A remote sensing image change detection method combining feature fusion and reinforcement difference

Abstract

The application discloses a remote sensing image change detection method combining feature fusion and reinforced difference, which comprises the following steps: inputting two-phase remote sensing images into a model after preprocessing; extracting features of the two-phase remote sensing images through a down-sampling channel and calculating a reinforced difference feature map; cascading the highest layer features of the two channels and the difference features as up-sampling input; establishing an auxiliary up-sampling channel to fuse the features into a main up-sampling channel in the up-sampling process, and cascading the reinforced difference features and shallow features of each scale in the down-sampling to corresponding layers of the up-sampling; obtaining an image change probability map through deconvolution and Sigmoid operation on the output of the last layer of the up-sampling; calculating a loss according to a real image change label by using a hybrid loss function based on binary cross entropy and Dice; and updating the network to finally generate a change detection result map through a back propagation algorithm. The application can effectively extract more abundant and detailed features of the two-phase remote sensing images, and can obtain a change detection result with higher precision than other networks.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of remote sensing, and particularly relates to a remote sensing image change detection technology combining feature fusion and strengthened difference, and particularly relates to a high-resolution remote sensing image change detection method combining deep separable convolution and strengthened change region BACKGROUND

[0002] Building change detection technology is an important research direction in the field of remote sensing. With the continuous development of deep learning technology, building change detection technologies based on deep learning methods have emerged. The U-shaped deep network derived from the Unet network is a commonly used network for building change detection.

[0003] Currently, change detection technologies based on the U-shaped deep network mainly focus on the research of change part extraction algorithms and the modification of network structures, and show not low change detection precision. However, the features extracted by various feature extraction methods in many methods are always more or less missing. The paper "Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation [J]. Springer International Publishing, 2015." first designs a U-shaped feature extraction network, and the strictly symmetrical structure of which can conveniently cascade deep and shallow layer information, but its structure is relatively simple, and the down-sampling only has 5 convolution operations. And in the up-sampling process, only the shallow layer information in the down-sampling process is cascaded, and the fused features are limited. The paper "Z. Zhou, M. Siddiquee, N. Tajbakhsh and J. Liang: UNet++: A Nested U-Net Architecture for Medical Image Segmentation" improves the basic network U-Net, and performs step-by-step fusion in the down-sampling stage. However, this operation greatly increases the experimental parameters, and the effective features extracted are relatively less.

[0004] Based on the above, the application provides a remote sensing image change detection method combining feature fusion and strengthened difference. SUMMARY

[0005] In view of the technical problems existing in the research on building change detection in the background art, the present invention provides a remote sensing image change detection method that combines feature fusion and enhanced difference, which can solve the problem of weak feature extraction capability in the literature listed in the background art. The present invention can extract richer features from images from two periods, supplement the feature extraction and restoration process, and obtain better change detection results.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] A method for detecting changes in remote sensing images that combines feature fusion and enhanced differences includes the following steps:

[0008] Step 1: First, preprocess the two remote sensing images, then input the two images into two downsampling channels with the same parameters to extract image features; and during the downsampling process, obtain the different scale enhancement difference features of the two channels.

[0009] Step 2: Concatenate the features obtained from downsampling of the two channels and the enhancement difference features obtained from the last downsampling. Use the concatenated image features as the input of the two upsampling channels. During the upsampling process, concatenate the shallow features at each scale, the enhancement difference features at each scale, and the features of the auxiliary upsampling channel to the same scale feature layer through short connections.

[0010] Step 3: After the upsampling operation is completed, the output of the last upsampling layer is deconvolved and Sigmoided to obtain the image change probability map. Then, a hybrid loss function based on binary cross-entropy and Dice is used to calculate the loss based on the real image change labels. The network parameters are continuously updated through the backpropagation algorithm. After multiple iterations, the model with the optimal relative parameters is obtained and the final change detection result map is generated.

