Remote sensing image change detection method and system based on dual-branch attention and boundary enhancement of Mama architecture

Abstract

The invention discloses a dual-branch attention and boundary enhanced remote sensing image change detection method and system based on a Mama framework, and belongs to the technical field of computer vision and remote sensing image processing. The method comprises the following steps: carrying out cutting, data set division and normalization preprocessing on a dual-tense remote sensing image; multi-scale features are extracted through a Mamba encoder, and global spatial dependence is modeled by using a state space model; constructing a boundary enhancement branch, and combining channel compression and up-sampling to generate a boundary probability graph; in the decoding stage, a space-time state space (STSS) module is introduced to fuse space-time features, and a dynamic channel attention mechanism is embedded to strengthen a change sensitive area; using deformable convolution to optimize edge offset, and performing weighted fusion on main branch and boundary branch output to generate a final change graph; and training the model through a composite loss function and optimizing parameters. According to the invention, the detection precision of a change area in a complex scene is significantly improved, and the problems of boundary blur, small target leak detection and background interference are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision and remote sensing image processing, and specifically relates to a remote sensing image change detection method and system based on Mamba architecture dual-branch attention and boundary enhancement. Background Art

[0002] Remote Sensing Change Detection (RSCD) aims to analyze the image difference information of the same geographical area at different time phases, and has important application value in urbanization monitoring, ecological environment assessment, disaster emergency response and other fields.

[0003] Traditional RSCD methods rely primarily on manual visual interpretation or spectral difference analysis (such as change vector analysis and image interpolation). These methods require manual threshold setting and rely on expert experience, resulting in low efficiency and high subjectivity. Subsequently, researchers introduced machine learning algorithms such as support vector machines (SVMs) and random forests to improve automation, but their feature representation capabilities remain limited by manually designed feature operators.

[0004] In recent years, deep learning-based methods have significantly improved RSCD performance. For example, Daudt et al. first introduced fully convolutional networks (FCNs) to this task and proposed twin network architectures such as FC-EF and FC-Siam-conc. However, due to the local limitations of convolution operations, these methods have difficulty modeling the global spatiotemporal correlations of changing regions in remote sensing images. Chen et al. further combined Transformer with CNN (such as the BIT network) to enhance the ability to model long-range dependencies through a self-attention mechanism. However, the complexity of traditional self-attention is at the square of the input size, making it difficult to adapt to the needs of high-resolution image processing.

[0005] The Mamba architecture, based on the state-space model (SSM), has gradually emerged in computer vision tasks due to its low linear computational complexity and global context modeling capabilities. However, when this architecture is directly applied to change detection in remote sensing images, it still exposes a series of practical challenges. For example, the model lacks an adaptive calibration mechanism for feature channels during the decoding process, making it easily affected by complex backgrounds such as shadows and seasonal changes in vegetation, resulting in frequent false alarms. At the same time, due to the local limitations of the upsampling process, the edges of the changing areas are prone to blurring or breaking, and the detection accuracy of fine boundaries such as building outlines and road shapes is significantly insufficient. In addition, existing methods are relatively crude in the design of multi-scale feature fusion, making it difficult to effectively capture the change signals of small targets, and the problem of missed detection is prominent. These problems seriously restrict the application value of the Mamba architecture in actual remote sensing scenarios.

[0006] In summary, there is an urgent need for a remote sensing image change detection method that maintains the efficiency of Mamba while improving detection accuracy through a lightweight attention mechanism and boundary enhancement strategy. Summary of the Invention

[0007] The purpose of the present invention is to address the problems of blurred boundaries, missed detection of small targets, and sensitivity to background interference in existing remote sensing image change detection methods when processing complex scenes. The present invention proposes a remote sensing image change detection method and system based on the Mamba architecture with dual-branch attention and boundary enhancement. Through dynamic channel calibration and edge enhancement strategies, the integrity detection capability of the changed area and the robustness of the algorithm are significantly improved.

