A remote sensing image interpretation method based on a lightweight visual base model

Abstract

The application provides a remote sensing image interpretation method based on a lightweight visual basic model, a double-time-phase remote sensing image change detection model Edge-CD constructed by the method comprises an EdgeSAM encoder and a boundary reinforcement decoding head; the EdgeSAM encoder adopts an EdgeSAM image encoder as a feature extraction backbone, and a space-time feature alignment module is embedded in the EdgeSAM image encoder, so that the perception ability of the model to real changes is effectively enhanced, and the common false change problem is relieved. A boundary enhancement module is designed in the boundary reinforcement decoding head, so that the fitting degree of the boundary contour of the detection result to the real change is improved. In addition, in view of the problem that the traditional loss function is insufficient for optimizing the edge region, a boundary perception loss function is designed, a higher weight is given to the neighborhood region of the real change boundary, the gradient response of the model in the boundary region is enhanced, the edge false detection and the edge missing detection are effectively relieved, and the segmentation precision and the geometric integrity of the change boundary are further improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent interpretation technology of remote sensing images, and in particular to a remote sensing image interpretation method based on a lightweight visual foundation model. Background Technology

[0002] Remote sensing image interpretation (i.e., remote sensing change detection) is a core technology for analyzing the spatiotemporal dynamic evolution of the Earth's surface. By comparing and analyzing multi-temporal remote sensing images, it can accurately identify the transformation of land cover types and spatial morphological changes, playing an irreplaceable role in key areas such as refined land resource management, dynamic monitoring of the ecological environment, urban expansion planning, emergency response to natural disasters, and regulation of agricultural production.

[0003] In recent years, deep learning technology has made significant progress in the field of remote sensing change detection, with methods based on convolutional neural networks (CNNs) and transformers being widely adopted. CNN-based methods typically use networks with shared weights to aggregate semantic change information from multi-temporal images; while transformer-based methods leverage their global perception capabilities to capture long-range dependencies in images through encoders. Furthermore, hybrid methods combining CNNs and transformers can more effectively model contextual information in the spatiotemporal domain. Although these methods can achieve relatively effective remote sensing change detection results, they still face two major bottlenecks in practical applications: 1. A strong dependence on a large amount of high-quality labeled data, and the resulting insufficient model generalization ability. Specifically, the annotation of remote sensing images has a very high professional threshold. High-quality labeled data is often extremely scarce, or even completely unavailable, leading to insufficient training samples or difficulty in guaranteeing annotation accuracy. Limited by the availability and representativeness of labeled data, current mainstream deep learning models experience a significant drop in interpretation performance and severely limited generalization ability when facing new scenes with distributions significantly different from the training set. 2. Real-world remote sensing scenes are inherently highly complex and dynamic. For example, urban areas typically contain dense building clusters, intricate road networks, fragmented green spaces and water bodies, with blurred boundaries and mixed land cover types. Agricultural areas exhibit crop rotation, irregular field boundaries, and a mix of irrigation facilities and temporary coverings, making it difficult to distinguish seasonal changes from actual land use changes. Natural ecological areas may simultaneously encompass multiple mixed land cover types such as forests, wetlands, bare rock, and landslides, exhibiting strong spatial heterogeneity and high semantic ambiguity. These complex spatial patterns mean that identifying changing areas not only faces the binary challenge of "change versus constancy" but also requires addressing deeper issues such as whether the changes have semantic significance. However, existing dual-temporal interpretation methods mostly employ static feature comparison or simple difference strategies, lacking the ability to model the contextual relationships of land cover, the evolutionary process of change, and spatiotemporal consistency. Furthermore, non-changing interferences introduced by imaging conditions such as cloud cover, changes in illumination angle, and sensor differences further exacerbate the interpretation difficulty.

[0004] To alleviate the heavy reliance of traditional deep learning methods on large-scale, high-quality labeled data, visual foundation models have been introduced into the field of intelligent interpretation of remote sensing images in recent years due to their strong generalization ability in general vision tasks. These models learn rich general visual representations through self-supervised or weakly supervised pre-training on massive and diverse natural images, providing a new technical path for tasks such as change detection in dual-temporal remote sensing images. However, despite their excellent performance in general scenarios, directly transferring visual foundation models to dual-temporal remote sensing image intelligent interpretation tasks still faces significant limitations, mainly in terms of computational efficiency and task adaptability. Regarding computational efficiency, visual foundation models, represented by SAM, employ heavy-duty image encoders based on ViT, with hundreds of millions of parameters. The inference process places extremely high demands on computational resources, memory bandwidth, and power consumption, severely restricting their feasibility for implementation in time-sensitive applications such as emergency response and real-time monitoring. Regarding adaptability to remote sensing tasks, the visual basic model has not been structurally or mechanistically adapted to the core requirements of dual-temporal remote sensing tasks, making it difficult to effectively suppress non-variable interferences caused by differences in illumination, atmosphere, and sensors. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] Based on this, the present invention provides a remote sensing image interpretation method based on a lightweight visual basic model to solve the problems mentioned in the background art, such as the difficulty of interpretation caused by non-changing interference introduced by imaging conditions, the insufficient generalization ability of the model caused by the strong dependence on a large amount of high-quality labeled data, the lack of modeling ability for contextual relationships, change and evolution processes and spatiotemporal consistency leading to poor interpretation results and low interpretation computation efficiency.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, this invention provides a remote sensing image interpretation method based on a lightweight visual foundation model, comprising:

[0009] Step S1: Construct the Edge-CD remote sensing image change detection model based on the lightweight visual foundation model EdgeSAM;

[0010] Step S2: Based on the remote sensing image dataset, train the Edge-CD model to obtain the trained Edge-CD model;

[0011] Step S3: Input remote sensing images from different time phases into the trained Edge-CD model to obtain the interpretation results;

[0012] In step S1, the Edge-CD model includes: an EdgeSAM encoder unit and a boundary enhancement decoding head unit; the EdgeSAM encoder unit includes: an EdgeSAM image encoder and a spatiotemporal feature alignment module; the boundary enhancement decoding head unit includes: a boundary enhancement module (EEM), a convolution module, a size unification module, and a prediction head; the EEM includes a SimAM attention module.

[0013] dual-temporal remote sensing images and The EdgeSAM encoder unit extracts spatiotemporal alignment features respectively. and ;Will and Features obtained by splicing Input the boundary enhancement decoding head unit to obtain the change detection results;

[0014] In the EdgeSAM encoder unit, the single-input EdgeSAM image encoder is improved into a twin structure with two dual-temporal remote sensing image inputs. That is, the dual-temporal remote sensing images... and The first EdgeSAM image encoder and the second EdgeSAM image encoder are input respectively, and the two EdgeSAM image encoders share weights.

[0015] In the EdgeSAM image encoder, multi-level feature maps are generated through 12 encoder modules. Each encoder module is considered as one layer; that is, the first encoder module of the first EdgeSAM image encoder and the first encoder module of the second EdgeSAM image encoder both belong to the first layer, the second encoder module of the first EdgeSAM image encoder and the second encoder module both belong to the second layer, and so on, for a total of 12 layers. This invention embeds a spatiotemporal feature alignment module after each encoder module in layers 1 to 11, while the last layer does not contain a spatiotemporal feature alignment module. The output of the last encoder module is used as the output of the EdgeSAM encoder unit. In other words, dual-temporal remote sensing images... and After passing through the first encoder module of the first EdgeSAM image encoder and the second EdgeSAM image encoder respectively, the two feature maps obtained by the first encoder module are input together into the first spatiotemporal feature alignment module. The two output features of the first spatiotemporal feature alignment module are then used as inputs into the second encoder module. This process continues until the output of the last encoder module is used as the output of the EdgeSAM encoder unit.

[0016] Specifically, the encoder module is a RepViT Block.

[0017] Specifically, when describing the structure of the spatiotemporal feature alignment module, it is assumed that the input of the module is... and The output is and ;

[0018] First, input and Feature extraction and enhancement are performed separately using a CBG module to obtain... and The CBG module includes, in sequence, convolution, normalization, and activation functions; subsequently, and Compared with the original input and Element-wise multiplication is performed to enhance the spatiotemporal feature information and obtain the result. and This process can be represented by the following formula:

[0019]

[0020] in, This indicates that the CBG module is processing the data. This indicates element-wise multiplication;

[0021] Then, and After each of the two processes undergoes a 1×1 convolution, we obtain... and After that, Spatiotemporal characteristics in and Perform element-wise addition, while... Spatiotemporal characteristics in and By adding element by element, we get and The following indicates:

[0022]

[0023] in, This represents a 1×1 convolution; This indicates element-wise addition.