[0011] In step 1, features are extracted using two downsampling channels, taking into account the differences between the two downsampling channels, as detailed below:

[0012] 1) Let T1 be the five downsampling convolutional modules in channel 1. i , i = 1, 2, 3, 4, 5, the i-th convolutional module T1 i The input and output are denoted as T1. i (In) and T1 i (Out); The input of the first convolutional module T11 is X1, that is, T11(In) = X1, and the input of subsequent convolutional modules T1... i Input T1 i (In) is MAX(T1) i-1(Out)), MAX() represents the max pooling operation;

[0013] 2) Let T2 be the five downsampling convolutional modules in channel 2. i , i = 1, 2, 3, 4, 5, the i-th convolutional module T2 i The input and output are denoted as T2. i (In) and T2 i (Out); The input of the first convolutional module T21 is X2, that is, T21(In) = X2, and the input of subsequent convolutional modules T2 i Input T2 i (In) is MAX(T2) i-1 (Out)), MAX() represents the max pooling operation;

[0014] 3) While downsampling the two channels, generate enhanced differential features (STCA) from the features at each scale of the two channels. i (Strengthen The Change Srea, STCA) is determined by the following formula:

[0015]

[0016]

[0017] STCA i =P i1 +P i2 (3)

[0018] Where i = 1, 2, 3, 4, 5, T1 i (Out) is the output of the i-th downsampling convolutional module in channel 1, T2 i (Out) represents the output of the i-th downsampling convolutional module in channel 2, P i1 P represents the output feature map of the i-th convolutional module in channel 1. i2 STCA represents the feature map output by the i-th convolutional module in channel 2. i This represents the i-th enhanced differential feature map generated by the two convolutional channels.

[0019] In step 1, specifically, the two-stage image input network → T11 and T21 convolutional layers → the first STCA module → the first layer convolutional output MAX(T11(Out)) and MAX(T21(Out)) → T12 and T22 convolutional layers → the second STCA module → the second layer convolutional output MAX(T12(Out)) and MAX(T22(Out)) → T13 and T23 convolutional layers → the third STCA module → the third layer convolutional output MAX(T13(Out)) and MAX(T23(Out)) → T14 and T24 convolutional layers → the fourth STCA module → the fourth layer convolutional output MAX(T14(Out)) and MAX(T24(Out)) → T15 and T25 convolutional layers → the fifth STCA module → the fifth layer convolutional output MAX(T15(Out)) and MAX(T25(Out)).

[0020] In step 2, the highest-level features T15(Out) and T25(Out) obtained by downsampling channel 1 and channel 2, along with the enhancement difference feature STCA5 of the highest-level features of the two channels, are concatenated to obtain the concatenated feature FUSE, i.e. Using FUSE as the input to both the primary and secondary upsampling channels for upsampling operations involves the following sub-steps:

[0021] Step 2.1: In the auxiliary upsampling channel, denote the four deconvolution modules as UA. i i = 1, 2, 3, 4, the i-th deconvolution module UA i The input and output are denoted as UA. i (In) and UA i (Out) The image size is gradually restored to the input size by passing through four deconvolution modules that are different from the main channel. The output feature map size of each layer is the same as the scale of the corresponding layer in the main channel, so that it can be cascaded into the main channel to supplement the features lost by the main channel deconvolution.

[0022] Step 2.2: In the main upsampling channel, the image is also sequentially passed through 4 deconvolution modules to restore the image scale to the input size; these 4 deconvolution modules are denoted as UM. i , i = 1, 2, 3, 4, the i-th deconvolution module UM i The input and output are denoted as UM. i (In) and UM i (Out).

[0023] In step 2.1, each deconvolution module UA i (i = 1, 2, 3, 4) The following operation is performed:

[0024] Step 2.1.1: Input UAi (In) Perform a deconvolution operation, denoted as F. Then, the deconvolution operation of the auxiliary upsampling channel is denoted as F1. The output after the deconvolution operation is F1(UA). i (In));

[0025] Step 2.1.2: Convert the output F1(UA) after the deconvolution operation i (In)) Perform depthwise separable convolution operations; depthwise separable convolution replaces the traditional convolution operation with two parts: channel-wise convolution and point-wise convolution. Compared with traditional convolution, it can greatly reduce the number of parameters while maintaining a small loss of accuracy.

[0026] Step 2.1.3: Receive the output from the previous step and perform the same operation; after the above three steps, obtain the UA of each layer of the auxiliary upsampling channel deconvolution module. i Output UA i (Out).

