Remote sensing image change detection method of difference enhancement network based on edge guidance

Abstract

The invention relates to a remote sensing image change detection method of a difference enhancement network based on edge guidance, and the constructed difference enhancement change detection network mainly comprises a backbone network, an edge enhancement module, a double-branch difference enhancement module, a multi-scale difference feature fusion module and a decoder. Edge information of the changed object is refined through an edge enhancement module, so that feature information with clear boundaries can be extracted for the changed object; a dual-branch difference enhancement module effectively learns the difference between the foreground and the background. The module captures global and local information of a dual-time image through space and channel branches, emphasizes a change area, inhibits noise interference, and enhances long-range dependence of difference. According to the method, the detection performance of the small object under the high-resolution remote sensing image is improved by using a multi-scale difference feature fusion method. Experimental results show that the detection effects of the proposed network on a plurality of public data sets are superior to those of several common remote sensing image change detection methods at present.

Description

Technical Field

[0001] The invention relates to the field of remote sensing images, and in particular to a remote sensing image change detection method based on an edge-guided difference enhancement network. Background Art

[0002] Over the past few decades, various change detection methods have been developed to handle a range of practical change detection tasks. Traditional change detection uses the pixel-level numerical differences between bi-temporal images to identify changes. Generally speaking, they consist of two main steps: (1) generating a difference image, which represents the degree of change through numerical calculation results. (2) analyzing the difference values ​​through threshold segmentation or clustering algorithms to obtain the final change map. Although traditional methods have achieved good results, these methods only extract the spectral information of the image and ignore the contextual relationship, so they are only applicable to medium and low resolution images. As optical sensors become more and more accurate, the complexity of objects in the scene in high-resolution optical remote sensing images, such as clear geometric structures, multi-scale, random distribution and uneven spectral characteristics, puts higher requirements on the accuracy of CD.

[0003] With the development of satellite technology, researchers can obtain more high-quality and high-resolution optical remote sensing images, which also provides the required data for the training of deep learning models. Deep learning-based change detection methods can use neural networks to extract deep features containing rich spectral and spatial details from remote sensing images, and then use these features to build semantic feature descriptions and detect change areas of interest. Deep learning-based change detection methods are mainly divided into "early fusion" and "late fusion". The methods that implement early fusion integrate bi-temporal information at the image level (i.e., network input), that is, stacking bi-temporal images along the channel dimension, and then encoding-decoding to generate change maps. In contrast, the methods that implement late fusion integrate bi-temporal information at the feature level. Generally speaking, the twin backbone network is used to extract bi-temporal features separately, and then the extracted features are processed and fused for downstream decision making. Image-to-image methods based on deep learning networks have greatly overcome the limitations of patch-to-pixel methods. They take the entire input image as the basic unit of analysis and introduce rich global, contextual, and semantic information into the method. Compared with patch-to-pixel methods, image-to-image methods can directly generate object-level rather than pixel-level results based on global and local semantic patterns.

[0004] Recent advances in deep learning have been extended to the field of change detection in remote sensing images. Convolutional neural networks (CNNs), such as the ResNet series and the UNet series, are often used as backbone networks to help extract discriminative hierarchical features from bi-temporal remote sensing images. Recurrent neural networks (RNNs) are good at processing time series data and can exploit the temporal dependencies between bi-temporal images in change detection. However, due to the lack of temporal information in bi-temporal high-resolution remote sensing images, these methods are limited in the temporal information they can exploit. Fully convolutional neural network (FCN) models attempt to capture long-range dependencies in context by deepening the network structure and using methods such as dilated convolutions. Various forms of attention mechanisms have also been introduced into the change detection task to reduce the interference of redundant features and enhance global information. However, due to the local nature of the convolution operation, methods that simply use channel attention or spatial attention are still insufficient in capturing long-range context associations in the spatiotemporal dimension. This deficiency may lead to uncertainty in edge pixels and missed detection of tiny objects. Therefore, the emergence of self-attention mechanisms provides a new idea for capturing long-range dependencies and obtaining more representative feature representations.