[0024] Specifically, when describing the structure of the boundary enhancement decoding head, firstly, the input features of the boundary enhancement decoding head unit are... The features are respectively processed through max pooling downsampling, an identity mapping layer, and two transposed convolutions of different sizes; among them, the features The feature dimension is C×H×W; the feature dimension output after max pooling downsampling is C×H / 2×W / 2, the feature dimension output after the identity mapping layer is C×H×W, the feature dimension output after the first transposed convolution is C / 2×2*H×2*W, and the feature dimension output after the second transposed convolution is C / 4×4*H×4*W, thus obtaining a hierarchical multi-scale feature. , , and Secondly, the features The feature representation is refined through convolutional modules to obtain features. ,in, Then, let , ;Will and The input EEM enhances the focus on key edge details of multi-scale features, and the output boundary enhancement features. ;Will and Input to EEM, output boundary enhancement features After that, After passing through the size unification module, the data is concatenated to obtain the final features. The final features are then input into the prediction head to output a change mask. The convolution module includes convolution, normalization, and ReLU activation functions. The size unification module uses 1×1 convolution and bilinear interpolation to unify the size to the same number of channels and spatial dimensions, providing a consistent scale for feature fusion and decoding. The prediction head maps the channel dimensions of the final features to the number of target categories through 1×1 convolutional layers, outputting the change detection results.

[0025] Specifically, when describing the boundary enhancement module, it is assumed that the input features of this module are... and The output is ;

[0026] First, and After element-wise subtraction, differential feature maps are generated through convolutional layers and the ReLU nonlinear activation function. It can be expressed by the following formula:

[0027]

[0028] At the same time, and The results are obtained by inputting the SimAM attention module respectively. and It can be represented by the following formula:

[0029]

[0030] The SimAM attention module uses the SimAM attention mechanism to enhance or suppress the input features of the module element by element using 3D attention weights to obtain the output features of the module.

[0031] Each and Edge features are extracted through a combination of average pooling, subtraction, convolution, and the sigmoid function. and The subtraction operation involves performing a difference calculation on the features before and after the average pooling operation to enhance edge regions. For example, it involves subtracting the features before the average pooling operation... Features after average pooling operation Difference calculation is performed to obtain The following formula represents:

[0032]

[0033] in, Indicates average pooling; Represents convolution; This represents the Sigmoid activation function;

[0034] Will , respectively with , After further enhancing the edge features through multiplication and addition operations, the data is input into the SimAM attention module again to obtain... , ;Will , Subtraction yields the difference edge features As shown below:

[0035]

[0036] Difference edge features and differences in characteristics After addition, the boundaries are further enhanced using the SimAM attention module to obtain enhanced boundary features. As shown below:

[0037] .

[0038] Specifically, in step S2, during the training phase, the image domain of the change mask is set as... The manual annotations of the actual changes are First, regarding the real labels Gaussian smoothing is applied; then, the prediction boundary is extracted using morphological gradients. Then, an expansion operation is performed on the predicted boundary to form a "boundary band" of a certain width, denoted as... And define the boundary-aware loss function. for:

[0039]

[0040] in, Indicates the boundary range of the actual area of ​​change; Represents points on the prediction boundary; Represents a point on the actual boundary; Points on the prediction boundary Inside the real boundary, Points on the prediction boundary Outside the true boundary; inside the true boundary represents a region of change, and outside the true boundary represents a region of no change;

[0041] Boundary-aware loss function Joint optimization is performed with the traditional region loss function; where the traditional region loss function uses binary cross-entropy loss, i.e. The joint loss function is then:

[0042]

[0043] in, This represents the balance coefficient.

[0044] Specifically, the remote sensing image dataset is MSRS-CD.

[0045] (III) Beneficial Effects

[0046] As can be seen from the above technical solution, the remote sensing image interpretation method based on a lightweight visual fundamental model proposed in this invention has the following beneficial effects:

[0047] 1. The change detection model Edge-CD constructed in this invention is based on the lightweight visual basic model EdgeSAM. While significantly reducing the computational overhead, it retains excellent general visual representation capabilities and also has strong few-sample generalization capabilities.

[0048] 2. The spatiotemporal feature alignment module designed in this invention effectively suppresses noise interference caused by irrelevant factors such as illumination angle, shadow distribution and sensor type differences through dynamic interaction and compensation of dual-temporal features, thereby improving the model's extraction quality of dual-temporal features.

[0049] 3. The boundary enhancement module (EEM) designed in this invention, by combining the SimAM attention mechanism, assigns higher attention weight to the edges of changing regions, enhancing the obtained boundary information and suppressing redundant information interference from invariant regions. In the boundary enhancement detection head unit, by constructing multi-scale features and combining them with the boundary enhancement module, the model is guided to focus on fine-grained details at changing edges, making the boundaries of changing regions clearer and improving the segmentation effect of changing boundaries.