[0027] In step 2.2, each deconvolution module UM i (i = 1, 2, 3, 4) The following operation is performed:

[0028] Step 2.2.1: For the input UM i (In) Perform a deconvolution operation, denoted as F. Then, the deconvolution operation of the main upsampling channel is denoted as F2. The output after the deconvolution operation is F2(UM). i (In));

[0029] Step 2.2.2: Convert the output T1 of the 5th–ith downsampling convolutional module of downsampling channel 1 and channel 2. 5-i (Out), T2 5-i (Out), Enhanced differential features of downsampling channel 1 and channel 2 (STCA) 5-i The deconvolution output F2(UM) in step 2.2.1 of the main channel i (In)), and the output UA of the corresponding scale auxiliary upsampling channel. i (Out) performs a cascading operation to obtain the cascading feature FUSE. i ,Right now When i=1, what is received is the deconvolution after the last downsampled fusion layer and the fused features of the corresponding layer.

[0030] Step 2.2.3: Apply the above cascaded feature FUSE iThe process is sequentially handled by two convolutional units, each consisting of a convolutional layer, a batch normalization layer, and a ReLU activation layer. The convolutional units used here are the same as those used in downsampling channel 1 and channel 2. The main upsampling channel is processed through steps 2.2.1 to 2.2.3 to obtain the deconvolution module UM. i Output UM i (Out).

[0031] The specific process of step 2 is as follows: FUSE → F(FUSE) → F1(UA1(In)) and F2(UM1(In)) deconvolution layers → UA1(Out), MAX(T14(Out)), MAX(T24(Out)), F2(UM1(In)), STCA4 are fused into the UM1 deconvolution main channel → UM1(Out) → F1(UA2(In)) and F2(UM2(In)) deconvolution layers → UA2(Out), MAX(T13(Out)), MAX(T23(Out)), F2(UM2(In)), STCA3 are fused into the UM2 deconvolution main channel In the middle → UM2(Out) → F1(UA3(In)) and F2(UM3(In)) deconvolution layer → UA3(Out), MAX(T12(Out)), MAX(T22(Out)), F2(UM3(In)), STCA2 are merged into the main channel of UM3 deconvolution → UM3(Out) → F1(UA4(In)) and F2(UM4(In)) deconvolution layer → UA4(Out), MAX(T11(Out)), MAX(T21(Out)), F2(UM4(In)), STCA1 are merged into the main channel of UM4 deconvolution → UM4(Out).

[0032] In step 3, after the upsampling operation is completed, the upsampling output feature map UM4(Out) is obtained. After one step of deconvolution and activation by the Sigmoid function, the change detection feature map is obtained. The loss is calculated based on the real image change label using a hybrid loss function based on binary cross-entropy and Dice. The network parameters are continuously updated through the backpropagation algorithm. After multiple iterations, the model with the optimal relative parameters is obtained and the final change detection result map is generated.

[0033] This has the following technical effects:

[0034] The technical solution proposed in this invention uses an auxiliary upsampling channel, which differs from traditional convolutional methods, to fuse features into the main upsampling channel, thereby acquiring richer feature information. Simultaneously, the proposed solution also considers the enhanced difference features at various scales between the two channels during downsampling, resulting in even richer extracted feature information. Through these measures, this invention achieves more accurate change detection results. Attached Figure Description

[0035] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0036] Figure 1 This is a network structure diagram of an embodiment of the present invention;

[0037] Figure 2 This is the first-period image X1 of the experimental data example image used in this embodiment of the invention;

[0038] Figure 3 This is the second-period image X2 of the experimental data example image used in this embodiment of the invention;

[0039] Figure 4 This is a diagram showing the actual changes between two phases of experimental data example images used in an embodiment of the present invention;

[0040] Figure 5 This is a change detection map of the example image obtained using experimental data, obtained by the control method (Unet++) of this invention;

[0041] Figure 6 This is a change detection map of the example image obtained using experimental data from the control method (Unet) of this invention;

[0042] Figure 7 This is a change detection map obtained by the present invention using example images of experimental data. Detailed Implementation

[0043] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0044] This embodiment uses the WHU Building Area Dataset, a commonly used building change detection dataset in the field of remote sensing image change detection, to conduct experiments and verify the effectiveness of the method proposed in this invention. This dataset contains 1213 building remote sensing images of 256×256 pixels each. Each image includes remote sensing images from two periods and a true change map of the images from the two periods. The training set consists of 600 images, and the test set consists of 613 images.