[0005] Thanks to the development of visual transformer (Vit) in the field of vision, many researchers have applied Vit and its variants to the field of remote sensing image change detection. For example, Zhang et al. designed a pure transformer change detection network Swin-SUNet, built a codec based on the Swin transformer module, and achieved good detection results. ChangeFormer extracts coarse-grained and fine-grained features from dual-temporal images by building a hierarchical transformer encoder for the dual network, and then uses a lightweight MLP decoder to fuse the multi-scale difference features calculated by the difference module. The hybrid model combining CNN and Transformer also has great potential in solving CD tasks. Bit-CD proposed by Chen et al. uses transformer as the encoder for spatiotemporal context modeling, which effectively enhances the features extracted by the convolutional network.

[0006] Although many deep learning models have been proposed and used for remote sensing image change detection tasks, they still have some problems. First, the change detection methods based on early fusion and late fusion either ignore the original information of the dual-temporal images or do not fully utilize the difference information between the dual-temporal images, so that the relationship between the dual-temporal images cannot be effectively constructed, resulting in irrelevant changes having an adverse effect on the detection results. Secondly, the boundary integrity and internal compactness of the changed objects are not fully considered, resulting in the loss of edge information in the predicted change map. Finally, the existing dual-branch network expands the model size, increases the computational cost, and is prone to overfitting problems. Summary of the invention

[0007] In view of the shortcomings of the prior art, the present invention proposes a remote sensing image change detection method based on edge-guided difference enhancement network.

[0008] Compared with the prior art, the present invention has the following beneficial effects:

[0009] 1. The present invention designs an edge enhancement module to refine the edge information of the changed object, so as to extract feature information with clear boundaries for the changed object.

[0010] 2. The present invention proposes a dual-branch difference enhancement module to effectively learn the difference between foreground and background. The module captures the global and local information of the dual-temporal image through spatial and channel branches, emphasizes the change area, suppresses noise interference, and enhances the long-range dependency of the difference.

[0011] 3. The present invention uses a multi-scale difference feature fusion method to improve the detection performance of small objects in high-resolution remote sensing images. Experimental results show that the detection effect of the proposed network on multiple public datasets is better than the current advanced methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a schematic diagram of the structure of the detection network proposed in the present invention;

[0013] Figure 2 is a schematic diagram of the structure of the edge enhancement module;

[0014] Figure 3 It is a schematic diagram of the structure of the dual-branch difference enhancement module;

[0015] Figure 4 It is a structural diagram of the multi-scale difference feature fusion module;

[0016] Figure 5 It is a visual comparison of feature fusion of different methods on the LEVIR-CD dataset;

[0017] Figure 6 It is a heat map visualization of feature maps at different stages of difference enhancement. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings. It should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, the description of well-known structures and technologies is omitted to avoid unnecessary confusion of the concept of the present invention.

[0019] The following is a detailed description with reference to the accompanying drawings.

[0020] Remote sensing image change detection (CD) is a challenging task that aims to detect semantic changes between pairs of satellite images at different temporal phases, which is of great significance for understanding surface changes. Although deep learning has achieved remarkable success in remote sensing change detection, current methods perform poorly in detecting subtle edge change regions and do not make full use of difference information, resulting in insufficient ability to distinguish pseudo-changes (such as illumination changes and perspective differences) and inability to fully capture the relationship and difference between bi-temporal image features. In order to accurately locate the areas where real changes have occurred and effectively learn the difference representation between foreground and background in the presence of limited samples and a large amount of interference, this paper proposes a multi-scale fusion change detection network based on edge guidance and difference enhancement, which aims to make full use of difference information to improve the accuracy of change detection in high-resolution remote sensing images.

[0021] Figure 1 The present invention proposes a structural schematic diagram of the detection network. The present invention proposes an edge-guided difference enhancement network for remote sensing image change detection method, constructs a difference enhancement change detection network, and the change detection network includes a backbone network, an edge enhancement module, a dual-branch difference enhancement module, a multi-scale difference feature fusion module and a decoder. The dual-time remote sensing high-resolution image is input into the change detection network, and the backbone network extracts features to obtain a multi-scale feature map. The edge information of the multi-scale feature map is enhanced by the edge enhancement module, and then the difference information is calculated by the dual-branch difference enhancement module to obtain a difference information feature map. Finally, the multi-scale difference feature fusion module assists the decoder to decode and obtain a binary change prediction map, which is used to display the surface or landscape changes that occur between different time points in the same area.

[0022] Step 1: Extract features from the high-resolution optical remote sensing image pair before and after the change, and generate an enhanced difference information feature map. Dual-temporal remote sensing images refer to: Figure 1 Before and after images of the changes.

