.37% 95.79% Orchards U-Net 88.30% 93.20% 92.00% SegNet 83.90% 88.90% 87.50% PSPNet 89.50% 93.80% 92.60% SMR-DeepLabV3+ 91.6% 96.09% 95.62% The performance comparison results in Table 5 demonstrate that SMR-DeepLabV3 + consistently outperforms other models, such as U-Net, SegNet, and PSPNet, across all land cover categories. SMR-DeepLabV3 + achieves good accuracy in terms of mIoU, PA, and Dice scores, particularly excelling in complex categories with intricate boundaries and challenging feature differentiation, such as hydraulic structures, dirt roads, and greenhouses. PSPNet exhibits performance comparable to or even exceeding SMR-DeepLabV3 + in some subcategories, such as water surfaces, dirt roads, and buildings, showcasing its potential in feature extraction and boundary delineation. U-Net also demonstrates stable performance in large-area land types like "water surfaces" and "hydraulic structures," highlighting its ability to capture extensive regional features. While SegNet generally underperforms compared to the other models, it still demonstrates competitive performance in PA and Dice scores for "buildings" and "dirt roads." This highlights SMR-DeepLabV3+'s boundary recognition and feature extraction capabilities, reinforcing its robustness in both natural and artificial environments while showcasing its performance, particularly in complex land cover types, for remote sensing image classification tasks. 4.4 Cultivated Land Monitoring Analysis To assess the effectiveness and accuracy of the SMR-DeepLabV3 + model in cultivated land recognition and area calculation, land plots from Zhaohua District, Guangyuan City, Sichuan Province, were selected as the case study area for 2023. This region, characterized by diverse cultivated land types and typical agricultural activities, is ideal for validating the model's performance. A total of 3,466 patches were randomly sampled, ensuring a diverse dataset for comprehensive model evaluation. To scientifically assess the accuracy of intelligent verification, this study established a robust evaluation criterion after multiple rounds of data validation and comparison: if at least 65% of the area identified by the model matches the pre-labeled land class information for changes, the land classification for that patch is considered accurate. This criterion accounts for potential errors in surveys while maintaining the reliability and practicality of the recognition results. Through statistical analysis of the recognition results for the 3,466 patches, the land was categorized into three types: paddy fields, irrigated land, and dryland. The recognition statistics are presented in Table 4 . Table 6 Cultivated Land Recognition Statistics Land Class Subclass Misidentified as Other Classes Total Patches Concordance Rate Cultivated Land Paddy Field 43 1329 80.3% Irrigated Land 62 1130 81.7% Dryland 23 1007 86.1% The paper also presents a portion of the verification results, which are divided into two categories: compliant and non-compliant patches. In the sample images, the yellow trapezoidal area represents the photo coverage, the blue contour outlines the area of the change patch, and the green region indicates the AI-recognized area. The ratio of the area within the photo coverage that aligns with the pre-change land class in the change patch is referred to as the compliance rate. Figures 11 and 12 respectively show samples of compliant and non-compliant patch identifications. The recognition results indicate significant improvements in land class identification, confirming the effectiveness of the SMR-DeepLabV3 + model. For the mismatches, two key factors were identified: first, human interference, such as the submission of distorted patch information; second, seasonal variations, particularly during the spring and summer months when dense vegetation growth can obscure surface features, thereby affecting the accuracy of land class determination. 5. Conclusion and Discussion Grain security is crucial to sustainable development, making the protection of cultivated land a top priority. Accurate land change surveys and verification processes are essential to maintain reliable cultivated land information. While handheld devices and drone photos provide rich data support, complex factors such as lighting conditions and angles pose challenges for traditional semantic segmentation models. To address these issues, this study improves the DeepLabV3 + model by embedding an attention mechanism, employing multimodal fusion techniques, and incorporating regularization terms into the loss function, leading to the development of the SMR-DeepLabV3 + model. A multimodal sample dataset was constructed, consisting of 88,268 annotated samples, including 53,767 photos from handheld devices and 34,501 from drones. The model’s classification accuracy was validated using PA and mIoU, with classification accuracy consistently exceeding 90%. Through comparative experiments, SMR-DeepLabV3 + achieves better performance in mIoU, PA, and Dice metrics compared to models such as U-Net, SegNet, and PSPNet, particularly excelling in boundary recognition and feature extraction.Additionally, a 3D spatial land area calculation method was proposed to convert pixel coordinates into physical areas, using coordinate transformation and camera calibration to calculate changes in patch areas. The SMR-DeepLabV3 + model was tested in Zhaohua District, Guangyuan City, Sichuan Province (3,346 patches in 2023), achieving recognition rates of 80.3%, 81.7%, and 86.1% for paddy fields, irrigated land, and dryland, respectively. The model demonstrated high accuracy in semantic segmentation, showing significant advantages in improving recognition accuracy and efficiency in handheld device photos. The model also has broad application potential in cultivated land verification, promoting more refined management of cultivated land resources. While this study has achieved promising results, several challenges remain in practical applications: 1. High Sample Production Costs: Data annotation is labor-intensive, and the model may struggle in extreme, unseen scenarios. 2. Model Interpretability: The decision-making process of the model lacks transparency, necessitating further improvements to enhance the interpretability of its predictions. To overcome these challenges, future work plans to introduce domain adaptation techniques, such as multi-domain feature alignment based on transfer learning, cross-view feature fusion with attention enhancement, dynamic domain adaptation, and self-supervised learning, to improve the model's generalization across different geographic regions and promote the automatic classification and intelligent analysis of land data. Such methods will improve the efficiency of online verification processes and provide robust, scientific decision support for land resource management, ecological protection, urban and rural planning, and other fields. Ultimately, these approaches will enhance both the efficiency and accuracy of data analysis. Declarations Clinical trial number: not applicable. Data supporting the findings of this study are available from the corresponding author upon reasonable request. Ethics, Consent to Participate, and Consent to Publish declarations: not applicable. The authors declare no competing interests. Funding: This research was funded by “Integrated dynamic monitoring of natural resources, Department of Natural Resources of Sichuan Province”, grant number N5100012024001267. Acknowledgments: All authors thanks the data sharing of Zhaohua District, Guangyuan City. Author's email address Zheng Fang, [email protected] ; Yuhang Gong, [email protected] ; Heli Zhu, [email protected] ; Canyang Shi, [email protected] ; Zhijia Gong, [email protected] ; Lin Tian, [email protected] . Author Contribution Z.F.: Conceptualization, methodology, formal analysis, writing—original draft preparation, supervision,project administration, funding acquisition;Y.G.: methodology, software, formal analysis, writing—review and editing, visualization;H.Z.:validation, investigation,resources, data curation, writing—review and editing;C.S.:validation, investigation,resources, data curation, writing—review and editing;Z.G.:validation, investigation,resources, data curation, writing—review and editing;L.T.:validation, investigation,resources, data curation, writing—review and editing;All authors have read and agreed to the published version of the manuscript. References Dong Z Y， Li J H， Zhang J， Yu J Q and An S. 2023. Cultivated land extraction from high-resolution remote sensing images based on BECU-Net model with edge enhancement. National Remote Sensing Bulletin， 27（12）：2847-2859 DOI： 10.11834/jrs.20222268. DU Guo-ming,GAI Zhao-xue*,WANG Hong-yan.Theoretical Explanation and Research Framework of Cultivated-land Fragmentation in China[J].Journal of Earth Sciences and Environment,2021,43(06):997-1008.[doi:10.19814/j.jese.2021.07029] ZHANG Xinchang, HUANG Jianfeng, NING Ting. Progress and Prospect of Cultivated Land Extraction from High-Resolution Remote Sensing Images[J]. Geomatics and Information Science of Wuhan University, 2023, 48(10): 1582-1590. DOI: 10.13203/j.whugis20230114 LI Liangjun; MU Zhe; Lv Lei; MIAODAN. Application analysis of “Internet+” evidence technology in land survey[J]. Surveying, Mapping and Spatial Geographic Information,2024,(01):200-202+205. Feng Siyuan. Application of UAV inclined photogrammetry in the third national land survey [J]. Mapping and Spatial Geographic Information, 2023, 46 (07): 154-156+160. Xiao H; Dong Shengguang; Tang Zhiqing; Tang Xianlong. Exploration and practice of construction of command center for natural resources investigation and monitoring in Hunan Province[J]. Natural Resources Informatization,2023,(03):9-14. ZHANG Yi;JI Bo;JING Ruan. Practice and thinking of natural resources intelligent sensing model based on tower-based video monitoring[J]. Land Resources Informatization,2021,(05):8-14. ZHANG Yichao, ZHENG Xiangtao, LU Xiaoqiang. Hyperspectral image classification method based on hierarchical transformer network[J]. Acta Geodaetica et Cartographica Sinica, 2023, 52(7): 1139-1147. DU Peijun, ZHANG Wei, ZHANG Peng, LIN Cong, GUO Shanchuan, HU Zezhou. A capsule network for hyperspectral image classification employing spatial-spectral feature[J]. Acta Geodaetica et Cartographica Sinica, 2023, 52(7): 1090-1104. SUN Genyun, WANG Xin, AN Na, ZHANG Aizhu. A method for large-scale and high-resolution impervious surface extraction based on multi-source remote sensing and deep learning[J]. Acta Geodaetica et Cartographica Sinica, 2023, 52(2): 272-282. Peng L K，Liu L C，Chen X H，Chen J，Cao X and Qiu Y A. 2021. Generalization ability of cloud detection network for satellite imagery based on DeepLabv3+. National Remote Sensing Bulletin， 25（5）：1169-1186 DOI： 10.11834/jrs.20210061. ZHENG Kai, LI Jiansheng, YANG Jianfeng, OUYANG Wen, WANG Gaojie, ZHANG Xun. A cloud and snow detection method of TH-1 image based on combined ResNet and DeepLabV3+[J]. Acta Geodaetica et Cartographica Sinica, 2020, 49(10): 1343-1353. CHEN Menghua,ZHANG Tongyun,ZHOU Zixiang,et al. A high-precision batch automatic extraction method for buildings based on DeepLabV3+[J/OL]. Journal of Surveying and Mapping Science and Technology,1-6[2024-09-11] JIANG Baode, HANG Wei, XU Shaofen, WU Yong. Multi-scale building instance refinement extraction from remote sensing images by fusing with decentralized adaptive attention mechanism[J]. Acta Geodaetica et Cartographica Sinica, 2023, 52(9): 1504-1514. LIN Yunhao, WANG Yanjun, LI Shaochun, CAI Hengfan. A coupled DeepLab and Transformer approach for fine classification of crop cultivation types in remote sensing[J]. Acta Geodaetica et Cartographica Sinica, 2024, 53(2): 353-366. ZHOU Xinxin, WU Yanlan, LI Mengya, ZHENG Zhiteng. Automatic Vegetation Extraction Method based on Feature Separation Mechanism with Deep Learning[J]. Journal of Geo-information Science, 2021, 23(9): 1675-1689 https://doi.org/10.12082/dqxxkx.2021.200641 Liu C T， Feng Q L， Liu J T， Wang Y， Shi T G， Li Y， Gong J H and Zhao H H. 2022. Urban green plastic cover extraction and spatial pattern changes in Jinan city based on DeepLabv3+ semantic segmentation model. National Remote Sensing Bulletin， 26（12）：2518-2530 DOI： 10.11834/jrs.20220101. Wu Q Q，Wang S，Wang B and Wu Y L. 2022. Road extraction method of high-resolution remote sensing image on the basis of the spatial information perception semantic segmentation model. National Remote Sensing Bulletin， 26（9）： 1872-1885 DOI： 10.11834/jrs.20210021. TIAN Fuyou，CAO Yupei，ZHAO Hang，WU Bingfang，ZENG Hongwei，LIU Yazhou，QIN Xingli，ZHANG Miao，ZHU Liang，ZHU Weiwei. XXXX. Agricultural Field Segmentation using Spatial Attention Mechanism and Multi-task Learning Strategy. National Remote Sensing Bulletin， XX（XX）：1-15 DOI： 10.11834/jrs.20243191. TAN Min, LIN Huijing, HAO Ming. Prior Knowledge Guided Deep Learning for Monitoring Buildings and Greenhouses within Cultivated Land[J]. Journal of Geo-information Science, 2023, 25(11): 2293-2302 https://doi.org/10.12082/dqxxkx.2023.230315 Zhu, X.X., et al. (2017) Deep Learning in Remote Sensing: A Comprehensive Review and List of Resources. IEEE Geoscience and Remote Sensing Magazine, 5, 8-36. Li Y, Zhang H, Xue X, et al. Deep learning for remote sensing image classification: A survey[J]. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 2018, 8(6): e1264. Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation[C]//Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. Springer International Publishing, 2015: 234-241. Kampffmeyer M, Salberg A B, Jenssen R. Semantic segmentation of small objects and modeling of uncertainty in urban remote sensing images using deep convolutional neural networks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition workshops. 2016: 1-9. Chen L C, Zhu Y, Papandreou G, et al. Encoder-decoder with atrous separable convolution for semantic image segmentation[C]//Proceedings of the European conference on computer vision (ECCV). 2018: 801-818. Pan Y, Zhang L. Dual attention deep learning network for automatic steel surface defect segmentation[J]. Computer‐Aided Civil and Infrastructure Engineering, 2022, 37(11): 1468-1487. Gu H, Wang P, Li Y, et al. MFP-DeepLabv3+: A Multi-scale Feature Fusion and Parallel Attention Network for Enhanced Bone Metastasis Segmentation[J]. Traitement du Signal, 2024, 41(2). Wang Y, Yang L, Liu X, et al. An improved semantic segmentation algorithm for high-resolution remote sensing images based on DeepLabv3+[J]. Scientific Reports, 2024, 14(1): 9716. Ullah M, Bais A, Wist T. TranSEF: Transformer Enhanced Self-Ensemble Framework for Damage Assessment in Canola Crops[J]. IEEE Transactions on AgriFood Electronics, 2024. Chen, Liang-Chieh, et al. Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), 2018, pp. 801-818. François Chollet. Xception: Deep Learning with Depthwise Separable Convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 1251-1258. Keskar, Nitish Shirish, and Richard Socher. Improving Generalization Performance by Switching from Adam to SGD. arXiv preprint arXiv:1712.07628, 2017. Loshchilov, Ilya, and Frank Hutter. Decoupled Weight Decay Regularization. In International Conference on Learning Representations (ICLR), 2019. Goodfellow I. Deep learning[J]. 2016. Hu, Jie, et al. "Squeeze-and-Excitation Networks." IEEE Transactions on Pattern Analysis and Machine Intelligence (2018). Baltrušaitis, T., et al. "Multimodal Machine Learning: A Survey and Taxonomy." IEEE Transactions on Pattern Analysis and Machine Intelligence (2019). Li, Z., et al. "Deep Learning for Remote Sensing Image Classification: A Comprehensive Review." IEEE Transactions on Geoscience and Remote Sensing (2020). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 01 Jul, 2025 Read the published version in Discover Sustainability → Version 1 posted Editorial decision: Revision requested 06 May, 2025 Reviews received at journal 15 Apr, 2025 Reviewers agreed at journal 14 Apr, 2025 Reviewers agreed at journal 11 Apr, 2025 Editor assigned by journal 09 Apr, 2025 Reviewers invited by journal 09 Apr, 2025 Submission checks completed at journal 08 Apr, 2025 First submitted to journal 24 Mar, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-5770339","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":440520099,"identity":"80c5029f-3eb4-44c9-a891-791e54510447","order_by":0,"name":"Zheng Fang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA40lEQVRIiWNgGAWjYDAC5gNQBnsDkDCwIEILWwKUwQPSayBBihYJMIMILfJtzM8efmGoleOXfH51w48CCQb+9u4EvFoY29jMjWUYjhtLzs4pu9kDdJjEmbMb8Gphlm8wk5ZgOJa44XZO2g0eoBYDiVz8WtjY2L+BtNRvuHkm7eYfYrTwsPGYSX5gqEkwuMF+7DZRtkiw8ZRJMxgcMJzZk8N2W8ZAgoegX+Tb2LdJ/qiok+dnP/7s5ps/NnL87b34tYAAM4/BYZAbDcAuJagcBBh/MNQBKfYHRKkeBaNgFIyCkQcAbOtA52lVWbgAAAAASUVORK5CYII=","orcid":"","institution":"Sichuan Institute of Land Science and Technology","correspondingAuthor":true,"prefix":"","firstName":"Zheng","middleName":"","lastName":"Fang","suffix":""},{"id":440520100,"identity":"c1eaf31f-3569-4f84-95b1-59be50734841","order_by":1,"name":"Yuhang Gong","email":"","orcid":"","institution":"Sichuan Institute of Land Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Yuhang","middleName":"","lastName":"Gong","suffix":""},{"id":440520101,"identity":"e6a57225-1a71-4571-a79a-5d3fff604748","order_by":2,"name":"Heli Zhu","email":"","orcid":"","institution":"Sichuan Institute of Land Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Heli","middleName":"","lastName":"Zhu","suffix":""},{"id":440520102,"identity":"e6ef4e70-65e8-4c3b-b688-db2b101d1b31","order_by":3,"name":"Canyang Shi","email":"","orcid":"","institution":"Sichuan Institute of Land Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Canyang","middleName":"","lastName":"Shi","suffix":""},{"id":440520104,"identity":"d3dbeed6-88aa-4e66-b2dd-91c800ce02d1","order_by":4,"name":"Zhijia Gong","email":"","orcid":"","institution":"Sichuan Institute of Land Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Zhijia","middleName":"","lastName":"Gong","suffix":""},{"id":440520107,"identity":"2edd17c6-b031-48e4-9e31-bbd7eecc011b","order_by":5,"name":"Lin Tian","email":"","orcid":"","institution":"Sichuan Institute of Land Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Lin","middleName":"","lastName":"Tian","suffix":""}],"badges":[],"createdAt":"2025-01-06 03:38:06","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5770339/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5770339/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1007/s43621-025-01461-w","type":"published","date":"2025-07-01T15:58:27+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":80306443,"identity":"06c85328-8528-4bbc-8beb-a61b8090c2ca","added_by":"auto","created_at":"2025-04-10 10:23:17","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":52004,"visible":true,"origin":"","legend":"\u003cp\u003eDeepLabV3+ Model Architecture\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5770339/v1/d26a6fddd6591e2ff9b9aa14.jpg"},{"id":80305980,"identity":"eeda0f1c-e734-49df-b1d1-1fb324c2e365","added_by":"auto","created_at":"2025-04-10 10:15:17","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":71412,"visible":true,"origin":"","legend":"\u003cp\u003eXception Architecture.\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5770339/v1/5620f47a7f57c7bb9a0cc11e.jpg"},{"id":80306446,"identity":"c69998c4-45a9-460a-9c40-fe2d3fa56359","added_by":"auto","created_at":"2025-04-10 10:23:17","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":46867,"visible":true,"origin":"","legend":"\u003cp\u003eSMR-DeepLabV3+ Network Architecture\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5770339/v1/ce6c80aebbc0de6dd9ac0a93.jpg"},{"id":80307404,"identity":"a97ab424-423d-46ed-8373-1afd39eff5a8","added_by":"auto","created_at":"2025-04-10 10:31:17","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":24519,"visible":true,"origin":"","legend":"\u003cp\u003eSENet Attention Mechanism Module.\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5770339/v1/5f94efb4dae59fb99c5488f1.jpg"},{"id":80305982,"identity":"8a506ec9-315c-4ab7-beb9-9b9a0cf5472f","added_by":"auto","created_at":"2025-04-10 10:15:17","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":52089,"visible":true,"origin":"","legend":"\u003cp\u003eMultimodal Fusion Schematic\u003csup\u003e[36]\u003c/sup\u003e.\u003c/p\u003e","description":"","filename":"5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5770339/v1/a0e664f1de471d9afb06c86c.jpg"},{"id":80307405,"identity":"02c28976-d3b1-4dd6-b74a-171bdac62615","added_by":"auto","created_at":"2025-04-10 10:31:17","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":102360,"visible":true,"origin":"","legend":"\u003cp\u003eDrone image example of Cultivated Land\u003c/p\u003e","description":"","filename":"6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5770339/v1/69a73c78fddec1ebc2337427.jpg"},{"id":80305997,"identity":"0cf674e9-a012-4418-b54c-1acf2fe4c04e","added_by":"auto","created_at":"2025-04-10 10:15:17","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":182594,"visible":true,"origin":"","legend":"\u003cp\u003eCultivated Land Annotation Example.\u003c/p\u003e","description":"","filename":"7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5770339/v1/c2561a7f3cfa75fbee76c61d.jpg"},{"id":80306452,"identity":"75c0415b-3e16-4d95-b9ee-b457ed36185d","added_by":"auto","created_at":"2025-04-10 10:23:17","extension":"jpg","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":121703,"visible":true,"origin":"","legend":"\u003cp\u003eOther Annotation Examples.\u003c/p\u003e","description":"","filename":"8.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5770339/v1/4500d9443bf76d8592cab64a.jpg"},{"id":80305989,"identity":"d6c863e8-6469-4fb8-ba47-7d14f3256638","added_by":"auto","created_at":"2025-04-10 10:15:17","extension":"jpg","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":102943,"visible":true,"origin":"","legend":"\u003cp\u003eSample Dataset.\u003c/p\u003e","description":"","filename":"9.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5770339/v1/e40fb539f974a6ae10b4c30c.jpg"},{"id":80305990,"identity":"3a50d626-30d1-431d-92d8-ca18206a8ce9","added_by":"auto","created_at":"2025-04-10 10:15:17","extension":"jpg","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":51595,"visible":true,"origin":"","legend":"\u003cp\u003eCorresponding Label Dataset.\u003c/p\u003e","description":"","filename":"10.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5770339/v1/c5bb94656eefea4f698814eb.jpg"},{"id":80306007,"identity":"a4b30c9c-bd11-44d4-aecd-bb8794851b18","added_by":"auto","created_at":"2025-04-10 10:15:18","extension":"jpg","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":233140,"visible":true,"origin":"","legend":"\u003cp\u003eSegmentation Results for Various Categories.\u003c/p\u003e","description":"","filename":"11.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5770339/v1/072fa292f2140edb7bf1f44d.jpg"},{"id":80306002,"identity":"66d29ed3-4d8d-44c7-b6a4-e0fd8ed6fe9a","added_by":"auto","created_at":"2025-04-10 10:15:17","extension":"jpg","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":77348,"visible":true,"origin":"","legend":"\u003cp\u003eCompliant Patch Identification.\u003c/p\u003e","description":"","filename":"12.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5770339/v1/e6ef51787ca3b1e86b99f07b.jpg"},{"id":80306459,"identity":"b018b82c-b624-499a-9ab8-7c01059ff8da","added_by":"auto","created_at":"2025-04-10 10:23:17","extension":"jpg","order_by":13,"title":"Figure 13","display":"","copyAsset":false,"role":"figure","size":83697,"visible":true,"origin":"","legend":"\u003cp\u003eNon-compliant Patch Identification.\u003c/p\u003e","description":"","filename":"13.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5770339/v1/fea34d2ce39d7429fb21d530.jpg"},{"id":86179176,"identity":"040c9e80-ddcd-4d92-a3bc-c271b29671ac","added_by":"auto","created_at":"2025-07-07 16:16:48","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2509301,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5770339/v1/96b1e796-0c87-488f-a75e-2454931db646.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"SMR-DeepLabV3+: A Refined Model for Cultivated Land Extraction from Cross-View Imagery Using an Attention Mechanism","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eCultivated land is a fundamental resource essential for human survival and agricultural production. The protection of cultivated land is intrinsically linked to food security, which is vital for the stable development of economics, politics, and societies[1]. As urbanization accelerates and pressure on land resources increases, timely monitoring of cultivated land dynamics has become crucial, particularly for detecting land loss and illegal land use[2]. Establishing a precise, comprehensive, and dynamic monitoring system for natural resources is key to promoting sustainable development.\u003c/p\u003e \u003cp\u003eAs cultivated land monitoring systems evolve, traditional paper-based surveys have gradually transitioned to integrated air-space-ground monitoring[3]. This approach leverages high-resolution remote sensing images[4], drone imagery[5], and ground-based mobile measurement equipment[6\u0026ndash;7], making cultivated land monitoring more comprehensive and efficient. With the rapid advancements in remote sensing and artificial intelligence (AI), cultivated land extraction methods have evolved from traditional pixel-based and object-oriented classification algorithms to intelligent, deep learning-based technologies[3].\u003c/p\u003e \u003cp\u003eDeep learning models, renown for their robust feature extraction capabilities and adaptability to complex environments, have been widely applied to multi-scale remote sensing image classification tasks[8\u0026ndash;9]. These applications include cloud-snow distinction[10\u0026ndash;12], building extraction[13\u0026ndash;14], crop identification[15\u0026ndash;16], dust cover detection[17], and road extraction[18]. In the context of cultivated land monitoring, deep learning has significantly improved the accuracy of land identification and segmentation, providing strong technical support for precise monitoring and management, while advancing the field towards more intelligent and automated solutions[19\u0026ndash;20].