[0008] The purpose of the present invention is achieved through the following technical solutions:

[0009] A binary change detection method for remote sensing images based on dual-branch attention and boundary enhancement. The specific steps are as follows:

[0010] Step 1: Data extraction and preprocessing: crop the original remote sensing image into non-overlapping sub-images of a predetermined size to ensure spatial alignment of the bi-temporal image and the label map; randomly divide the sub-images into training, validation, and test sets, and perform image normalization and data augmentation.

[0011] Step 2: Feature extraction; Main branch encoding, first extract the multi-scale features of the dual-temporal images in the training set through the Mamba encoder, the encoder contains multiple levels, each level gradually reduces the feature map resolution through downsampling operation, and in each level, the feature map is modeled for global spatial dependency through the state space model; boundary enhancement branch encoding, for the high-dimensional features output by the main branch, the spatial resolution of the feature map is gradually restored through channel compression and upsampling operations to generate a single-channel boundary probability map P main ;

[0012] Step 3: Decoding and spatiotemporal relationship modeling: Model the spatiotemporal relationship of the multi-temporal features output by the spatiotemporal state space (STSS) module encoder stage, integrate the feature maps of different levels through the fusion module, and embed the channel attention mechanism to enhance the change-sensitive features;

[0013] Step 4: Edge optimization and probability map generation; the main branch uses the Sobel operator to capture the initial edge response E base , combined with deformable convolution to optimize edge offset and generate the final edge map E final ;

[0014] Step 5: The final edge graph E generated by the main branch final The boundary probability map P generated by the boundary enhancement branch mainThe final change map is generated through weighted fusion. The error between the predicted result and the actual label is fed back into the network to adjust the network parameters and obtain a trained change detection model.

[0015] Step 6: Input the validation set into the trained change detection model to preliminarily evaluate the model performance and make parameter adjustments. If the validation set indicators meet the standards, proceed to step 7. If the indicators do not meet the standards, return to step 1.

[0016] Step 7: Input the images in the test set into the finalized change detection model to obtain the final prediction results for subsequent analysis and application of remote sensing images.

[0017] Furthermore, in step 1, the original remote sensing image is cropped into non-overlapping sub-images of size 256×256, and the sub-image set obtained after cropping is divided into the following types according to the conventional deep learning task ratio: the ratio of training set, validation set and test set is 6:2:2.

[0018] Furthermore, the Mamba encoder in step 2 is an encoder with two groups of weight sharing, and each group of encoding parts includes four encoding stages; each Mamba encoder is composed of a cascade of four levels of Mamba blocks, which gradually extract multi-scale features; a convolution layer with a stride of 2 is used before each level of Mamba block for 2x downsampling, and the level output size is one-half, one-quarter, one-eighth and one-sixteenth of the height and width of the original image respectively; within the Mamba block, the input feature map is spatially serialized and expanded, and the global spatial dependency is modeled through a state space model; the features of the pre-event and post-event images extracted from the four encoding stages will be used for subsequent decoders responsible for specific tasks; the boundary enhancement branch coding uses a 1×1 convolution layer to compress the number of channels of the feature map to 1 through the boundary prediction layer module.

[0019] Furthermore, the step 3 includes four decoding stages. At the beginning of each decoding stage, the spatiotemporal state space STSS module models the spatiotemporal relationship of the multi-temporal features input in the encoder stage, and the spatiotemporal token generation module rearranges the input multi-temporal features and inputs them into three VSS modules; the output of the STSS module in the current stage is then integrated with the information of the feature map of the previous stage through the fusion module; the attention mechanism SE module is embedded in the fusion module, and the number of channels of the low-level feature map is mapped to the same as the high-level feature map through a 1×1 convolution layer. Figure 1 Then, the high-level and low-level feature maps are summed; finally, the obtained feature map is smoothed by the residual layer; a new step is added between the fusion module and the upsampling layer, and the smoothed feature map undergoes global average pooling, MLP compression and restoration, and channel weighting steps in the SE module to enhance the network's attention to important features.