[0023] Step 11: In the encoding stage, the dual-phase high-resolution remote sensing image is first cropped into non-overlapping 256×256 image blocks, and then flipped horizontally and vertically for data enhancement. The batch size is set to 16. First, a pair of weight-sharing backbone networks are used to extract the multi-scale feature map sets of the high-resolution optical remote sensing image before and after the change. The multi-scale feature map sets generated by the two backbone networks are: {F1, F2, F3, F4, F5}, {F1′, F2′, F′, F′, F5′}. The feature maps with the same subscript have the same size, and their sizes are (16, 16, 128, 128), (16, 24, 64, 64), (16, 32, 32, 32), (16, 96, 16, 16), (16, 320, 8, 8), respectively. The values ​​in parentheses are batch size, number of channels, length, and width from left to right. Dual-phase remote sensing images refer to: Figure 1 The before and after images in . The type of change is learned based on the labels of the training data.

[0024] Step 12: Perform edge feature enhancement on the two multi-scale feature map sets extracted in step S11 to achieve the redistribution of spatial and semantic information; the edge enhancement module aims to improve the effect of subsequent tasks by enhancing the edge information of each feature map in the input multi-scale feature map set. Figure 2 It is a structural diagram of the edge enhancement module.

[0025] Step 121: The multi-scale feature map is passed to an anisotropic Sobel operator module to extract a high-frequency edge feature map.

[0026] Step 122: The multi-scale feature maps are subjected to average pooling and maximum pooling operations respectively to generate corresponding dimension reduction vectors.

[0027] Step 123: After summing the dimension-reduced vectors, the vectors are processed by a multi-layer perceptron and a ReLU activation function, and a set of weight coefficients are generated by a Sigmoid activation function.

[0028] Step 124: The multi-scale feature map is input into a depth-separable convolution process to obtain a semantic feature map, and the semantic feature map is multiplied by the weight coefficient to dynamically adjust the weight value of the semantic feature map.

[0029] Step 125: The adjusted semantic feature map is concatenated with the high-frequency edge feature map in the channel dimension, and the number of channels is reduced to 1 through a 1x1 convolution layer. After being processed by a standard convolution layer, a batch normalization layer, and a ReLU activation function, an edge-enhanced feature map is generated.

[0030] Each feature map in the two multi-scale feature map sets will pass through the edge enhancement module, and finally the edge enhanced feature map sets: {E1, E2, E3, E4, E5} and {E1′, E2′, E3′, E4′, E5′} are obtained, which are then used as the input of the dual-branch difference enhancement module.

[0031] Step 13: Calculate the difference features of the two sets of multi-scale feature maps after edge enhancement. By using the spatial attention mechanism and channel attention mechanism in the channel and space respectively, the difference features are recalibrated and enhanced in both spatial and channel dimensions, so that they can focus on the real change area.

[0032] like Figure 3 As shown, the features Figure 1 and Features Figure 2 Respectively represent {E i ,E i ′}, where i = {1, 2, 3, 4, 5}. Step 131: In the channel branch, firstly transform the feature Figure 1 and Features Figure 2 They are concatenated in the channel dimension and then processed by convolution to obtain the feature map z_c. The channel attention mechanism then processes the feature map z_c to generate the channel-enhanced feature map c_out.

[0033] Step 132: The spatial branch performs element-by-element subtraction and absolute value operation on the two input feature maps to obtain a feature map z_s. Then, the feature map z_s is processed using a spatial attention mechanism to generate a spatially enhanced feature map s_out.

[0034] Step 133: Multiply the feature map c_out and the feature map s_out to obtain a difference information feature map, thereby effectively capturing and enhancing the subtle differences between the input features.

[0035] After step S13, five difference information feature maps of different scales are obtained, namely, a first difference information feature map, a second difference information feature map, a third difference information feature map, a fourth difference information feature map, and a fifth difference information feature map, which are used in the decoding stage.

[0036] Step 2: Fuse the four multi-scale difference information feature maps, including: the second difference information feature map, the third difference information feature map, the fourth difference information feature map, and the fifth difference information feature map. Use the multi-scale difference feature fusion module to assist the encoding stage, and generate an auxiliary change map for deep supervision to speed up the network training.

[0037] The multi-scale difference feature fusion module aims to effectively fuse multi-scale features to improve the performance of the network when processing complex image tasks.