\u003c/p\u003e \u003cp\u003eSemantic segmentation algorithms based on deep learning offer an innovative approach to cultivated land extraction and monitoring. Convolutional neural networks (CNNs), the core technology behind these algorithms, automatically extract spatial and spectral features from images, significantly improving classification accuracy, particularly in multi-scale feature extraction[21]. Popular models such as AlexNet, VGG, and ResNet have been widely applied to land cover classification[22], demonstrating high robustness and generalization ability. Fully Convolutional Networks (FCN) and U-Net architectures are commonly employed in semantic segmentation tasks and have quickly gained traction in remote sensing image classification[23]. DeepLabV3+, which incorporates atrous convolution and Atrous Spatial Pyramid Pooling (ASPP), has been particularly effective in identifying complex land boundaries and improving classification accuracy and refinement[24]. Due to its network structure optimizations, DeepLabV3\u0026thinsp;+\u0026thinsp;has demonstrated high accuracy across standard datasets[25].\u003c/p\u003e \u003cp\u003eImprovements have been proposed for the DeepLabV3\u0026thinsp;+\u0026thinsp;model[26\u0026ndash;29], such as DAF-DeepLabV3+, MFP-DeepLabV3+,etc. Building upon these advancements, this study introduces the SMR-DeepLabV3\u0026thinsp;+\u0026thinsp;model to enhance inference speed, reduce memory consumption, and lower data preparation costs. The SMR-DeepLabV3\u0026thinsp;+\u0026thinsp;model is adopted as the base model for cultivated land extraction and its capabilities are enhanced by integrating the SENet attention mechanism. This addition improves the network's ability to focus on local details and enhances overall feature learning. Furthermore, the model incorporates cross-perspective information, utilizing drone images, remote sensing imagery, and mobile phone photos to build a more diverse and comprehensive sample dataset. This enables a more holistic understanding of cultivated land environments. The model's effectiveness was validated using cultivated land patches, with results showing that it accurately identifies cultivated land areas and types. These findings support the refined management of cultivated land resources.\u003c/p\u003e"},{"header":"2. SMR-DeepLabV3 + Model Integration with Attention Mechanism","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Overview of DeepLabV3+\u003c/h2\u003e \u003cp\u003eDeepLabV3\u0026thinsp;+\u0026thinsp;is an enhanced version of the DeepLab series of semantic segmentation algorithms. It builds upon the strengths of DeepLabV3, introducing key innovations and optimizations to overcome common challenges in semantic segmentation, such as the loss of boundary details and the integration of multi-scale features, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eDeepLabV3\u0026thinsp;+\u0026thinsp;utilizes an Encoder-Decoder[30] architecture that significantly improves segmentation accuracy by combining high-level semantic features with low-level spatial information, particularly enhancing boundary detail preservation. The decoder employs multi-path feature fusion, integrating features from different stages of the encoder with those generated by the decoder progressively, layer by layer. This approach helps retain fine-resolution details while capturing multi-scale contextual information, leading to more precise object boundary recognition and localization.\u003c/p\u003e \u003cdiv id=\"Sec4\" class=\"Section3\"\u003e \u003ch2\u003e2.1.1 Lightweight Backbone Network (Xception)\u003c/h2\u003e \u003cp\u003eDeepLabV3\u0026thinsp;+\u0026thinsp;employs Xception as its backbone network. Introduced by Google in 2017[31], it is a lightweight architecture that replaces standard convolution operations with depthwise separable convolutions, which consist of spatial convolutions followed by channel-wise convolutions. This approach significantly reduces the number of parameters and computational complexity, while enhancing the model's capacity for feature representation. Xception (in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) is composed of 36 convolutional layers, organized into three stages: Entry, Middle, and Exit flows. These stages handle down-sampling, feature learning, and feature aggregation, respectively, providing the foundational structure for feature extraction in DeepLabV3+. The use of residual connections between convolutional layers helps preserve important feature information, allowing the network to maintain a balance between computational efficiency and high performance. This makes Xception particularly well-suited for high-precision semantic segmentation tasks, where accurate feature extraction is critical.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section3\"\u003e \u003ch2\u003e2.1.2 Combined Stochastic Gradient Descent and Adam Optimizers\u003c/h2\u003e \u003cp\u003eIn deep learning, optimizers adjust model parameters to minimize the loss function, directly impacting model performance. DeepLabV3\u0026thinsp;+\u0026thinsp;adopts a combined strategy that leverages the strengths of both Stochastic Gradient Descent (SGD) and Adam optimizers. SGD excels in handling non-convex optimization problems and has good generalization ability, while Adam achieves faster convergence through adaptive learning rates[32]. SGD, one of the fundamental optimization algorithms in deep learning, updating weights based on gradients computed from each training batch, making it suitable for achieving robust generalization. In contrast, Adam accelerates convergence by combining momentum-based gradient descent with adaptive learning rate adjustments, allowing it to perform well in scenarios involving sparse data or noisy gradients[33].\u003c/p\u003e \u003cp\u003eBy combining these optimizers, DeepLabV3\u0026thinsp;+\u0026thinsp;benefits from the rapid convergence of Adam and the fine-tuning capabilities of SGD, which often yields better final results. This SGD\u0026thinsp;+\u0026thinsp;Adam strategy alternates or stages the two optimizers, using Adam for early-stage optimization and then switching to SGD for refined adjustments. This approach improves both training efficiency and the model's generalization ability, leading to better overall performance.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section3\"\u003e \u003ch2\u003e2.1.3 Loss Function (Cross-Entropy Loss)\u003c/h2\u003e \u003cp\u003eThe loss function is a fundamental component in machine learning and deep learning, used to quantify the error between the model\u0026rsquo;s predictions and the actual values. Selecting an appropriate loss function is crucial for optimizing model performance. Cross-entropy loss is commonly used in classification tasks, as it measures the divergence between the predicted probability distribution and the true labels[34]. A lower cross-entropy value signifies better model performance by indicating a closer match between the predicted and true distributions. Cross-entropy loss is often used alongside the Softmax function, which normalizes the output into a probability distribution, allowing cross-entropy to compute the error more effectively by comparing the true labels to the predicted probabilities. This combination ensures that the model can optimize its parameters more efficiently during training.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Optimized SMR-DeepLabV3\u0026thinsp;+\u0026thinsp;Model\u003c/h2\u003e \u003cp\u003eThis paper introduces the SMR-DeepLabV3\u0026thinsp;+\u0026thinsp;model, which builds upon and optimizes the DeepLabV3\u0026thinsp;+\u0026thinsp;architecture at multiple levels. The architecture of the SMR-DeepLabV3\u0026thinsp;+\u0026thinsp;model is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003e2.2.1 Attention Mechanism Integration (Squeeze-and-Excitation Networks)\u003c/h2\u003e \u003cp\u003eTo improve feature representation and enhance semantic segmentation performance, this paper integrates the Squeeze-and-Excitation Networks (SENet) module (as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e), an attention mechanism, into the DeepLabV3\u0026thinsp;+\u0026thinsp;model. This introduces an efficient and novel approach to feature enhancement. The core idea behind SENet is to assign different weights to each feature channel, allowing the model to highlight critical features while suppressing irrelevant background noise\u003csup\u003e[35]\u003c/sup\u003e. SENet consists of two steps: \"Squeeze\" and \"Excitation.\" First, global average pooling is applied to compress spatial information. Then, fully connected layers and a Sigmoid activation function generate weight values for each channel, optimizing feature extraction by emphasizing essential channels and downplaying less relevant ones.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe SENet module evaluates the significance of feature channels at a global scale, enabling the model to prioritize important features by assigning adaptive weights. This attention mechanism complements DeepLabV3+'s multi-scale feature fusion strategy, significantly improving the efficiency of blending features across multiple layers. By incorporating SENet, the model effectively integrates shallow spatial details with deeper semantic information, addressing the previous limitations of insufficient deep feature blending. This enhancement leads to better preservation of fine-grained details and a more accurate understanding of global semantic structures, ultimately improving quality and accuracy of image segmentation.In the extraction of cultivated land, SENet's attention mechanism enhances the model's ability to capture subtle differences between various types of cultivated land, thereby improving segmentation accuracy.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003e2.2.2 Multimodal Information Fusion\u003c/h2\u003e \u003cp\u003eMultimodal fusion refers to the integration of two or more information modalities, each representing different sources or forms of information (as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e)\u003csup\u003e[36]\u003c/sup\u003e. These modalities can exhibit redundancy or complementarity, and effectively combining them results in richer and more comprehensive feature representations. In multimodal research, modality fusion enhances the understanding and expression of complex scenes by leveraging the strengths of different modalities, enabling the capture of diverse multidimensional information. This approach has become an essential method for improving model performance in various tasks.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn the DeepLabV3\u0026thinsp;+\u0026thinsp;model, multimodal fusion incorporates images from different perspectives, including mobile phone photos, drone images, and remote sensing data[37]. Each modality provides distinct advantages: mobile phone photos provide rich details, drone images offer a broader field of view, and remote sensing images capture macro-environmental context. By exploiting the complementarity of these modalities, the model significantly improves segmentation performance. To achieve this, data from different modalities are normalized and spatially registered, ensuring they are processed within a unified representation space. High-level semantic features from each modality are then fused using mechanisms such as the ASPP module and attention mechanisms. This multimodal fusion strategy improves the accuracy of segmentation boundaries and enhances the model's robustness, allowing it to adapt more effectively to complex and diverse scenes.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section3\"\u003e \u003ch2\u003e2.2.3 Regularization Term to Optimize Loss Function\u003c/h2\u003e \u003cp\u003eThe application of loss functions and regularization is crucial in cultivated land information extraction tasks. This paper introduces regularization techniques to optimize the loss function by imposing constraints on model parameters, thus preventing overfitting, and improving the model\u0026rsquo;s generalization capability. L1 regularization encourages sparsity in the model by effectively selecting important features, while L2 regularization enhances stability and robustness by reducing the magnitude of parameter values. The core concept of regularization is to add a penalty term to the loss function to avoid overfitting the training data, and to ensure better performance on unseen data.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"3. Data Acquisition and Processing","content":"\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Study Area and Experimental Data\u003c/h2\u003e \u003cp\u003eThe study area for this research is Zhaohua District, Guangyuan City, located in the northern region of the Sichuan Basin, along the upper reaches of the Jialing River. Cultivated land in this region is mainly distributed across the river valley plains and hilly areas. Due to the challenging terrain, the cultivated land is fragmented, characteristic of mountainous agricultural landscapes. Zhaohua District\u0026rsquo;s cultivated land in primarily consists of paddy fields and dryland, with major crops including rice, corn, and wheat. To ensure food security, the Zhaohua District enforces stringent protection policies, focusing on maintaining high-standard cultivated land and regulating land transfer systems to prevent illegal land occupation. Zhaohua District was selected as the study area to investigate land information extraction. The experimental data include photos captured by handheld devices and drones.