[0020] Furthermore, the smoothed feature map is globally average pooled in the SE module, and the formula is as follows:

[0021] w c =GAP(F)

[0022] Among them, F is the feature map, F∈R H×W×C ;

[0023] Generate channel weights through two fully connected layers and Sigmoid function:

[0024] w c =σ(W2(δ(W1z)))

[0025] Where z∈R C Represents the feature vector obtained by global average pooling; represents the weight matrix of the first fully connected layer, r is the scaling factor; represents the weight matrix of the second fully connected layer; δ(·) represents the ReLU activation function; σ(·) represents the Sigmoid activation function, which is used to generate channel weights in the range of [0,1]; w c ∈R C Represents the weight of each channel;

[0026] Multiply the weights by the original features channel by channel to obtain the enhanced feature map F out :

[0027] F out =w c ⊙F

[0028] Furthermore, the step 4 uses the Sobel operator to capture the initial edge response:

[0029] E base =Sobel(P main )

[0030] Optimize edge offset using deformable convolution:

[0031]

[0032] Among them, E final represents the output boundary graph; E base represents the input feature map; K represents the convolution kernel size; p k represents the offset on the standard convolution kernel; Δp k represents the learnable offset; w k Represents the convolution kernel weight; p0 represents the convolution center point;

[0033] Furthermore, the step 5 converts the boundary graph E final With the main branch output Pmain Pixel-by-pixel weighting, the formula is:

[0034] P final =P main ⊙(1+α·E final )

[0035] Where α∈[0.5,1.0].

[0036] Furthermore, the final predicted change map obtained in step 6 is compared with the true label map in the test set, and the probability TP of actually being predicted as positive, the probability TN of actually being predicted as negative, the probability FP of actually being predicted as positive, and the probability FN of actually being predicted as positive are respectively counted to calculate the actual performance indicators of our trained remote sensing image change detection network.

[0037] A computer device / equipment / system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of a remote sensing image binary change detection method based on dual-branch attention and boundary enhancement.

[0038] A computer-readable storage medium stores a computer program / instruction, which, when executed by a processor, implements the steps of a remote sensing image binary change detection method based on dual-branch attention and boundary enhancement.

[0039] The beneficial effects of the present invention are:

[0040] The specific advantages of the remote sensing image change detection method of the present invention are reflected in the following three aspects: A dynamic channel attention mechanism is embedded in the feature decoding stage. Channel weight vectors are generated through global average pooling of cross-layer feature maps, automatically enhancing the feature responses of change-sensitive channels. For example, when processing building changes in vegetation-covered areas, this mechanism significantly enhances attention to high-frequency changes such as roof edges, while suppressing interference from low-frequency background noise such as diffuse cloud reflections, reducing the risk of missed detections at the feature expression level. An independent boundary enhancement branch enables targeted optimization of the edges of changing regions. A deformable convolutional layer replaces the traditional fixed sampling grid. The position of feature sampling points is dynamically adjusted based on the preliminary change probability map output by the main branch, achieving sub-pixel localization of building fracture edges and narrow road structures. A global feature extraction network is constructed based on the Mamba dual-temporal encoder, utilizing a state-space model to replace the traditional self-attention mechanism. This design uses discretized state equations to model the spatiotemporal dependencies of image sequences, efficiently capturing the long-range context of large-scale images while maintaining linear computational complexity. Compared to the Transformer, training memory usage is reduced and inference speed is improved at the same input resolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is the overall network architecture diagram of the remote sensing image change detection method based on the dual-branch attention and boundary enhancement of the Mmaba architecture of the present invention.

[0042] Figure 2 It is a detailed diagram of the encoder structure in the remote sensing image change detection method of the present invention.

[0043] Figure 3 It is a detailed diagram of the decoder structure in the remote sensing image change detection method of the present invention.

[0044] Figure 4 It is a detailed structural diagram of each module of the decoder part of the present invention.

[0045] Figure 5 This is a diagram of the boundary-enhanced branch structure of the present invention.

[0046] Figure 6 This is a flow chart of the SE module and loss calculation of the present invention. DETAILED DESCRIPTION

[0047] The present invention will be further described below with reference to the accompanying drawings.