[0038] The multi-scale difference feature fusion module consists of a side chain layer and a feature fusion layer. The side chain layer includes a convolution layer, a batch normalization layer, a ReLU activation layer, and a compression extraction layer, which are used to process features of different scales. The feature fusion layer is used to fuse and output the features processed by all side chain layers to obtain a multi-scale feature fusion map. The structure of the multi-scale difference feature fusion module is as follows: Figure 4 shown.

[0039] Step 21: The multi-scale difference feature fusion module first uses the side chain layer to process the difference information feature maps of different scales. Each side chain layer includes a convolution layer, a batch normalization layer, a ReLU activation function and a compression extraction layer. The compression extraction layer compresses the feature map through an average pooling operation, and then uses a fully connected layer and an activation function to obtain a weight matrix. The weight matrix is ​​multiplied by the difference information feature to adaptively adjust the relationship between channels so that the network pays attention to the features that really change.

[0040] Step 22: The four processed multi-scale difference feature maps are upsampled to a uniform size and concatenated in the channel dimension. Finally, the concatenated feature maps are convolutionally processed to obtain multi-scale fused feature maps. This design enables the network to make full use of multi-scale difference information and generate difference features for auxiliary decoding and deep supervised learning.

[0041] Step 3: In the decoding stage, the fifth difference information feature map in the encoding stage is used as input, and the first difference information feature map, the second difference information feature map, the third difference information feature map, the fourth difference information feature map and the multi-scale fusion feature map are used to assist in decoding, making full use of the difference information to generate a more accurate change result. The decoding process is as follows: Figure 1 shown.

[0042] Step 31: Take the fifth difference information feature map as the input of the decoder, first upsample the fifth difference information feature map to the same resolution as the fourth difference information feature map, concatenate the fifth difference information feature map and the fourth difference information feature map in the channel dimension, and then perform an edge enhancement to obtain a first difference edge enhanced feature map, and then use the first difference edge enhanced feature map as the input of the next decoding stage, repeat three times, and obtain the final fourth difference edge enhanced feature map.

[0043] Step 32: The multi-scale fusion feature map obtained by multi-scale fusion is added element by element to the fourth difference edge enhancement feature map, and then a convolution operation with a convolution kernel size of 1 is performed to change the channel dimension of the feature map to 1, and then the sigmoid activation function is used to limit the value of each element between 0 and 1, so as to generate a binary change prediction map through the threshold.

[0044] Step 33: Perform a 1×1 convolution and a sigmoid activation function on the multi-scale fusion feature map to generate an auxiliary change prediction map for deep supervision.

[0045] Step 34: After the decoder stage, a binary change prediction map and an auxiliary change prediction map are generated. Then, the binary cross entropy loss is calculated for the two prediction results and the corresponding labels. The loss values ​​are weighted and summed according to the set weights, which are used as the basis for back propagation to update the model parameters. In each iteration, the model parameters are optimized through continuous forward propagation and back propagation, and the training will continue until the preset number of iterations is met. Specifically, the binary cross loss of the auxiliary change prediction map and the label image is the auxiliary loss.

[0046] In order to better evaluate the effectiveness of the method proposed in the present invention, the present invention also uses quantitative evaluation indicators for comparison, specifically: precision, recall, F1 score and intersection over union. Precision is the ratio of the number of pixels predicted correctly by the model to the total number of pixels. Recall is an indicator that measures the ability of a model to correctly identify all instances of a certain category. The F1 score is an indicator that measures the accuracy of a binary classification model. It is the harmonic average of precision and recall, and is used to consider these two indicators simultaneously. The intersection over union measures the degree of overlap between the prediction and the true annotation. Table 1 is a comparison of the quantitative evaluation results.

[0047] Table 1 Comparison of quantitative evaluation indicators of different methods

[0048] Precision Law Recall F1 score Intersection and Union Method of the present invention 0.9021 0.9204 0.9112 0.8369 Method 1 0.8520 0.8597 0.8558 0.7480 Method 2 0.8613 0.8977 0.8791 0.7843 Method 3 0.8483 0.8893 0.8684 0.7673 Method 4 0.9138 0.8708 0.8918 0.8044 Method 5 0.8850 0.9172 0.9008 0.8195 Method 6 0.8734 0.9135 0.8930 0.8067 Method 7 0.8615 0.8813 0.8713 0.7719 Method 8 0.8896 0.9264 0.9077 0.8309 Method 9 0.9047 0.9160 0.9103 0.8369

[0049] Method 1 is the FC-EF method, method 2 is the FC-Siam-conc method, method 3 is the FC-Siam-diff method, method 4 is the MSPSNET method, method 5 is the BIT method, method 6 is the SNUNet method, method 7 is the DTCDSCN method, method 8 is the ICIFNet method, and method 9 is the ELGCNet method.