[0050] 4. To address the issue of insufficient optimization of edge regions by traditional region loss functions, a boundary-aware loss function was designed. This function, based on conventional pixel-level supervision, assigns higher weights to the neighborhood regions of truly changing boundaries, enhancing the model's gradient response in boundary regions and effectively mitigating false positives and false negatives at edges, further improving the segmentation accuracy and geometric integrity of changing boundaries. A joint loss optimization mechanism combining boundary-aware loss and traditional region loss strengthens boundary guidance at both the feature fusion and output prediction levels, significantly improving the localization accuracy and boundary integrity of changing regions. Attached Figure Description

[0051] The features and advantages of the invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the invention in any way. In the drawings:

[0052] Figure 1 This is a schematic diagram of the Edge-CD model for detecting changes in remote sensing images based on a lightweight visual fundamental model, as described in this invention.

[0053] Figure 2 This is a schematic diagram of the spatiotemporal feature alignment module of the present invention;

[0054] Figure 3 This is a schematic diagram of the boundary enhancement decoding head unit of the present invention;

[0055] Figure 4 This is a schematic diagram of the boundary reinforcement module of the present invention;

[0056] Figure 5 This is a schematic diagram illustrating the operational principle of the boundary-aware loss function of this invention.

[0057] Figure 6 This is a schematic diagram illustrating the detection effect of the Edge-CD model of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] This invention provides a remote sensing image interpretation method based on a lightweight visual foundation model, comprising:

[0060] Step S1: Construct the Edge-CD remote sensing image change detection model based on the lightweight visual foundation model (EdgeSAM);

[0061] like Figure 1 As shown, the Edge-CD model includes: an EdgeSAM encoder unit and a boundary enhancement decoding head unit; wherein, the EdgeSAM encoder unit includes: an EdgeSAM image encoder and a spatiotemporal feature alignment module; the boundary enhancement decoding head unit includes: a boundary enhancement module (EEM), a convolution module, a size unification module and a prediction head; the EEM includes a SimAM attention module.

[0062] dual-temporal remote sensing images and The EdgeSAM encoder unit extracts spatiotemporal alignment features respectively. and ;Will and Features obtained by concatenation Input the boundary enhancement decoding head unit to obtain the change detection result.

[0063] 1. EdgeSAM encoder unit

[0064] EdgeSAM (Edge Segment Anything Model), a lightweight visual foundational model designed for general image segmentation tasks, originally only supported a single image input, making it difficult to directly adapt to the dual-temporal feature modeling requirements of dual-temporal remote sensing image interpretation tasks. Therefore, in the EdgeSAM encoder unit, the single-input EdgeSAM image encoder was improved into a twin structure with two dual-temporal remote sensing image inputs. In other words, the dual-temporal remote sensing images... and The first EdgeSAM image encoder and the second EdgeSAM image encoder are input respectively, and the two EdgeSAM image encoders share weights. EdgeSAM is built on an efficient convolutional neural network, which significantly reduces computational overhead while retaining excellent general visual representation capabilities.

[0065] In the EdgeSAM image encoder, multi-level feature maps are generated through 12 encoder modules (i.e., RepViT Blocks). Each encoder module is considered as one layer; that is, the first encoder module of the first EdgeSAM image encoder and the first encoder module of the second EdgeSAM image encoder both belong to the first layer, the second encoder module of the first EdgeSAM image encoder and the second encoder module both belong to the second layer, and so on, for a total of 12 layers. This invention embeds a spatiotemporal feature alignment module after each encoder module in layers 1 to 11. The last layer (layer 12) does not contain a spatiotemporal feature alignment module; the output of the last encoder module (i.e., the 12th encoder module) is used as the output of the EdgeSAM encoder unit. In other words, dual-temporal remote sensing images... and After passing through the first encoder module of the first EdgeSAM image encoder and the second EdgeSAM image encoder respectively, the two feature maps obtained by the first encoder module (layer 1) are input together into the first spatiotemporal feature alignment module. The two output features of the first spatiotemporal feature alignment module are then used as inputs into the second encoder module (layer 2). This process continues until the output of the last encoder module is used as the output of the EdgeSAM encoder unit.

[0066] Since EdgeSAM is not specifically designed for change detection tasks, it cannot accurately capture the homogeneous and heterogeneous features of two images over time when processing bi-temporal images. This poses a challenge to combining lightweight visual base models with change detection tasks. To address this issue and further improve the model's adaptability to multi-scale ground features, complex backgrounds, and weak change signals in remote sensing images, this invention designs a spatiotemporal feature alignment module, such as... Figure 2 As shown. Assume the input to the spatiotemporal feature alignment module is... and The output is and In this module, Spatiotemporal features injected into In the middle, at the same time Spatiotemporal features injected into In this process, the feature cross-injection mechanism not only enhances the model's ability to understand changes between different time points, but also effectively unifies the spatiotemporal features of the two time phases.