[0048] To address the problems of blurred boundaries, missed small targets, and sensitivity to background interference in existing remote sensing image change detection methods when processing complex scenes, this paper proposes a dual-branch decoder network based on the Mamba architecture, which achieves high-precision change detection through channel attention-guided feature selection and boundary enhancement mechanism.

[0049] The present invention provides a remote sensing image binary change detection method based on dual-branch attention and boundary enhancement. The overall network framework is based on Figure 1 As shown in , the network consists of two sets of weight-sharing encoders (four-layer encoding), one set of decoders (four-layer decoding main branches) and one set of boundary branches. It should be noted that Figure 1 The jump connection between the encoder and decoder is omitted in the figure. The details can be found in Figure 2 and Figure 3 Understand in. Figure 1 The two green modules represent the feature images at different times of the input. The next two groups of four purple modules each represent the four stages of the encoding module. It should be noted that the weights of the two groups of encoder modules are shared. Next, the boundary enhancement branch structure diagram is adopted. The main branch corresponds to a group of four dark blue decoding stages. This decoder is improved and the SE module is introduced in all four decoding stages. Figure 3As can be seen in the figure, the boundary branch contains two bright blue modules, namely the depthwise separable convolution and boundary prediction layers, which are boundary enhancement branch structures. The main branch ultimately outputs the main change map of the yellow module, and the boundary branch ultimately outputs the boundary map of the yellow module. The two are weightedly fused to generate the final change map of the orange module.

[0050] The present invention carries out a systematic processing flow for the original change detection remote sensing image, including image segmentation, data set partitioning and preprocessing. First, due to the high resolution of the original remote sensing image, directly inputting it into the change detection model not only consumes a lot of computational resources, but may also cause local change information to be ignored. Therefore, we crop the original image into non-overlapping sub-images of size 256×256 to enhance the model's ability to learn local area changes. During cropping, two phases of remote sensing images and their corresponding change label maps are processed simultaneously to ensure spatial alignment of the three. The sub-image set obtained after cropping is divided into the ratio of training set, validation set and test set according to the conventional deep learning task ratio: the ratio of 6:2:2. The division method is guaranteed to be random division. Data preprocessing operations such as data augmentation and normalization are performed before the image is input into the change detection network.

[0051] The initial input to the overall network architecture is a set of two feature images of the same region at different times, selected from the training, validation, and test sets. All images in the training set are used to train the deep learning change detection network. All images in the validation set are used to verify that the trained network achieves the desired results. All images in the test set are used to output the final performance metrics of the network model. This illustrates the respective roles of these three datasets in the change detection task.

[0052] The encoder structure of the present invention is shown in FIG. Figure 2 As shown, this is a detailed diagram of the encoder structure of the present invention, the four stages of the encoding part Stage I to Stage IV, and their detailed composition is as follows Figure 2 , each stage first downsamples the input data, and then uses multiple visual state space VSS modules to fully model the spatial context information. The structure of the VSS module is also as follows Figure 2 As shown. The input first passes through a linear embedding layer, and the output is divided into two information streams. One stream passes through a 3×3 depth convolution (DWConv) layer, followed by a Silu activation function, and then enters the core SS2D module (that is, the integration of S6 and cross-scanning mechanism). The output of the SS2D module passes through a layer normalization (LN) layer and is then added to the output of other streams that have passed Silu activation. This combination produces the final output of the VSS block. Finally, the features of the four stages are used in the subsequent decoder to perform specific tasks. It should be noted that in the encoding part, the image features of different heights, widths and numbers of channels output by the four stages are input into the decoder structure through jump connections. Figure 3 Finally, the features of the pre-event and post-event images extracted from the four stages will be used for subsequent decoders responsible for specific tasks. Among them, each Mamba encoder is composed of 4 levels of Mamba blocks (En_block1~4) cascaded, gradually extracting multi-scale features. A convolutional layer with a stride of 2 is used before each level of Mamba block for 2x downsampling, and the level output size is one-half, one-quarter, one-eighth and one-sixteenth of the height and width of the original image, respectively. Inside the Mamba block, the input feature map is spatially serialized and expanded, and the global spatial dependency is modeled through the state space model.