[0045] like Figure 1 The diagram shown illustrates the main framework of the method proposed in this invention. This invention provides a remote sensing image change detection method that combines feature fusion and enhanced difference analysis, comprising the following steps:

[0046] Step 1: First, preprocess the two sets of remote sensing images. Then, input the first set of remote sensing images into the first channel and the second set into the second channel. Both channels use the same downsampling parameters to extract remote sensing image features. During the downsampling process, obtain the different scale enhancement features between the two channels. Specifically:

[0047] 1) Extract the features of the first remote sensing image X1 using channel 1, and then perform downsampling operations on X1 sequentially through five convolutional modules: denote the five downsampling convolutional modules as T1. i , i = 1, 2, 3, 4, 5, the i-th convolutional module T1 i The input and output are denoted as T1. i (In) and T1 i (Out). It should be noted that the convolution method and stride can be adjusted according to the specific application. The input of the first convolutional module T11 in channel 1 is X1, i.e., T11(In) = X1, and subsequent convolutional modules T1... i Input T1 i (In) is MAX(T1) i-1 (Out)), where i = 2, 3, 4, 5, and MAX() represents the max pooling operation.

[0048] 2) Extract features from the second remote sensing image X2 using channel 2, and then perform downsampling operations on X2 sequentially through five convolutional modules: denoted as T2 for these five downsampling convolutional modules. i , i = 1, 2, 3, 4, 5, the i-th convolutional module T2 i The input and output are denoted as T2. i (In) and T2 i (Out). It should be noted that the convolution method and stride can be adjusted according to the specific application. The input to the first convolutional module T21 in channel 2 is X2, i.e., T21(In) = X2, and subsequent convolutional modules T2... i Input T2 i (In) is MAX(T2) i-1 (Out)), where i = 2, 3, 4, 5, and MAX() represents the max pooling operation.

[0049] Each convolutional module in Channel 1 and Channel 2 includes two identical convolutional units. Each convolutional unit comprises a convolutional layer, a batch normalization layer, and a ReLU activation layer. Table 1 shows the specific parameters used by the convolutional units in each convolutional module of Channel 1 and Channel 2 in this embodiment of the invention, including kernel size, number of kernels, stride, and activation function. It should be noted that other parameters such as convolution parameters, activation methods, and pooling methods in Channel 1 and Channel 2 can be adjusted according to specific applications.

[0050] Table 1 shows the specific parameters of the convolutional units used in each convolutional module of the two channels in this embodiment of the invention.

[0051]

[0052] 3) Generating Enhanced Differential Features: The differential features between two remote sensing images play an important role in change detection. To make the extracted features richer and more detailed, we consider using features from each scale of the two channels to generate enhanced differential features (STCA). i And determined by the following formula:

[0053]

[0054]

[0055] STCA i =P i1 +P i2 (3)

[0056] Where i = 1, 2, 3, 4, 5, T1 i (Out) is the output of the i-th downsampling convolutional module in channel 1, T2 i (Out) is the output of the i-th downsampling convolutional module in channel 2. P i1 P represents the output feature map of the i-th convolutional module in channel 1. i2 STCA represents the feature map output by the i-th convolutional module in channel 2. i Let represent the i-th enhanced differential feature map generated by the two convolutional channels. In the above equation, (1) and (2) respectively represent the process of superimposing the pre-classification information onto the original feature map to enhance the differential parts. Concatenating these enhanced differential feature maps during upsampling can better highlight the changing parts and improve the accuracy of change detection.

[0057] Step 2: Concatenate the highest-level features obtained from downsampling of the two channels, along with the enhanced difference features of the highest-level features of the two channels. Use the concatenated features as input for upsampling to begin the upsampling operation. During upsampling, concatenate the shallow features at each scale in the two downsampling channels, along with the enhanced difference features at each scale in the two channels, to the upsampling feature layer of the same scale through short connections. Specifically:

[0058] First, the highest-level features T15(Out) and T25(Out) obtained by downsampling from channels 1 and 2, along with the enhancement difference feature STCA5 of the highest-level features of the two channels, are concatenated to obtain the concatenated feature FUSE, i.e. Secondly, FUSE is used as the input to both the main upsampling channel and the auxiliary upsampling channel. In the auxiliary upsampling channel, the image is gradually restored to the input size by passing through four deconvolution modules different from those in the main channel. The output feature map size of each layer is the same as the corresponding layer size in the main channel, so that it can be cascaded into the main channel to supplement the features lost during the main channel's deconvolution. These four deconvolution modules are denoted as UA. ii = 1, 2, 3, 4, the i-th deconvolution module UA i The input and output are denoted as UA. i (In) and UA i (Out). In the main upsampling channel, the image is also sequentially processed through four deconvolution modules to restore the image scale to the input size; these four deconvolution modules are denoted as UM. i , i = 1, 2, 3, 4, the i-th deconvolution module UM i The input and output are denoted as UM. i (In) and UM i (Out). The input to the first deconvolution module is F(FUSE), i.e., UA1(In) = F(FUSE), UM1(In) = F(FUSE), where F represents the deconvolution operation.