[0067] Specifically, first, input and Feature extraction and enhancement are performed separately using a CBG module to obtain... and The CBG module includes, in sequence, convolution, normalization, and activation functions; subsequently, and Compared with the original input and Element-wise multiplication is performed to enhance the spatiotemporal feature information and obtain the result. and This process can be represented by the following formula:

[0068]

[0069] in, This indicates that the CBG module is processing the data. This indicates element-wise multiplication.

[0070] Then, and After each of the two processes undergoes a 1×1 convolution, we obtain... and This operation aims to capture and highlight spatiotemporal features that are crucial for change detection; subsequently, Spatiotemporal characteristics in and Perform element-wise addition, while... Spatiotemporal characteristics in and By adding element by element, we get and This achieves spatiotemporal feature alignment between two temporal images. As shown below:

[0071]

[0072] in, This represents a 1×1 convolution; This indicates element-wise addition.

[0073] The spatiotemporal feature alignment module uses a channel-space collaborative attention mechanism to dynamically model the semantic correspondence between bi-temporal images on feature maps at different levels generated by multiple encoder modules. This enables adaptive compensation of cross-temporal features, effectively suppresses pseudo-change interference caused by changes in imaging conditions such as illumination differences, cloud and fog obstruction, and sensor offset, and significantly enhances the model's sensitivity to real change signals.

[0074] Through the above operations, spatiotemporal information between the two-phase remote sensing images can be effectively exchanged. This ensures that the lightweight visual base model can accurately align spatiotemporal features during the feature extraction stage, thereby improving the effectiveness and reliability of feature extraction.

[0075] 2. Boundary Enhancement Decoding Head Unit

[0076] EdgeSAM extracts rich global semantic features, but conventional decoding heads struggle to fully process the boundary information, resulting in the inability to generate fine-grained and accurate change masks. To address this, this invention incorporates a boundary enhancement decoding head with an edge enhancement module (EEM), enabling the generated change mask to meet the requirements of fine-grained remote sensing monitoring.

[0077] After completing the dual-temporal feature extraction and adaptive interaction, the fused features (concatenated features) are fed into a decoder head equipped with a boundary enhancement mechanism. The structure of the boundary enhancement decoder head is as follows: Figure 3 As shown. First, the input features of the boundary enhancement decoding head unit are... (Feature dimensions are C×H×W) The features are processed through max pooling downsampling, an identity mapping layer, and two transposed convolutions of different sizes. The feature dimension output after max pooling downsampling is C×H / 2×W / 2, the feature dimension output after the identity mapping layer is C×H×W, the feature dimension output after the first transposed convolution is C / 2×2*H×2*W, and the feature dimension output after the second transposed convolution is C / 4×4*H×4*W. This yields multi-scale features with a hierarchical structure. , , and Secondly, the features The feature representation is refined through convolutional modules to obtain features. Then, let , ;Will and The input EEM enhances the focus on key edge details of multi-scale features, and the output boundary enhancement features. ;Will and Input to EEM, output boundary enhancement features After that, After passing through the size unification module, the data is concatenated to obtain the final features. These final features are then input into the prediction head, which outputs a change mask. The convolutional module includes convolution, normalization, and ReLU activation functions. The size unification module uses 1×1 convolution and bilinear interpolation to uniformly adjust the size to the same number of channels and spatial dimensions, providing a consistent scale for feature fusion and decoding. The prediction head maps the channel dimensions of the final features to the number of target classes using 1×1 convolutional layers, outputting the change detection results.

[0078] This boundary enhancement decoding head employs max pooling, identity mapping layers, and two transposed convolutions of different sizes to extract high-level semantic information and low-level spatial details layer by layer, generating multi-scale feature representations that combine local texture and global structure. To accurately capture the geometric boundary information of changing regions, a boundary enhancement module is introduced during the decoding process. This module uses boundary attention guidance to dynamically focus on potential changing boundary regions, improving the model's accuracy in depicting irregular, fine-grained changing boundaries such as the edges of newly built roads and building outlines. Finally, after feature optimization by multiple convolutional modules, the multi-scale features are scale-normalized by a size unification module. The processed features are then input into the prediction head, ultimately outputting a pixel-level change detection result map.

[0079] 3. Boundary Enhancement Module

[0080] The EdgeSAM image encoder, combined with a spatiotemporal feature alignment module, possesses high-level semantic understanding capabilities, effectively identifying real-world changing objects in remote sensing images. However, it still suffers from insufficient sensitivity when processing local boundary features of changing objects, leading to blurring and breaks in boundary details. This invention employs a boundary enhancement module designed using the SimAM attention mechanism and feature difference operations to strengthen the boundaries of extracted features, ensuring the model can more accurately identify and depict the boundaries of target regions. The SimAM attention mechanism adaptively derives 3D attention weights for feature maps without introducing additional parameters. Specifically, inspired by classical neuroscience theory, the SimAM attention mechanism quantifies the salience of neurons by constructing an energy function, thereby automatically identifying their importance. Based on this, another significant advantage of the SimAM attention module is that its core operations are derived from the defined energy function, requiring no complex structural design, thus perfectly meeting the model's stringent requirements for parameter efficiency.