[0053] The decoder structure of the present invention is as follows Figure 3 The detailed diagram of each component of the decoder is shown in Figure 4 As shown, in Figure 3 In this paper, the network fully learns the spatiotemporal relationships from the extracted multi-temporal features in four stages and obtains accurate binary change maps. At the beginning of each stage, the spatiotemporal relationships of the multi-temporal features are first modeled using the STSS module. In the STSS module, a spatiotemporal tagger module rearranges the input multi-temporal features and then inputs them into three VSS modules. Each block is responsible for learning a spatiotemporal relationship. The output of the STSS module in the current stage is then integrated with the feature map information of the previous stage through a fusion module. After passing through the upsampling layer, the feature map is input to the next stage.

[0054] refer to Figure 3 As can be seen, the decoder in this invention introduces the SE module in all four decoding stages, which is placed between the fusion module and the upsampling module. The SE module is an attention mechanism used in convolutional neural networks. It enhances the network's response to important features and suppresses unimportant features by adaptively recalibrating the importance of each channel. Its core idea is to enhance the ability to express features through two steps: compression and excitation, thereby improving network performance.

[0055] The boundary enhancement branch junction reference of the present invention Figure 5 It can be seen that it adds a boundary enhancement branch on the basis of the main branch to improve the ability to identify the boundaries of objects in the image, which is especially suitable for tasks such as image segmentation and target detection. The structure diagram contains multiple processing steps to enhance boundary information and optimize boundary prediction. Figure 5 In the encoding stage, the input feature map comes from Stage IV, which contains the high-dimensional feature information of the image, which is crucial for subsequent boundary enhancement operations.

[0056] The input high-dimensional feature information first undergoes depthwise separable convolution. Depthwise convolution is the first step of the convolution operation, in which each input channel is independently convolved. The convolution kernel size of this convolution operation is 3×3, which means that each input channel is convolved with its own convolution kernel. Since there is no cross-channel information exchange, it is relatively computationally efficient. Pointwise convolution is the second step of the convolution operation, in which each pixel is processed using a 1×1 convolution kernel to integrate the features of each channel. The number of channels is compressed to 64. This operation reduces the number of channels and retains a more compact information representation.

[0057] The upsampling module then passes through the feature map. The upsampling process aims to gradually increase the spatial resolution of the feature map, restore the spatial structure of the image, and adapt it to the final boundary prediction. The upsampling operation is performed using bilinear interpolation, and each upsampling doubles the resolution of the feature map. The first upsampling doubles the spatial resolution, the second upsampling further increases the spatial resolution, and the third upsampling further restores the spatial information. Through these multiple upsampling, the spatial resolution of the feature map is gradually increased, which helps to obtain more detailed spatial information in the final boundary prediction.

[0058] At the end of the upsampling path, a 1×1 convolution is used to compress the number of channels in the feature map to 1, outputting a single-channel boundary probability map. This is the boundary prediction layer module. This allows the network to learn the boundary regions of objects in the image, thereby improving task accuracy, which is particularly important in segmentation and boundary detection. The boundary enhancement branch structure enhances image boundary information through operations such as depthwise separable convolution, upsampling, and boundary prediction. It enhances the boundary features of the image and ultimately outputs a boundary probability map. The boundary probability map is weightedly fused with the main change map to generate the final change map, significantly improving the model's performance in image segmentation and object detection tasks. By generating ground truth boundary values, the model can be effectively trained to learn how to predict and recognize object boundaries.

[0059] SE module and loss calculation process of the present invention Figure 6As can be seen, the input feature map undergoes global average pooling, MLP compression and restoration, and channel weighting in the SE module, enhancing the network's focus on important features and suppressing the influence of unimportant features. This mechanism, through global information compression and channel importance incentives, can significantly improve network performance in various computer vision tasks. It can adaptively adjust the weight of each channel, allowing the network to focus more on important features and suppress unimportant ones. This improves the network's responsiveness to important features, resulting in better performance in various tasks. It also provides a globally adaptive feature selection mechanism that does not rely on hand-designed rules, but instead automatically selects which channels are more important through learning. In the loss calculation flowchart, the weighted sum of the cross-entropy loss and the Lovasz loss generates the main branch loss, which is then combined with the boundary branch loss to generate the total loss.