[0050] Figure 5 It is a visual comparison of feature fusion of different methods on the LEVIR-CD dataset. Specifically, Figure 5 (a) is the image at time T0, Figure 5 (b) is the image at time T1. Figure 5 (c) to Figure 5 (h) is an existing remote sensing image change detection method, specifically, Figure 5 (c) is the MSPSNet method, Figure 5 (d) is the BIT method, Figure 5 (e) is the SNUNet method, Figure 5 (f) is the DTCDSCN method, Figure 5 (g) is the ICIFNet method, Figure 5 (h) is the ELGCNet method. Figure 5 (i) is the method proposed by the present invention. Figure 5 (j) is the label image. Small building samples were selected from the LEVIR-CD dataset for visual comparison. In the presence of strong illumination changes, the proposed method shows significantly lower false detection and missed detection, and the boundaries between buildings in the predicted image are clearer.

[0051] Figure 6 It is a heat map visualization comparison of feature maps at different stages of difference enhancement. Specifically, Figure 6 (a) is the image at time T0, Figure 6 (b) is the image at time T1, Figure 6 (c) is the label image, Figure 6 (d) is the first difference information feature map, Figure 6 (e) is the fifth difference information feature map, Figure 6 (f) is the output feature of the last layer of the decoder. Through feature visualization, we can see that the proposed model has been learning the representation of changes at different stages. As the network depth increases, the feature map after difference enhancement shows a strong ability to identify the changed and unchanged areas in the input feature map, and can accurately locate the changed area and better preserve the integrity and spatial details of the changed area.

[0052] It should be noted that the above specific embodiments are exemplary, and those skilled in the art can come up with various solutions inspired by the disclosure of the present invention, and these solutions also belong to the disclosure scope of the present invention and fall within the protection scope of the present invention. Those skilled in the art should understand that the present invention description and its drawings are illustrative and do not constitute a limitation of the claims. The protection scope of the present invention is defined by the claims and their equivalents.