\u003c/p\u003e \u003cdiv id=\"Sec13\" class=\"Section3\"\u003e \u003ch2\u003e3.1.1 Photos from Handheld Devices\u003c/h2\u003e \u003cp\u003eHandheld devices, such as smartphones and tablets, are now indispensable tools in everyday life. The high-resolution cameras enable the easy capture of high-quality images, providing a valuable data source for various applications. In the context of cultivated land survey and verification, photos taken with handheld devices can quickly gather on-site information, such as land use patterns and crop growth conditions, making them a critical source of data for this study.\u003c/p\u003e \u003cp\u003eHowever, handheld device photos may suffer from issues such as noise or blur due to hand movement, lens contamination, or suboptimal shooting conditions. To ensure accurate extraction and identification of cultivated land information, it is essential to mitigate these issues while preserving important features.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section3\"\u003e \u003ch2\u003e3.1.2 Drone Photos\u003c/h2\u003e \u003cp\u003eDrone imagery provides a new perspective for identifying cultivated land resources. Equipped with high-resolution cameras, drones can capture detailed real-time images over large areas. Unlike traditional satellite imagery, drone photos offer finer spatial detail, enhancing data collection efficiency and providing a more comprehensive dataset for deep learning models, ultimately improving model accuracy and stability. Drones also offer flexibility in adjusting flight altitude and angles, enabling targeted shots and better adaptation to complex cultivated land environments. However, variations in crop growth can cause blurred boundaries in the imagery, making fine segmentation challenging for deep-learning models.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Sample Annotation\u003c/h2\u003e \u003cp\u003eSample annotation is a crucial step in deep learning-based semantic segmentation, as the quality of annotations directly influences model training and prediction outcomes. In this study, a total of 88,268 samples were annotated, including 53,767 images from handheld devices and 34,501 images from drones, details in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. These annotated datasets are essential for optimizing the model's training process and improving prediction accuracy.\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\u003eSample Annotation Data Quantity\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=\"left\" 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\u003eType\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCategory\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAnnotated Samples (Count)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"7\" rowspan=\"8\"\u003e \u003cp\u003eHandheld Device\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMixed\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3056\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBuilding\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2478\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eConstruction\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e8179\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRoads\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e7872\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTrees\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e10333\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWater Bodies\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e5671\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ecultivated Land\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e8340\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIndoor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e8340\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"5\" rowspan=\"6\"\u003e \u003cp\u003eDrone Photos\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBuilding\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e16664\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGreenhouse\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1869\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCultivated land\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e4136\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePhotovoltaic Fields\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1796\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRoads\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e6855\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWater Bodies\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3183\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\u003eIn the intelligent verification method for cultivated land surveys based on deep learning semantic segmentation, precise sample annotation is a key element in constructing and training the model. The annotated categories are primarily as follows:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eCultivated Land: This category includes crop areas such as rice fields, wheat fields, and cornfields, as well as idle cultivated land and agricultural production facilities like greenhouses (shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eOther categories: These include buildings, water bodies, trees, construction sites, and roads, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003cp\u003eTo ensure the accuracy of the sample annotations, a multi-level review mechanism was established. Each sample was annotated independently by at least two individuals, followed by cross-checks and expert reviews to ensure consistency and precision. Additionally, by observing land-use changes over multiple time periods, more accurate annotations were made for different land categories. A quality threshold was applied to the annotated samples, accompanied by quantitative evaluations and rigorous quality control procedures. During model training, feedback was incorporated to continuously optimize the annotations, ensuring that adjustments were made where necessary. Finally, a sufficient and balanced distribution of samples across all categories was maintained to mitigate the risk of overfitting or underfitting caused by data imbalance.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Data Preprocessing\u003c/h2\u003e \u003cp\u003eData preprocessing is a critical step in developing an efficient and accurate model. By utilizing open-source computer vision libraries like OpenCV, along with the Python programming language, raw data samples were meticulously processed and optimized to extract their potential and enhance model performance.\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eStrict Quality Control of Data: The initial stage of data preprocessing involved removing duplicate, erroneous, or incomplete samples. Following the preliminary annotation, a thorough quality check was conducted on segmentation masks to ensure pixel-level accuracy corresponding to the correct land type, as shown in Figs.\u0026nbsp;8 and \u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e9\u003c/span\u003e. These checks focused on boundary precision, category accuracy, and the completeness of annotated areas. Any discrepancies were promptly addressed and corrected to ensure high-quality data for training.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eLabel Processing: Annotation tools such as Labelme and COCO were employed to label the images, producing JSON or XML format files that were later converted into segmentation masks. Rigorous quality assurance ensured that each annotation had well-defined boundaries and precise category labels, avoiding mislabeling, omissions, or incomplete areas. The consistency of image and mask sizes was verified, and discrepancies were corrected through cropping, scaling, or other adjustments to maintain alignment and data integrity.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eData Augmentation: Data augmentation techniques were applied to diversify the training samples, simulating complex real-world scenarios and thereby expanding the dataset. This process is essential in cultivated land image extraction, where images are often affected by variables such as lighting, seasonal changes, and varying perspectives. The study employed geometric transformations (e.g., flipping, rotation, scaling, cropping) and color space transformations (e.g., brightness, contrast, and saturation adjustments) to generate more diverse samples. These augmentations addressed the issue of limited sample diversity in the cultivated land dataset, enriching the dataset and improving the model's generalization ability.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eDataset Splitting: A high-quality dataset is essential for constructing and fine-tuning deep learning models. The dataset was split into training, validation, and test sets; the training set for model learning, the validation set for tuning model parameters, and the test set for evaluating model performance. Given the diversity of cultivated land types and the complexity of environmental conditions, careful dataset splitting was crucial. By ensuring consistent data distribution and preventing overfitting, the model's ability to generalize to unseen data was significantly improved, contributing to the advancement of intelligent algorithms for cultivated land extraction.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e3.4 3D Spatial Land Area Calculation\u003c/h2\u003e \u003cp\u003eThe optimized SMR-DeepLabV3\u0026thinsp;+\u0026thinsp;model must not only achieve accurate image segmentation but also address the challenge of converting image pixel coordinates into actual physical areas. This process involves transforming between multiple coordinate systems, including pixel, physical, camera, and world coordinate systems. The actual area of the target region can be calculated using camera calibration techniques and the pinhole camera model. First, the segmented cultivated land regions in the pixel coordinate system retain boundary information. Then, using the camera's intrinsic and extrinsic parameters, a three-step transformation matrix is applied to derive the boundary points in the world coordinate system.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Results and Evaluation","content":"\u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Model Evaluation Metrics\u003c/h2\u003e \u003cp\u003eTo assess the performance of the enhanced SMR-DeepLabV3\u0026thinsp;+\u0026thinsp;model in segmenting cultivated land images, two widely used and complementary metrics were employed: the confusion matrix and pixel classification accuracy (PA), as well as Mean Intersection over Union (mIoU) and the Dice coefficient. These metrics provide a comprehensive assessment of the model\u0026rsquo;s segmentation accuracy and overall performance.\u003c/p\u003e \u003cdiv id=\"Sec20\" class=\"Section3\"\u003e \u003ch2\u003e4.1.1 Confusion Matrix\u003c/h2\u003e \u003cp\u003eThe confusion matrix is a valuable tool in classification and semantic segmentation tasks, especially for evaluating the model\u0026rsquo;s performance across different categories. In a binary classification task, a basic confusion matrix is represented as shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, where True Positive (TP) indicates the number of correctly predicted positive cases, False Negative (FN) indicates the number of actual positive cases incorrectly predicted as negative, False Positive (FP) is the number of actual negative cases incorrectly predicted as positive, and True Negative (TN) represents the number of correctly predicted negative cases.\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\u003eConfusion matrix\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=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConfusion Matrix\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePredicted Class 1\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePredicted Class 0\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eActual Class 1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eFN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eActual Class 0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFP\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section3\"\u003e \u003ch2\u003e4.1.2 Pixel Classification Accuracy (PA)\u003c/h2\u003e \u003cp\u003ePixel classification accuracy (PA) is a straightforward evaluation metric that measures the accuracy of the model\u0026rsquo;s pixel-level classification. It calculates the proportion of correctly classified pixels for each category. Specifically, PA is computed as the total number of correctly classified pixels divided by the total number of pixels:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:\\text{P}\\text{A}=\\left(\\frac{\\text{T}\\text{P}}{\\text{T}\\text{P}+\\text{F}\\text{N}}\\right).