[0081] The module structure is as follows: Figure 4 As shown. Assume the input features of the boundary enhancement module are... and The output is .

[0082] First, and After element-wise subtraction, differential feature maps are generated through convolutional layers and the ReLU nonlinear activation function. It can be expressed by the following formula:

[0083]

[0084] At the same time, and The results are obtained by inputting the SimAM attention module respectively. and It can be represented by the following formula:

[0085]

[0086] The SimAM attention module uses the SimAM (Simple Attention Module) attention mechanism to enhance or suppress the input features of the module element-wise using 3D attention weights, thereby obtaining the output features of the module.

[0087] and It contains important edge information of the initial input features of this module. (The remaining text appears to be incomplete and requires further context.) and Edge features are extracted through a combination of average pooling, subtraction, convolution, and the sigmoid function. and In this process, edge regions typically have higher gradient values, so average pooling is used to smooth the feature map. The subtraction operation involves calculating the difference between the features before and after the average pooling operation to enhance edge regions. For example, the features before the average pooling operation... Features after average pooling operation Difference calculation is performed to obtain The following formula represents:

[0088]

[0089] in, Indicates average pooling; Represents convolution; This represents the Sigmoid activation function.

[0090] and It improved the contrast of edge features. , respectively with , After further enhancing the edge features through multiplication and addition operations, the data is input into the SimAM attention module again to obtain... , ;Will , Subtraction yields the difference edge features As shown below:

[0091]

[0092] Difference edge features and differences in characteristics After addition, the boundaries are further enhanced using the SimAM attention module to obtain enhanced boundary features. As shown below:

[0093]

[0094] Thanks to the improvement in temporal consistency by the spatiotemporal feature alignment module and the enhancement of boundary details by the boundary reinforcement module, Edge-CD achieves coordinated optimization of the discrimination accuracy of real changes and boundary integrity while maintaining a low number of parameters.

[0095] Step S2: Based on the remote sensing image dataset, train the Edge-CD model to obtain the trained Edge-CD model;

[0096] In this embodiment, a self-made remote sensing image dataset, MSRS-CD, is used and proportionally divided into training, validation, and test sets. This dataset, collected from urban and suburban changes in southern Chinese cities between 2019 and 2023, contains 842 pairs of remote sensing images, each 1024×1024 pixels in size with a spatial resolution of 0.5 meters. The MSRS-CD dataset was released in 2025, and the main types of changes include new construction, suburban expansion, vegetation changes, and road construction. The number of image pairs in the training, validation, and test sets are 588, 84, and 169, respectively.

[0097] The training set is input into the Edge-CD model for training.

[0098] In dual-temporal remote sensing image change detection tasks, accurate localization of changed regions depends not only on the model's ability to discriminate the semantics of ground features, but also on the geometric clarity and topological integrity of the change boundaries. However, traditional training methods typically employ regional supervision signals and optimize using cross-entropy loss or Dice loss. While these loss functions achieve good performance on overall segmentation metrics, they are essentially statistical measures of pixel-level classification results, lacking explicit modeling of spatial structure and boundary geometric consistency. Their fundamental limitation is that all positive sample pixels are treated equally, regardless of whether they are located inside the changed region or near the boundary. This is insufficient in remote sensing scenarios. When faced with complex structures such as narrow roads or broken building complexes, prediction results often exhibit problems such as blurred boundaries, jagged breaks, local misalignment, and even topological distortion, thus affecting subsequent remote sensing image monitoring and practical decision-making.

[0099] To alleviate the aforementioned problems, this invention proposes a boundary-aware loss mechanism. By constructing a boundary-aware loss function, constraints are imposed on changing boundaries during the model training phase, thereby guiding the network to focus more intently on learning the boundary features of changing regions. The core idea of ​​this edge feature constraint strategy is to increase the weight of boundary pixels. The specific steps are as follows: The image domain of the change mask is set as... The manual annotations of the actual changes are First, regarding the real labels Gaussian smoothing is applied to suppress noise during the annotation process; then, the prediction boundary is extracted using morphological gradients. Then, an expansion operation is performed on the predicted boundary to form a "boundary band" of a certain width, denoted as... And define the boundary-aware loss function. for:

[0100]