Claims

1. A remote sensing image change detection method based on edge-guided difference enhancement network, characterized in that: A difference enhancement change detection network is constructed, the change detection network includes a backbone network, an edge enhancement module, a dual-branch difference enhancement module, a multi-scale difference feature fusion module and a decoder, a dual-temporal remote sensing high-resolution image is input into the change detection network, features are extracted through the backbone network to obtain a multi-scale feature map, the edge information of the multi-scale feature map is enhanced through the edge enhancement module, and then the difference information is calculated through the dual-branch difference enhancement module to obtain a difference information feature map, and finally the multi-scale difference feature fusion module is used to assist the decoder to decode and obtain a binary change prediction map. The specific method includes: Step 1: Extract features from the high-resolution optical remote sensing image pairs before and after the change, and generate an enhanced difference information feature map; Step 11: In the encoding stage, the dual-phase high-resolution remote sensing image is cropped into non-overlapping 256×256 image blocks, and then horizontally and vertically flipped for data enhancement. Then, a pair of weight-sharing backbone networks are used to extract multi-scale feature maps of the image before and after the change, respectively. Specifically, the set of multi-scale feature maps of the image before the change is {F1, F2, F3, F4, F5}, and the set of multi-scale feature maps of the image after the change is {F1′, F2′, F′, F′, F5′}; Step 12: Input the two multi-scale feature map sets extracted in step 11 into the edge enhancement module, and perform edge feature enhancement respectively to achieve the redistribution of spatial and semantic information. Each feature map in the two multi-scale feature map sets will pass through the edge enhancement module, and finally obtain the edge enhanced feature map sets of the image before and after the change: {E1, E2, E3, E4, E5} and {E1′, E2′, E3′, E4′, E5′}, which serve as the input of the dual-branch difference enhancement module; Step 13: Calculate the difference features of the two edge enhancement feature map sets, and recalibrate and enhance the difference features in the two dimensions of space and channel by using the spatial attention mechanism and the channel attention mechanism in the channel and space respectively, and finally obtain 5 difference information feature maps of different scales, specifically: the first difference information feature map, the second difference information feature map, the third difference information feature map, the fourth difference information feature map, and the fifth difference information feature map; Step 2: The four multi-scale difference information feature maps are fused through a multi-scale difference feature fusion module, specifically including: the second difference information feature map, the third difference information feature map, the fourth difference information feature map, and the fifth difference information feature map. The multi-scale difference feature fusion module is composed of a side chain layer and a feature fusion layer. The side chain layer includes a convolution layer, a batch normalization layer, a ReLU activation layer, and a compression extraction layer. Step 21: First, the difference information feature maps of different scales are processed using the side chain layer. Each side chain layer includes a convolution layer, a batch normalization layer, a ReLU activation function, and a compression extraction layer. The compression extraction layer compresses the difference information feature map through an average pooling operation, and then a fully connected layer and an activation function are used to obtain a weight matrix. The weight matrix and the difference information feature map are multiplied to obtain four processed difference information feature maps. Step 22: The four processed difference information feature maps are input into the feature fusion layer, upsampled to a uniform size, and concatenated together in the channel dimension; Step 23: The concatenated feature map is subjected to a convolution operation to obtain a multi-scale fused feature map; Step 3: In the decoding stage, the fifth difference information feature map is used as input, and the first difference information feature map, the second difference information feature map, the third difference information feature map, the fourth difference information feature map and the multi-scale fusion feature map are used to assist in decoding; Step 31: take the fifth difference information feature map as the input of the decoder, first upsample the fifth difference information feature map to the same resolution as the fourth difference information feature map, concatenate the fifth difference information feature map and the fourth difference information feature map in the channel dimension, and then perform an edge enhancement to obtain a first difference edge enhancement feature map, and then use the first difference edge enhancement feature map as the input of the next decoding stage, repeat three times, and obtain the final fourth difference edge enhancement feature map; Step 32: Add the multi-scale fusion feature map and the fourth difference edge enhancement feature map element by element, and then perform a convolution operation with a convolution kernel size of 1 to change the feature map channel dimension to 1, and then use the sigmoid activation function to limit the value of each element between 0 and 1, so as to generate a binary change prediction map through a threshold; Step 33: Perform a 1×1 convolution and a sigmoid activation function on the multi-scale fusion feature map to generate an auxiliary change prediction map; Step 34: After the decoder stage, a binary change prediction map and an auxiliary change prediction map are generated. Then, the binary cross entropy loss is calculated for these two prediction maps and the corresponding label images, and the model parameters are updated by back propagation until the training ends after the preset number of iterations is met.

2. The remote sensing image change detection method according to claim 1, characterized in that: The processing process of obtaining the edge enhancement feature map from the multi-scale feature map in step 2 includes: Step 121: The multi-scale feature map is passed to an anisotropic Sobel operator module to extract a high-frequency edge feature map; Step 122: The multi-scale feature maps are subjected to average pooling and maximum pooling operations respectively to generate corresponding dimension reduction vectors; Step 123: After summing the dimension-reduced vectors, the vectors are processed by a multi-layer perceptron and a ReLU activation function, and a set of weight coefficients are generated by a Sigmoid activation function; Step 124: input the multi-scale feature map into a depth-separable convolution process to obtain a semantic feature map, and multiply the multi-scale feature map by the weight coefficient to dynamically adjust the weight value of the semantic feature map; Step 125: The adjusted semantic feature map is concatenated with the high-frequency edge feature map in the channel dimension, the number of channels is reduced to 1 through a 1x1 convolution layer, and processed through a standard convolution layer, a batch normalization layer, and a ReLU activation function to generate an edge enhanced feature map.

3. The remote sensing image change detection method according to claim 2, characterized in that: The calculation of the difference information feature map in step 13 specifically includes: Step 131: Feature map 1 and feature map 2 represent {E i ,E i ′}, where i = {1, 2, 3, 4, 5}. In the channel branch, feature map 1 and feature map 2 are first concatenated in the channel dimension, and then convolution is performed to obtain feature map z_c. The channel attention mechanism then processes feature map z_c to generate a channel-enhanced feature map c_out. Step 132: In the spatial branch, perform element-by-element subtraction and absolute value operation on feature map 1 and feature map 2 to obtain feature map z_s, and then process feature map z_s using the spatial attention mechanism to generate a spatially enhanced feature map s_out; Step 133: Multiply the feature map c_out and the feature map s_out to obtain a difference information feature map.

Images