[0060] Experimental analysis of the present invention

[0061] In order to verify the effect of the method of the present invention, the following two experimental analyses were performed:

[0062] 1. Experimental Analysis of SYSU Dataset

[0063] The comparison results with the other ten models on the SYSU dataset are shown in Table 1. The experimental results show that C-Mmaba, the method of the present invention, performs excellently in multiple key performance indicators.

[0064] C-Mamba achieved an exceptional precision of 88.24%, ranking first among all models. Precision measures the proportion of pixels predicted by the model as changed that are actually changed. C-Mamba's high score on this metric demonstrates its ability to effectively reduce false positives and accurately identify changed regions. In comparison, other methods such as MambaBCD (87.91%) and BIT (81.04%) achieved lower precision, demonstrating C-Mamba's superiority in accurately identifying changed regions. In terms of recall, C-Mamba achieved 79.81%, ranking first among all models, demonstrating its ability to comprehensively identify true changed regions. Recall measures whether a model correctly identifies all actual changed regions, and C-Mamba performed exceptionally well in this regard. Although STANet achieved a slightly higher recall (80.51%), C-Mamba's balanced precision and recall made it stand out overall. C-Mamba achieved an F1 score of 83.76%, placing it among the top models. The F1 score combines precision and recall. C-Mamba's performance on this metric demonstrates that it strikes a good balance between optimizing precision and recall, reducing false positives while effectively capturing true areas of change. In comparison, MambaBCD (83.34%) and other models are slightly inferior to C-Mamba, further demonstrating C-Mamba's overall performance advantage. C-Mamba excels in the Intersection over Union (IoU) metric, achieving 72.51%, surpassing all other models. C-Mamba achieves 92.71% in Overall Accuracy (OA), surpassing all other models and further highlighting its strong advantage in overall image recognition. C-Mamba achieves a Kappa coefficient of 79.16%, ranking first among all models. C-Mamba's high Kappa coefficient demonstrates its high reliability and consistency in multi-category change detection.

[0065] The outstanding performance of the method of the present invention is due to the advanced technologies adopted in its model architecture, such as attention-guided feature selection and boundary enhancement mechanism, which significantly improve the model's adaptability and recognition accuracy to complex remote sensing image changing scenes.

[0066] Table 1 Quantitative comparison results with other SOTA models on the SYSU dataset.

[0067]

[0068] 2. Experimental Analysis of the LEVIR-CD+ Dataset

[0069] Experimental results on the LEVIR-CD+ dataset demonstrate that C-Mamba excels across multiple key performance metrics, surpassing seven other models. C-Mamba achieves a high precision score of 88.73%, ranking first among all models. This metric measures the proportion of pixels predicted by the model as changed that are actually changed. C-Mamba's superior precision demonstrates its ability to effectively reduce false positives and accurately identify actual changed regions. MambaBCD follows closely behind with an accuracy of 87.64%, slightly inferior to C-Mamba. Other models, such as BIT and FC-Siam-Diff, achieve accuracies of 84.67% and 80.51%, respectively, demonstrating C-Mamba's clear advantage in accurately identifying changed regions. C-Mamba's recall of 86.72% ranks first among all models, demonstrating outstanding performance. The F1 score is the harmonic mean of precision and recall, which comprehensively considers these two key metrics. C-Mamba's outstanding performance demonstrates that it strikes an ideal balance between precision and recall. Interference over Union (IoU) is a metric that evaluates the degree of overlap between a model's predicted and true change regions. A higher IoU indicates that C-Mamba has a significant advantage in segmenting change regions. C-Mamba achieved a high OA (overall accuracy) score of 99.08%, leading all other models. Overall accuracy measures the model's overall classification accuracy. C-Mamba's high OA score demonstrates its ability to effectively identify change regions in complex remote sensing imagery and also achieves high accuracy in classifying non-change regions. Other models, such as MambaBCD and SwinSUNet, achieved OA scores of 98.96% and 98.41%, respectively, which are close but still slightly lower than C-Mamba. C-Mamba achieved a Kappa coefficient of 87.10%, ranking first among all models. The Kappa coefficient measures the consistency between a model's predictions and the true labels. A higher Kappa coefficient indicates a model's high stability and reliability in classification consistency. C-Mamba's high Kappa coefficient demonstrates its high consistency in detecting change regions. The Kappa coefficients of other methods such as MambaBCD and SwinSUNet are 86.56% and 84.58% respectively. Although they also perform well, they are still slightly inferior to C-Mamba.