[0101] in, Indicates the boundary range of the actual area of ​​change; Represents points on the prediction boundary; Represents a point on the actual boundary; Points on the prediction boundary Inside the real boundary, Points on the prediction boundary Outside the true boundary; inside the true boundary represents the region of change, and outside the true boundary represents the region of no change. A schematic diagram illustrating the operational principle of the boundary-aware loss function is shown below. Figure 5 As shown:

[0102] To further enhance the synergy between boundary information and semantic features, the boundary-aware loss function is... Joint optimization is performed with the traditional region loss function. To guide model learning, the boundary-aware loss function aims to make the predicted boundary infinitely close to the true boundary; the traditional region loss function aims to make the predicted changed region infinitely close to the true changed region. During training, let the change probability map output by the model (i.e., the change detection result) be... The main task of change detection (traditional region loss) adopts binary cross-entropy loss. The joint loss function is then:

[0103]

[0104] in, The represents the balance coefficient, used to coordinate the trade-off between semantic classification accuracy and boundary geometric fidelity. This joint loss function assigns higher weights to pixels near the boundary, explicitly penalizing boundary positional shifts and shape distortions, thereby driving the model to learn sharper and more accurate boundary responses. Through this boundary-aware loss mechanism, the Edge-CD model achieves geometric boundary constraints during the training phase.

[0105] Step S3: Input remote sensing images from different time phases into the trained Edge-CD model to obtain the interpretation results;

[0106] The test set is input into the trained Edge-CD model, and Precision, Recall, Intersectionover Union (IoU), Overall Accuracy (OA), and F1 score are selected as evaluation metrics.

[0107] Among them, Precision focuses on the accuracy of the model's predictions of positive examples, Recall emphasizes the model's ability to capture positive examples, IoU is a comprehensive metric that measures the overlap between the predicted and actual regions of change, OA evaluates the overall accuracy of the model's classification, and F1 score balances the relationship between Precision and Recall. These evaluation metrics are calculated as follows:

[0108]

[0109]

[0110]

[0111]

[0112]

[0113] in, This indicates a true positive, which is the number of pixels that were predicted to change and actually did change. This indicates a true negative, meaning the number of pixels that are predicted to remain unchanged and actually remain unchanged. This indicates a false negative, meaning the number of pixels predicted to have changed but actually remaining unchanged. This indicates a false positive, which is the number of pixels that were predicted to remain unchanged but actually changed. Figure 6 This is a partial schematic diagram of the Edge-CD detection results.

[0114] The average test results obtained are shown below:

[0115]

[0116] Experimental results confirm that Edge-CD performs well in suppressing spurious changes, maintaining boundary clarity and structural integrity, and achieves good overall performance, providing a new method for high-precision remote sensing change detection.

[0117] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A remote sensing image interpretation method based on a lightweight visual fundamental model, characterized in that, include: Step S1: Construct the Edge-CD remote sensing image change detection model based on the lightweight visual foundation model EdgeSAM; Step S2: Based on the remote sensing image dataset, train the Edge-CD model to obtain the trained Edge-CD model; Step S3: Input remote sensing images from different time phases into the trained Edge-CD model to obtain the interpretation results; In step S1, the Edge-CD model includes: an EdgeSAM encoder unit and a boundary enhancement decoding head unit; the EdgeSAM encoder unit includes: an EdgeSAM image encoder and a spatiotemporal feature alignment module; the boundary enhancement decoding head unit includes: a boundary enhancement module (EEM), a convolution module, a size unification module, and a prediction head; the EEM includes a SimAM attention module. dual-temporal remote sensing images and The EdgeSAM encoder unit extracts spatiotemporal alignment features respectively. and ;Will and Features obtained by splicing Input the boundary enhancement decoding head unit to obtain the change detection results; In the EdgeSAM encoder unit, the single-input EdgeSAM image encoder is improved into a twin structure with two dual-temporal remote sensing image inputs. That is, the dual-temporal remote sensing images... and The first EdgeSAM image encoder and the second EdgeSAM image encoder are input respectively, and the two EdgeSAM image encoders share weights. In the EdgeSAM image encoder, multi-level feature maps are generated through 12 encoder modules. Each encoder module is considered as one layer; that is, the first encoder module of the first EdgeSAM image encoder and the first encoder module of the second EdgeSAM image encoder both belong to the first layer, the second encoder module of the first EdgeSAM image encoder and the second encoder module both belong to the second layer, and so on, for a total of 12 layers. This invention embeds a spatiotemporal feature alignment module after each encoder module in layers 1 to 11, while the last layer does not contain a spatiotemporal feature alignment module. The output of the last encoder module is used as the output of the EdgeSAM encoder unit. In other words, dual-temporal remote sensing images... and After passing through the first encoder module of the first EdgeSAM image encoder and the second EdgeSAM image encoder respectively, the two feature maps obtained by the first encoder module are input together into the first spatiotemporal feature alignment module. The two output features of the first spatiotemporal feature alignment module are then used as inputs into the second encoder module. This process continues until the output of the last encoder module is used as the output of the EdgeSAM encoder unit.