\\left(1\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section3\"\u003e \u003ch2\u003e4.1.3 Mean Intersection over Union (mIoU)\u003c/h2\u003e \u003cp\u003eMean Intersection over Union (mIoU) is a critical metric for image segmentation tasks, particularly for evaluating how well the model delineates segmentation boundaries for each category. It is based on the IoU, which is the ratio of the intersection area between the predicted segmentation and the actual labelled area to their union:\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:\\text{m}\\text{I}\\text{O}\\text{U}=\\left(\\frac{\\text{T}\\text{P}}{\\text{T}\\text{P}+\\text{F}\\text{P}+\\text{F}\\text{N}}\\right)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e.\u003c/p\u003e \u003cp\u003emIoU is the average of IoU values across all categories, providing a more comprehensive evaluation than PA, particularly in complex scenes.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec23\" class=\"Section3\"\u003e \u003ch2\u003e4.1.4 Dice Coefficient\u003c/h2\u003e \u003cp\u003eThe Dice coefficient is another important metric for assessing the overlap between the model's predicted segmentation and the actual ground truth labels. It is particularly useful in semantic segmentation tasks and serves as a variant of the Jaccard similarity coefficient. The Dice coefficient is more sensitive to small objects, as it measures the proportion of overlapping areas between predicted and actual regions. The Dice coefficient is computed as:\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:\\text{D}\\text{i}\\text{c}\\text{e}=\\left(\\frac{2\\text{T}\\text{P}}{2\\text{T}\\text{P}+\\text{F}\\text{P}+\\text{F}\\text{N}}\\right).\\:\\left(3\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Classification Accuracy of the SMR-DeepLabV3\u0026thinsp;+\u0026thinsp;Model\u003c/h2\u003e \u003cp\u003eThe training dataset used in this study consists of six key categories typically encountered in cultivated land surveys, including water bodies, hydraulic structures, roads, cultivated land, trees, construction sites, and buildings. This diversity ensures the model is trained on a wide range of features.\u003c/p\u003e \u003cp\u003eThe improved SMR-DeepLabV3\u0026thinsp;+\u0026thinsp;model was fine-tuned for cultivated land survey tasks, using pre-trained parameters to initialize the Xception backbone network. Critical hyperparameters, such as learning rate, batch size, and optimizer settings, were carefully monitored and adjusted to achieve optimal performance within a limited training cycle.\u003c/p\u003e \u003cp\u003eAn independent test set, comprising 10,000 handheld device images and 10,000 drone images, was used to evaluate the model\u0026rsquo;s performance. Three core evaluation metrics\u0026mdash;mIoU, PA, and Dice coefficient\u0026mdash;were used to quantify segmentation accuracy for each land type. Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e presents the detailed performance of the model across the various land categories.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eModel Performance Across Various Land Types\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"9\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLand Type\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSubcategory\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTP\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eFN\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eFP\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eTN\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003emIOU\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003ePA\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eDice\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eWater Bodies\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWater Surface\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e8985\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e98.54%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e99.42%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e99.27%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHydraulic Structures\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e27\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e8952\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e95.42%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e97.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e97.66%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eRoads\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCement Roads\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e8980\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e98.03%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e99.14%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e99%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCement Ground\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e26\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e8955\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e95.67%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e98.13%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e97.79%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDirt Roads\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e8913\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e91.95%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e93.37%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e95.81%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBuildings\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBuildings\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e8973\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e97.36%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e98.77%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e98.66%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConstruction Sites\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eConstruction Sites\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e8978\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e97.87%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e98.95%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e98.92%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eCultivated Land\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCrops\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e29\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e27\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e8944\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e94.69%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e97.16%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e97.27%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIdle Cultivated Land\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e26\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e8946\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e94.88%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e97.49%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e97.37%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGreenhouses\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e22\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e8957\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e95.89%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e97.82%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e97.90%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eTrees\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eForestland\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e49\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e39\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e8912\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e91.92%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e95.37%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e95.79%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOrchards\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e41\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e8909\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e91.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e96.09%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e95.62%\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\u003eFigure 10 illustrates segmentation results for various categories.\u003c/p\u003e\u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eModel Performance (Mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\"\u0026plusmn;\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLand Type\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSubcategory\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003emIOU\u003c/p\u003e \u003cp\u003e(Mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePA(Mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eDice(Mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003enote\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eWater Bodies\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWater Surface\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e98.54% \u0026plusmn;0.18%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e99.42% \u0026plusmn;0.12%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e99.27% \u0026plusmn;0.10%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ehigh-consistency categories effectively suppress noise in multimodal data\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eHydraulic Structures\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e95.42% \u0026plusmn;0.45%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e97.40% \u0026plusmn;0.35%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e97.66% \u0026plusmn;0.30%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eelongated structures cause localized prediction fluctuations.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eRoads\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCement Roads\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e98.03% \u0026plusmn;0.25%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e99.14% \u0026plusmn;0.15%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e99.00% \u0026plusmn;0.18%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ehomogeneous regions demonstrate high stability with manageable shadow interference.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCement Ground\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e95.67% \u0026plusmn;0.60%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e98.13% \u0026plusmn;0.40%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e97.79% \u0026plusmn;0.45%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ecracks and stains contribute to localized false detection rate variations.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDirt Roads\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e91.95% \u0026plusmn;0.85%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e93.37% \u0026plusmn;1.20%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e95.81% \u0026plusmn;0.75%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eboundary ambiguity and small target omissions result in greater variance.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBuildings\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBuildings\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e97.36% \u0026plusmn;0.30%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e98.77% \u0026plusmn;0.25%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e98.66% \u0026plusmn;0.22%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eregular shapes exhibit strong robustness, with minor fluctuations in reflective areas.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConstruction Sites\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eConstruction Sites\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e97.87% \u0026plusmn;0.35%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e98.95% \u0026plusmn;0.28%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e98.92% \u0026plusmn;0.30%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003edynamic objects (e.g., vehicles) cause localized instability.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eCultivated Land\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCrops\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e94.69% \u0026plusmn;0.70%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e97.16% \u0026plusmn;0.55%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e97.27% \u0026plusmn;0.60%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003edifferences in growth stages lead to intra-class feature variations.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIdle Cultivated Land\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e94.88% \u0026plusmn;0.65%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e97.49% \u0026plusmn;0.50%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e97.37% \u0026plusmn;0.55%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003emisclassification rates fluctuate near grassland boundaries.