[0070] Table 2 Quantitative comparison results with other SOTA models on the LEVIR-CD+ dataset.

[0071]

[0072] Experimental results on both the SYSU and LEVIR-CD+ datasets demonstrate that C-Mamba performs exceptionally well across multiple metrics, including precision, recall, F1 score, IoU, overall accuracy, and Kappa coefficient, outperforming all other models. By incorporating innovative techniques such as boundary enhancement and global feature fusion using multi-scale attention machines, C-Mamba demonstrates outstanding performance in remote sensing image change detection. Its efficient computational power and robustness make it a significant breakthrough in the field of change detection, promising promising applications.

[0073] The detection method of the present invention described above can be converted into software program instructions, which can be implemented by running a software analysis system including a processor and a memory, or can be implemented by computer instructions stored in a non-transitory computer-readable storage medium.

[0074] Finally, the method of the present invention is only a preferred embodiment and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A binary change detection method for remote sensing images based on dual-branch attention and boundary enhancement, characterized by: The specific steps are as follows: Step 1: Data extraction and preprocessing: crop the original remote sensing image into non-overlapping sub-images of a predetermined size to ensure spatial alignment of the bi-temporal image and the label map; randomly divide the sub-images into training, validation, and test sets, and perform image normalization and data augmentation. Step 2: Feature extraction: Main branch encoding, first extracting multi-scale features of the bitemporal images in the training set through the Mamba encoder. The encoder consists of multiple layers, each of which gradually reduces the resolution of the feature map through downsampling operations. In each layer, the global spatial dependency of the feature map is modeled using a state-space model. Boundary enhancement branch encoding, for the high-dimensional features output by the main branch, gradually restores the spatial resolution of the feature map through channel compression and upsampling operations, and generates a single-channel boundary probability map P main ; Step 3: Decoding and spatiotemporal relationship modeling; The spatiotemporal relationship of multi-temporal features output by the encoder stage of the spatiotemporal state space (STSS) module is modeled, the feature maps of different levels are integrated through the fusion module, and the channel attention mechanism is embedded to enhance the change-sensitive features; Step 4: Edge optimization and probability map generation; the main branch uses the Sobel operator to capture the initial edge response E base , combined with deformable convolution to optimize edge offset and generate the final edge map E final ; Step 5: The final edge graph E generated by the main branch final The boundary probability map P generated by the boundary enhancement branch main The final change map is generated through weighted fusion. The error between the predicted result and the actual label is fed back into the network to adjust the network parameters and obtain a trained change detection model. Step 6: Input the validation set into the trained change detection model to preliminarily evaluate the model performance and make parameter adjustments. If the validation set indicators meet the standards, proceed to step 7; If the indicator does not meet the standard, return to step 1; Step 7: Input the images in the test set into the finalized change detection model to obtain the final prediction results for subsequent analysis and application of remote sensing images.

2. The method for binary change detection in remote sensing images based on dual-branch attention and boundary enhancement according to claim 1, characterized in that: In step 1, the original remote sensing image is cropped into non-overlapping sub-images of size 256×256. The sub-image set obtained after cropping is divided into the following types according to the conventional deep learning task ratio: the ratio of training set, validation set and test set is 6:2:

2.