2. The method according to claim 1, characterized in that, The encoder module is a RepViT Block.

3. The method according to claim 2, characterized in that, When describing the structure of the spatiotemporal feature alignment module, assume that the input of the module is... and The output is and ; First, input and Feature extraction and enhancement are performed separately using a CBG module to obtain... and The CBG module includes, in sequence, convolution, normalization, and activation functions; subsequently, and Compared with the original input and Element-wise multiplication is performed to enhance the spatiotemporal feature information and obtain the result. and This process can be represented by the following formula: in, This indicates that the CBG module is processing the data. This indicates element-wise multiplication; Then, and After each of the two processes undergoes a 1×1 convolution, we obtain... and After that, Spatiotemporal characteristics in and Perform element-wise addition, while... Spatiotemporal characteristics in and By adding element by element, we get and The following indicates: in, This represents a 1×1 convolution; This indicates element-wise addition.

4. The method according to claim 3, characterized in that, When describing the structure of the boundary enhancement decoding head, firstly, the input features of the boundary enhancement decoding head unit are... The convolutions are processed through max pooling downsampling, identity mapping layer, and two transposed convolutions of different sizes. Among them, features The feature dimension is C×H×W; the feature dimension output after max pooling downsampling is C×H / 2×W / 2, the feature dimension output after the identity mapping layer is C×H×W, the feature dimension output after the first transposed convolution is C / 2×2*H×2*W, and the feature dimension output after the second transposed convolution is C / 4×4*H×4*W, thus obtaining a hierarchical multi-scale feature. , , and Secondly, the features The feature representation is refined through convolutional modules to obtain features. ,in, Then, let , ;Will and The input EEM enhances the focus on key edge details of multi-scale features, and the output boundary enhancement features. ;Will and Input to EEM, output boundary enhancement features After that, After passing through the size unification module, the data is concatenated to obtain the final features. The final features are then input into the prediction head to output a change mask. The convolution module includes convolution, normalization, and ReLU activation functions. The size unification module uses 1×1 convolution and bilinear interpolation to unify the size to the same number of channels and spatial dimensions, providing a consistent scale for feature fusion and decoding. The prediction head maps the channel dimensions of the final features to the number of target categories through 1×1 convolutional layers, outputting the change detection results.

5. The method according to claim 4, characterized in that, When describing the boundary enhancement module, assume that the input features of the module are... and The output is ; First, and After element-wise subtraction, differential feature maps are generated through convolutional layers and the ReLU nonlinear activation function. It can be expressed by the following formula: At the same time, and The results are obtained by inputting the SimAM attention module respectively. and It can be represented by the following formula: The SimAM attention module uses the SimAM attention mechanism to enhance or suppress the input features of the module element by element using 3D attention weights, thereby obtaining the output features of the module. Each and Edge features are extracted through a combination of average pooling, subtraction, convolution, and the sigmoid function. and The subtraction operation involves performing a difference calculation on the features before and after the average pooling operation to enhance edge regions. For example, it involves subtracting the features before the average pooling operation... Features after average pooling operation Difference calculation is performed to obtain The following formula represents: in, Indicates average pooling; Represents convolution; This represents the Sigmoid activation function; Will , respectively with , After further enhancing the edge features through multiplication and addition operations, the data is input into the SimAM attention module again to obtain... , ;Will , Subtraction yields the difference edge features As shown below: Difference edge features and differences After addition, the boundaries are further enhanced using the SimAM attention module to obtain enhanced boundary features. As shown below: 。 6. The method according to claim 5, characterized in that, In step S2, during the training phase, the image domain of the change mask is set to... The manual annotations of the actual changes are First, regarding the real labels Gaussian smoothing is applied; then, the prediction boundary is extracted using morphological gradients. Then, an expansion operation is performed on the predicted boundary to form a "boundary band" of a certain width, denoted as... And define the boundary-aware loss function. for: in, Indicates the boundary range of the actual area of ​​change; Represents points on the prediction boundary; Represents a point on the actual boundary; Points on the prediction boundary Inside the real boundary, Points on the prediction boundary Outside the true boundary; inside the true boundary represents a region of change, and outside the true boundary represents a region of no change; Boundary-aware loss function Joint optimization is performed with the traditional region loss function; where the traditional region loss function uses binary cross-entropy loss, i.e. The joint loss function is then: in, This represents the balance coefficient.

7. The method according to any one of claims 1-6, characterized in that, The remote sensing image dataset is MSRS-CD.

Images