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGreenhouses\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e95.89% \u0026plusmn;0.50%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e97.82% \u0026plusmn;0.40%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e97.90% \u0026plusmn;0.45%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003efilm reflections are manageable, but support structure shadows affect localized stability.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eTrees\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eForestland\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e91.92% \u0026plusmn;1.10%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e95.37% \u0026plusmn;0.90%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e95.79% \u0026plusmn;0.85%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003ecanopy gaps and lighting changes cause significant fluctuations.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOrchards\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c3\"\u003e \u003cp\u003e91.60% \u0026plusmn;1.20%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c4\"\u003e \u003cp\u003e96.09% \u0026plusmn;0.95%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\"\u0026plusmn;\" colname=\"c5\"\u003e \u003cp\u003e95.62% \u0026plusmn;1.00%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003evariations in planting density and species diversity result in segmentation inconsistencies.\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\u003eTable\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e illustrate the standard deviation values to measure consistency. The node sections are error analysis. The experimental results demonstrate that the improved SMR-DeepLabV3\u0026thinsp;+\u0026thinsp;model performs exceptionally well in cultivated land survey image segmentation tasks. It achieves high overall classification accuracy and demonstrates precise segmentation across various categories. These quantitative findings offer valuable insights into the model's capabilities, laying a solid foundation for further optimization of cultivated land image segmentation techniques.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec25\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Comparative Analysis with other models\u003c/h2\u003e \u003cp\u003eTo quantify the improvements of SMR-DeepLabV3+, this study compares it with other models such as U-Net, SegNet, and PSPNet. Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e provides a detailed comparison of each model's performance in terms of mIoU, PA, and Dice.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ecomparative analysis with other models\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLand Type\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSubcategory\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003emodel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003emIOU\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePA\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eDice\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"7\" rowspan=\"8\"\u003e \u003cp\u003eWater Bodies\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eWater Surface\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eU-Net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e96.80%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e98.50%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e98.30%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSegNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e94.20%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e96.90%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e96.50%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePSPNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e97.10%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e98.80%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e98.50%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSMR-DeepLabV3+\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e98.54%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e99.42%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e99.27%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eHydraulic Structures\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eU-Net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e92.50%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e95.80%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e95.20%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSegNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e89.30%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e93.10%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e92.70%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePSPNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e93.80%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e96.20%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e95.90%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSMR-DeepLabV3+\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e95.42%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e97.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e97.66%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"11\" rowspan=\"12\"\u003e \u003cp\u003eRoads\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eCement Roads\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eU-Net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e96.20%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e98.00%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e97.80%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSegNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e92.70%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e95.40%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e95.10%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePSPNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e96.90%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e98.50%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e98.20%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSMR-DeepLabV3+\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e98.03%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e99.14%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e99%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eCement Ground\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eU-Net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e93.10%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e96.20%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e95.60%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSegNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e88.50%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e92.70%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e91.90%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePSPNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e94.30%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e96.80%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e96.20%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSMR-DeepLabV3+\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e95.67%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e98.13%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e97.79%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eDirt Roads\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eU-Net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e88.40%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e90.10%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e92.30%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSegNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e83.60%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e86.90%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e88.50%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePSPNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e89.70%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e91.50%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e93.80%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSMR-DeepLabV3+\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e91.95%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e93.37%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e95.81%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eBuildings\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eBuildings\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eU-Net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e95.30%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e97.20%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e97.00%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSegNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e91.80%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e94.60%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e94.20%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePSPNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e96.10%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e97.90%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e97.60%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSMR-DeepLabV3+\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e97.36%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e98.77%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e98.66%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eConstruction Sites\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eConstruction Sites\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eU-Net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e94.70%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e96.80%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e96.50%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSegNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e90.10%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e93.20%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e92.90%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePSPNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e95.20%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e97.10%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e96.90%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSMR-DeepLabV3+\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e97.87%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e98.95%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e98.92%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"11\" rowspan=\"12\"\u003e \u003cp\u003eCultivated Land\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eCrops\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eU-Net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e91.50%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e94.30%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e94.00%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSegNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e87.20%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e90.80%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e90.50%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePSPNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e92.80%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e95.10%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e94.90%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSMR-DeepLabV3+\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e94.69%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e97.16%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e97.27%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eIdle Cultivated Land\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eU-Net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e91.80%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e94.70%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e94.30%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSegNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e87.60%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e91.20%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e90.90%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePSPNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e93.10%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e95.40%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e95.10%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSMR-DeepLabV3+\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e94.88%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e97.49%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e97.37%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eGreenhouses\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eU-Net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e93.20%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e95.60%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e95.30%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSegNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e89.40%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e92.80%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e92.10%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePSPNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e94.50%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e96.30%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e96.00%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSMR-DeepLabV3+\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e95.89%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e97.82%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e97.90%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"7\" rowspan=\"8\"\u003e \u003cp\u003eTrees\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eForestland\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eU-Net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e88.70%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e92.50%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e92.10%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSegNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e84.10%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e88.30%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e87.90%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePSPNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e89.90%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e93.10%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e92.80%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSMR-DeepLabV3+\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e91.92%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e95.37%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e95.79%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eOrchards\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eU-Net\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e88.30%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e93.20%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e92.00%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSegNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e83.90%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e88.90%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e87.50%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePSPNet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e89.50%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e93.80%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e92.60%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSMR-DeepLabV3+\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e91.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e96.09%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e95.62%\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 performance comparison results in Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e demonstrate that SMR-DeepLabV3\u0026thinsp;+\u0026thinsp;consistently outperforms other models, such as U-Net, SegNet, and PSPNet, across all land cover categories. SMR-DeepLabV3\u0026thinsp;+\u0026thinsp;achieves good accuracy in terms of mIoU, PA, and Dice scores, particularly excelling in complex categories with intricate boundaries and challenging feature differentiation, such as hydraulic structures, dirt roads, and greenhouses.\u003c/p\u003e \u003cp\u003ePSPNet exhibits performance comparable to or even exceeding SMR-DeepLabV3\u0026thinsp;+\u0026thinsp;in some subcategories, such as water surfaces, dirt roads, and buildings, showcasing its potential in feature extraction and boundary delineation. U-Net also demonstrates stable performance in large-area land types like \"water surfaces\" and \"hydraulic structures,\" highlighting its ability to capture extensive regional features. While SegNet generally underperforms compared to the other models, it still demonstrates competitive performance in PA and Dice scores for \"buildings\" and \"dirt roads.\"\u003c/p\u003e \u003cp\u003eThis highlights SMR-DeepLabV3+'s boundary recognition and feature extraction capabilities, reinforcing its robustness in both natural and artificial environments while showcasing its performance, particularly in complex land cover types, for remote sensing image classification tasks.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec26\" class=\"Section2\"\u003e \u003ch2\u003e4.4 Cultivated Land Monitoring Analysis\u003c/h2\u003e \u003cp\u003eTo assess the effectiveness and accuracy of the SMR-DeepLabV3\u0026thinsp;+\u0026thinsp;model in cultivated land recognition and area calculation, land plots from Zhaohua District, Guangyuan City, Sichuan Province, were selected as the case study area for 2023. This region, characterized by diverse cultivated land types and typical agricultural activities, is ideal for validating the model's performance. A total of 3,466 patches were randomly sampled, ensuring a diverse dataset for comprehensive model evaluation.\u003c/p\u003e \u003cp\u003eTo scientifically assess the accuracy of intelligent verification, this study established a robust evaluation criterion after multiple rounds of data validation and comparison: if at least 65% of the area identified by the model matches the pre-labeled land class information for changes, the land classification for that patch is considered accurate. This criterion accounts for potential errors in surveys while maintaining the reliability and practicality of the recognition results. Through statistical analysis of the recognition results for the 3,466 patches, the land was categorized into three types: paddy fields, irrigated land, and dryland. The recognition statistics are presented in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab6\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eCultivated Land Recognition Statistics\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLand Class\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSubclass\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMisidentified as Other Classes\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTotal Patches\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eConcordance Rate\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003eCultivated Land\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePaddy Field\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e43\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1329\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e80.3%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIrrigated Land\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1130\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e81.7%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDryland\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e1007\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e86.1%\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 paper also presents a portion of the verification results, which are divided into two categories: compliant and non-compliant patches. In the sample images, the yellow trapezoidal area represents the photo coverage, the blue contour outlines the area of the change patch, and the green region indicates the AI-recognized area. The ratio of the area within the photo coverage that aligns with the pre-change land class in the change patch is referred to as the compliance rate. Figures\u0026nbsp;11 and \u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e12\u003c/span\u003e respectively show samples of compliant and non-compliant patch identifications.\u003c/p\u003e \u003cp\u003eThe recognition results indicate significant improvements in land class identification, confirming the effectiveness of the SMR-DeepLabV3\u0026thinsp;+\u0026thinsp;model. For the mismatches, two key factors were identified: first, human interference, such as the submission of distorted patch information; second, seasonal variations, particularly during the spring and summer months when dense vegetation growth can obscure surface features, thereby affecting the accuracy of land class determination.\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Conclusion and Discussion","content":"\u003cp\u003eGrain security is crucial to sustainable development, making the protection of cultivated land a top priority. Accurate land change surveys and verification processes are essential to maintain reliable cultivated land information. While handheld devices and drone photos provide rich data support, complex factors such as lighting conditions and angles pose challenges for traditional semantic segmentation models. To address these issues, this study improves the DeepLabV3\u0026thinsp;+\u0026thinsp;model by embedding an attention mechanism, employing multimodal fusion techniques, and incorporating regularization terms into the loss function, leading to the development of the SMR-DeepLabV3\u0026thinsp;+\u0026thinsp;model.\u003c/p\u003e \u003cp\u003eA multimodal sample dataset was constructed, consisting of 88,268 annotated samples, including 53,767 photos from handheld devices and 34,501 from drones. The model\u0026rsquo;s classification accuracy was validated using PA and mIoU, with classification accuracy consistently exceeding 90%. Through comparative experiments, SMR-DeepLabV3\u0026thinsp;+\u0026thinsp;achieves better performance in mIoU, PA, and Dice metrics compared to models such as U-Net, SegNet, and PSPNet, particularly excelling in boundary recognition and feature extraction.Additionally, a 3D spatial land area calculation method was proposed to convert pixel coordinates into physical areas, using coordinate transformation and camera calibration to calculate changes in patch areas.\u003c/p\u003e \u003cp\u003eThe SMR-DeepLabV3\u0026thinsp;+\u0026thinsp;model was tested in Zhaohua District, Guangyuan City, Sichuan Province (3,346 patches in 2023), achieving recognition rates of 80.3%, 81.7%, and 86.1% for paddy fields, irrigated land, and dryland, respectively. The model demonstrated high accuracy in semantic segmentation, showing significant advantages in improving recognition accuracy and efficiency in handheld device photos. The model also has broad application potential in cultivated land verification, promoting more refined management of cultivated land resources.\u003c/p\u003e \u003cp\u003eWhile this study has achieved promising results, several challenges remain in practical applications: 1. High Sample Production Costs: Data annotation is labor-intensive, and the model may struggle in extreme, unseen scenarios. 2. Model Interpretability: The decision-making process of the model lacks transparency, necessitating further improvements to enhance the interpretability of its predictions. To overcome these challenges, future work plans to introduce domain adaptation techniques, such as multi-domain feature alignment based on transfer learning, cross-view feature fusion with attention enhancement, dynamic domain adaptation, and self-supervised learning, to improve the model's generalization across different geographic regions and promote the automatic classification and intelligent analysis of land data. Such methods will improve the efficiency of online verification processes and provide robust, scientific decision support for land resource management, ecological protection, urban and rural planning, and other fields. Ultimately, these approaches will enhance both the efficiency and accuracy of data analysis.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eClinical trial number: not applicable.\u003c/p\u003e\n\u003cp\u003eData supporting the findings of this study are available from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003cp\u003eEthics, Consent to Participate, and Consent to Publish declarations: not applicable.\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research was funded by \u0026ldquo;Integrated dynamic monitoring of natural resources, Department of Natural Resources of Sichuan Province\u0026rdquo;, grant number N5100012024001267.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll authors thanks the data sharing of Zhaohua District, Guangyuan City.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor\u0026apos;s email address\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eZheng Fang, [email protected];\u003c/p\u003e\n\u003cp\u003eYuhang Gong, [email protected];\u003c/p\u003e\n\u003cp\u003eHeli Zhu, [email protected];\u003c/p\u003e\n\u003cp\u003eCanyang Shi, [email protected];\u003c/p\u003e\n\u003cp\u003eZhijia Gong, [email protected];\u003c/p\u003e\n\u003cp\u003eLin Tian, [email protected].\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eZ.F.: Conceptualization, methodology, formal analysis, writing\u0026mdash;original draft preparation, supervision,project administration, funding acquisition;Y.G.: methodology, software, formal analysis, writing\u0026mdash;review and editing, visualization;H.Z.:validation, investigation,resources, data curation, writing\u0026mdash;review and editing;C.S.:validation, investigation,resources, data curation, writing\u0026mdash;review and editing;Z.G.:validation, investigation,resources, data curation, writing\u0026mdash;review and editing;L.T.:validation, investigation,resources, data curation, writing\u0026mdash;review and editing;All authors have read and agreed to the published version of the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col class=\"decimal_type\"\u003e\n\u003cli\u003eDong Z Y， Li J H， Zhang J， Yu J Q and An S. 2023. Cultivated land extraction from high-resolution remote sensing images based on BECU-Net model with edge enhancement. National Remote Sensing Bulletin， 27（12）：2847-2859 DOI： 10.11834/jrs.20222268.