3. The method for binary change detection in remote sensing images based on dual-branch attention and boundary enhancement according to claim 1, characterized in that: In step 2, the Mamba encoder is composed of two groups of weight-sharing encoders, each group of encoding parts includes four encoding stages; each Mamba encoder is composed of a cascade of four levels of Mamba blocks, which gradually extract multi-scale features; a convolutional layer with a stride of 2 is used before each level of Mamba block for 2x downsampling, and the level output size is one-half, one-quarter, one-eighth, and one-sixteenth of the height and width of the original image, respectively; within the Mamba block, the input feature map is spatially serialized and expanded, and global spatial dependencies are modeled using a state-space model; the features of the pre-event and post-event images extracted from the four encoding stages will be used for subsequent decoders responsible for specific tasks; the boundary enhancement branch coding uses a 1×1 convolutional layer to compress the number of channels of the feature map to 1 through a boundary prediction layer module.

4. The method for binary change detection in remote sensing images based on dual-branch attention and boundary enhancement according to claim 1, characterized in that: The step 3 includes four decoding stages. At the beginning of each decoding stage, the spatiotemporal state space STSS module models the spatiotemporal relationship of the multi-temporal features input in the encoder stage, and the spatiotemporal token generation module rearranges the input multi-temporal features and inputs them into three VSS modules; the output of the STSS module in the current stage is then integrated with the information of the feature map of the previous stage through the fusion module; the attention mechanism SE module is embedded in the fusion module, and the number of channels of the low-level feature map is mapped to be consistent with the high-level feature map through a 1×1 convolution layer; then, the high-level and low-level feature maps are summed; finally, the obtained feature map is smoothed by a residual layer; a new layer is added between the fusion module and the upsampling layer, and the smoothed feature map undergoes global average pooling, MLP compression and restoration, and channel weighting steps in the SE module to enhance the network's attention to important features.

5. The method for binary change detection in remote sensing images based on dual-branch attention and boundary enhancement according to claim 4, characterized in that: The smoothed feature map is globally average pooled in the SE module, and the formula is as follows: w c =GAP(F) Among them, F is the feature map, F∈R H×W×C ; Generate channel weights through two fully connected layers and Sigmoid function: w c =σ(W2(δ(W1z))) Where z∈R C Represents the feature vector obtained by global average pooling; represents the weight matrix of the first fully connected layer, r is the scaling factor; represents the weight matrix of the second fully connected layer; δ(·) represents the ReLU activation function; σ(·) represents the Sigmoid activation function, which is used to generate channel weights in the range of [0,1]; w c ∈R C Represents the weight of each channel; Multiply the weights by the original features channel by channel to obtain the enhanced feature map F out : F out =w c ⊙F。 6. The method for binary change detection in remote sensing images based on dual-branch attention and boundary enhancement according to claim 1, characterized in that: Step 4 uses the Sobel operator to capture the initial edge response: AND base =Sobel(P main ) Optimize edge offset using deformable convolution: Among them, E final represents the output boundary graph; E base represents the input feature map; K represents the convolution kernel size; p k represents the offset on the standard convolution kernel; Δp k represents the learnable offset; w k Represents the convolution kernel weight; p0 represents the convolution center point.

7. The method for binary change detection in remote sensing images based on dual-branch attention and boundary enhancement according to claim 1, characterized in that: The step 5 transforms the boundary graph E final With the main branch output P main Pixel-by-pixel weighting, the formula is: P final =P main ⊙(1+α·E final ) Where α∈[0.5,1.0].

8. The method for binary change detection in remote sensing images based on dual-branch attention and boundary enhancement according to claim 1, characterized in that: The final predicted change map obtained in step 6 is compared with the true label map in the test set, and the probability TP of actually being predicted as positive, the probability TN of actually being predicted as negative, the probability FP of actually being predicted as positive, and the probability FN of actually being predicted as negative are respectively calculated to calculate the actual performance indicators of our trained remote sensing image change detection network.

9. A computer device / apparatus / system comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images