\u003c/li\u003e\n\u003cli\u003eDU Guo-ming,GAI Zhao-xue*,WANG Hong-yan.Theoretical Explanation and Research Framework of Cultivated-land Fragmentation in China[J].Journal of Earth Sciences and Environment,2021,43(06):997-1008.[doi:10.19814/j.jese.2021.07029]\u003c/li\u003e\n\u003cli\u003eZHANG Xinchang, HUANG Jianfeng, NING Ting. Progress and Prospect of Cultivated Land Extraction from High-Resolution Remote Sensing Images[J]. Geomatics and Information Science of Wuhan University, 2023, 48(10): 1582-1590. DOI: 10.13203/j.whugis20230114\u003c/li\u003e\n\u003cli\u003eLI Liangjun; MU Zhe; Lv Lei; MIAODAN. Application analysis of \u0026ldquo;Internet+\u0026rdquo; evidence technology in land survey[J]. Surveying, Mapping and Spatial Geographic Information,2024,(01):200-202+205.\u003c/li\u003e\n\u003cli\u003eFeng Siyuan. Application of UAV inclined photogrammetry in the third national land survey [J]. Mapping and Spatial Geographic Information, 2023, 46 (07): 154-156+160.\u003c/li\u003e\n\u003cli\u003eXiao H; Dong Shengguang; Tang Zhiqing; Tang Xianlong. Exploration and practice of construction of command center for natural resources investigation and monitoring in Hunan Province[J]. Natural Resources Informatization,2023,(03):9-14.\u003c/li\u003e\n\u003cli\u003eZHANG Yi;JI Bo;JING Ruan. Practice and thinking of natural resources intelligent sensing model based on tower-based video monitoring[J]. Land Resources Informatization,2021,(05):8-14.\u003c/li\u003e\n\u003cli\u003eZHANG Yichao, ZHENG Xiangtao, LU Xiaoqiang. Hyperspectral image classification method based on hierarchical transformer network[J]. Acta Geodaetica et Cartographica Sinica, 2023, 52(7): 1139-1147.\u003c/li\u003e\n\u003cli\u003eDU Peijun, ZHANG Wei, ZHANG Peng, LIN Cong, GUO Shanchuan, HU Zezhou. A capsule network for hyperspectral image classification employing spatial-spectral feature[J]. Acta Geodaetica et Cartographica Sinica, 2023, 52(7): 1090-1104.\u003c/li\u003e\n\u003cli\u003eSUN Genyun, WANG Xin, AN Na, ZHANG Aizhu. A method for large-scale and high-resolution impervious surface extraction based on multi-source remote sensing and deep learning[J]. Acta Geodaetica et Cartographica Sinica, 2023, 52(2): 272-282.\u003c/li\u003e\n\u003cli\u003ePeng L K，Liu L C，Chen X H，Chen J，Cao X and Qiu Y A. 2021. Generalization ability of cloud detection network for satellite imagery based on DeepLabv3+. National Remote Sensing Bulletin， 25（5）：1169-1186 DOI： 10.11834/jrs.20210061.\u003c/li\u003e\n\u003cli\u003eZHENG Kai, LI Jiansheng, YANG Jianfeng, OUYANG Wen, WANG Gaojie, ZHANG Xun. A cloud and snow detection method of TH-1 image based on combined ResNet and DeepLabV3+[J]. Acta Geodaetica et Cartographica Sinica, 2020, 49(10): 1343-1353.\u003c/li\u003e\n\u003cli\u003eCHEN Menghua,ZHANG Tongyun,ZHOU Zixiang,et al. A high-precision batch automatic extraction method for buildings based on DeepLabV3+[J/OL]. Journal of Surveying and Mapping Science and Technology,1-6[2024-09-11]\u003c/li\u003e\n\u003cli\u003eJIANG Baode, HANG Wei, XU Shaofen, WU Yong. Multi-scale building instance refinement extraction from remote sensing images by fusing with decentralized adaptive attention mechanism[J]. Acta Geodaetica et Cartographica Sinica, 2023, 52(9): 1504-1514.\u003c/li\u003e\n\u003cli\u003eLIN Yunhao, WANG Yanjun, LI Shaochun, CAI Hengfan. A coupled DeepLab and Transformer approach for fine classification of crop cultivation types in remote sensing[J]. Acta Geodaetica et Cartographica Sinica, 2024, 53(2): 353-366.\u003c/li\u003e\n\u003cli\u003eZHOU Xinxin, WU Yanlan, LI Mengya, ZHENG Zhiteng. Automatic Vegetation Extraction Method based on Feature Separation Mechanism with Deep Learning[J]. Journal of Geo-information Science, 2021, 23(9): 1675-1689 https://doi.org/10.12082/dqxxkx.2021.200641\u003c/li\u003e\n\u003cli\u003eLiu C T， Feng Q L， Liu J T， Wang Y， Shi T G， Li Y， Gong J H and Zhao H H. 2022. Urban green plastic cover extraction and spatial pattern changes in Jinan city based on DeepLabv3+ semantic segmentation model. National Remote Sensing Bulletin， 26（12）：2518-2530 DOI： 10.11834/jrs.20220101.\u003c/li\u003e\n\u003cli\u003eWu Q Q，Wang S，Wang B and Wu Y L. 2022. Road extraction method of high-resolution remote sensing image on the basis of the spatial information perception semantic segmentation model. National Remote Sensing Bulletin， 26（9）： 1872-1885 DOI： 10.11834/jrs.20210021.\u003c/li\u003e\n\u003cli\u003eTIAN Fuyou，CAO Yupei，ZHAO Hang，WU Bingfang，ZENG Hongwei，LIU Yazhou，QIN Xingli，ZHANG Miao，ZHU Liang，ZHU Weiwei. XXXX. Agricultural Field Segmentation using Spatial Attention Mechanism and Multi-task Learning Strategy. National Remote Sensing Bulletin， XX（XX）：1-15 DOI： 10.11834/jrs.20243191.\u003c/li\u003e\n\u003cli\u003eTAN Min, LIN Huijing, HAO Ming. Prior Knowledge Guided Deep Learning for Monitoring Buildings and Greenhouses within Cultivated Land[J]. Journal of Geo-information Science, 2023, 25(11): 2293-2302 https://doi.org/10.12082/dqxxkx.2023.230315\u003c/li\u003e\n\u003cli\u003eZhu, X.X., et al. (2017) Deep Learning in Remote Sensing: A Comprehensive Review and List of Resources. IEEE Geoscience and Remote Sensing Magazine, 5, 8-36.\u003c/li\u003e\n\u003cli\u003eLi Y, Zhang H, Xue X, et al. Deep learning for remote sensing image classification: A survey[J]. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 2018, 8(6): e1264.\u003c/li\u003e\n\u003cli\u003eRonneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation[C]//Medical image computing and computer-assisted intervention\u0026ndash;MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. Springer International Publishing, 2015: 234-241.\u003c/li\u003e\n\u003cli\u003eKampffmeyer M, Salberg A B, Jenssen R. Semantic segmentation of small objects and modeling of uncertainty in urban remote sensing images using deep convolutional neural networks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition workshops. 2016: 1-9.\u003c/li\u003e\n\u003cli\u003eChen L C, Zhu Y, Papandreou G, et al. Encoder-decoder with atrous separable convolution for semantic image segmentation[C]//Proceedings of the European conference on computer vision (ECCV). 2018: 801-818.\u003c/li\u003e\n\u003cli\u003ePan Y, Zhang L. Dual attention deep learning network for automatic steel surface defect segmentation[J]. Computer‐Aided Civil and Infrastructure Engineering, 2022, 37(11): 1468-1487.\u003c/li\u003e\n\u003cli\u003eGu H, Wang P, Li Y, et al. MFP-DeepLabv3+: A Multi-scale Feature Fusion and Parallel Attention Network for Enhanced Bone Metastasis Segmentation[J]. Traitement du Signal, 2024, 41(2).\u003c/li\u003e\n\u003cli\u003eWang Y, Yang L, Liu X, et al. An improved semantic segmentation algorithm for high-resolution remote sensing images based on DeepLabv3+[J]. Scientific Reports, 2024, 14(1): 9716.\u003c/li\u003e\n\u003cli\u003eUllah M, Bais A, Wist T. TranSEF: Transformer Enhanced Self-Ensemble Framework for Damage Assessment in Canola Crops[J]. IEEE Transactions on AgriFood Electronics, 2024.\u003c/li\u003e\n\u003cli\u003eChen, Liang-Chieh, et al. Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), 2018, pp. 801-818.\u003c/li\u003e\n\u003cli\u003eFran\u0026ccedil;ois Chollet. Xception: Deep Learning with Depthwise Separable Convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 1251-1258.\u003c/li\u003e\n\u003cli\u003eKeskar, Nitish Shirish, and Richard Socher. Improving Generalization Performance by Switching from Adam to SGD. arXiv preprint arXiv:1712.07628, 2017.\u003c/li\u003e\n\u003cli\u003eLoshchilov, Ilya, and Frank Hutter. Decoupled Weight Decay Regularization. In International Conference on Learning Representations (ICLR), 2019.\u003c/li\u003e\n\u003cli\u003eGoodfellow I. Deep learning[J]. 2016.\u003c/li\u003e\n\u003cli\u003eHu, Jie, et al. \u0026quot;Squeeze-and-Excitation Networks.\u0026quot; IEEE Transactions on Pattern Analysis and Machine Intelligence (2018).\u003c/li\u003e\n\u003cli\u003eBaltru\u0026scaron;aitis, T., et al. \u0026quot;Multimodal Machine Learning: A Survey and Taxonomy.\u0026quot; IEEE Transactions on Pattern Analysis and Machine Intelligence (2019).\u003c/li\u003e\n\u003cli\u003eLi, Z., et al. \u0026quot;Deep Learning for Remote Sensing Image Classification: A Comprehensive Review.\u0026quot; IEEE Transactions on Geoscience and Remote Sensing (2020).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"discover-sustainability","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"disu","sideBox":"Learn more about [Discover Sustainability](https://www.springer.com/43621)","snPcode":"","submissionUrl":"","title":"Discover Sustainability","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Discover Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Cultivated Land, Deep Learning, Semantic Segmentation, DeepLabV3+, Attention Mechanism","lastPublishedDoi":"10.21203/rs.3.rs-5770339/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5770339/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe protection and monitoring of cultivated land, as a vital resource, are of utmost importance. Rapid and accurate extraction of cultivated land significantly enhances the quality and efficiency of comprehensive natural resource monitoring. In recent years, advanced technologies for cultivated land extraction, particularly deep learning-based semantic segmentation, have yielded impressive results. The DeepLabV3\u0026thinsp;+\u0026thinsp;model has demonstrated exceptional performance across multiple datasets. However, conventional approaches primarily rely on high-resolution remote sensing imagery, which may not always be feasible in practical scenarios requiring cross-perspective image analysis (e.g., images captured by handheld devices or drones), such as in cultivated land survey verification. This paper introduces a new algorithm, SMR-DeepLabV3+, which integrates an attention mechanism into the DeepLabV3\u0026thinsp;+\u0026thinsp;model, along with regularization enhancements. The model was trained using 88,268 meticulously annotated samples and evaluated on 3,466 cultivated land patches from Zhaohua District, Guangyuan City, Sichuan Province. The results showed recognition rates of 80.3% for paddy fields, 81.7% for irrigated land, and 86.1% for dryland. The SMR-DeepLabV3\u0026thinsp;+\u0026thinsp;model significantly improves the precision of cultivated land extraction and area estimation, offering a practical and effective solution for identifying different types of cultivated land. This study contributes to more refined management of cultivated land resources and provides reliable support for assessing the effectiveness of cultivated land protection policies.\u003c/p\u003e","manuscriptTitle":"SMR-DeepLabV3+: A Refined Model for Cultivated Land Extraction from Cross-View Imagery Using an Attention Mechanism","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-04-10 10:15:12","doi":"10.21203/rs.3.rs-5770339/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-05-07T01:55:43+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-04-15T07:57:29+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"134320053135652124857666292002778334479","date":"2025-04-14T09:12:56+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"154887506392195024242844688860059290424","date":"2025-04-11T14:24:22+00:00","index":"hide","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-04-09T09:02:19+00:00","index":"","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-04-09T09:01:35+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-04-08T12:29:02+00:00","index":"","fulltext":""},{"type":"submitted","content":"Discover Sustainability","date":"2025-03-24T15:56:08+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"discover-sustainability","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"disu","sideBox":"Learn more about [Discover Sustainability](https://www.springer.com/43621)","snPcode":"","submissionUrl":"","title":"Discover Sustainability","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Discover Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"72e4412a-0abb-4fdc-a9d2-10fdd7c370ab","owner":[],"postedDate":"April 10th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2025-07-07T16:06:56+00:00","versionOfRecord":{"articleIdentity":"rs-5770339","link":"https://doi.org/10.1007/s43621-025-01461-w","journal":{"identity":"discover-sustainability","isVorOnly":false,"title":"Discover Sustainability"},"publishedOn":"2025-07-01 15:58:27","publishedOnDateReadable":"July 1st, 2025"},"versionCreatedAt":"2025-04-10 10:15:12","video":"","vorDoi":"10.1007/s43621-025-01461-w","vorDoiUrl":"https://doi.org/10.1007/s43621-025-01461-w","workflowStages":[]},"version":"v1","identity":"rs-5770339","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5770339","identity":"rs-5770339","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}