[0059] Step 2.1: In the auxiliary upsampling channel, each deconvolution module UA i (i = 1, 2, 3, 4) includes the following steps:

[0060] Step 2.1.1: Input UA i (In) Perform a deconvolution operation, denoted as F, and the auxiliary upsampling channel as F1. The output after the deconvolution operation is F1(UA). i (In)). For specific details on deconvolution, please refer to the literature "Matthew D. Zeiler, and Dilip Krishnan and Graham W. Taylor and Rob Fergus Dept, "DECONVOLUTIONAL NETWORKS" 2010 IEEE Computer Society Conference on ComputerVision and Pattern Recognition."

[0061] Table 2 provides the specific parameters of each deconvolution module in the auxiliary upsampling channel of this invention, including kernel size, number of kernels, stride, and activation function. It should be noted that the number of convolution modules used, the number of convolutional units used in each module, and the parameters and activation methods in each convolutional unit can all be adjusted according to specific applications.

[0062] Table 2. Specific parameters of each deconvolution module in the auxiliary upsampling channel in this embodiment of the invention.

[0063]

[0064] Step 2.1.2: Convert the output F1(UA) after the deconvolution operation i(In)) Perform depthwise separable convolution operations; depthwise separable convolution replaces the traditional convolution operation with two parts: channel-wise convolution and pointwise convolution. Compared with traditional convolution, it can greatly reduce the number of parameters while maintaining a small loss of accuracy. For details on depthwise separable convolution operations, please refer to the paper "CHOLLET F.XCEPTION: DEEP LEARNING WITH DEPTHWISE SEPARABLE CONVOLUTIONS[J].IEEE,2017";

[0065] Step 2.1.3: Receive the output from the previous step and perform the same operation; after the above steps, the auxiliary upsampling channel deconvolution module UA is obtained. i Output UA i (Out).

[0066] Step 2.2: In the upsampling main channel, each deconvolution module UM i (i = 1, 2, 3, 4) contains the following three sub-steps:

[0067] Step 2.2.1: For the input UM i (In) Perform a deconvolution operation. If the deconvolution operation is denoted as F, then the deconvolution operation of the main upsampling channel is denoted as F2. The output after the deconvolution operation is F2(UM). i (In));

[0068] Step 2.2.2: Convert the output T1 of the 5th–ith downsampling convolutional module of downsampling channel 1 and channel 2. 5-i (Out), T2 5-i (Out), Enhanced differential features of downsampling channel 1 and channel 2 (STCA) 5-i The deconvolution output F2(UM) of the main channel in step one i (In)), and the output UA of the corresponding scale auxiliary upsampling channel. i (Out) performs a cascading operation to obtain the cascading feature FUSE. 5–i ,Right now When i is received, the received feature is the downsampled fused feature.

[0069] Step 2.2.3: Apply the above cascaded feature FUSE 5–i The data is processed sequentially through two convolutional units. Each convolutional unit consists of three parts: a convolutional layer, a batch normalization layer, and a ReLU activation layer. The convolutional units used here are the same as those used in downsampling channel 1 and channel 2.

[0070] After the above steps, the main upsampling channel yields the deconvolution module UM. i Output UM i (Out).

[0071] After upsampling is completed, the upsampled output UM4(Out) is obtained. The size of UM4(Out) has been restored to the same size as the input image.

[0072] Step 3: After the upsampling operation is completed, the upsampling output feature map UM4(Out) is obtained. After one step of deconvolution and activation by the Sigmoid function, the change detection feature map is obtained. The loss is calculated based on the real image change label using a hybrid loss function based on binary cross-entropy and Dice. The network parameters are continuously updated through the backpropagation algorithm. After multiple iterations, the model with the optimal relative parameters is obtained and the final change detection result map is generated.

[0073] Specifically, firstly, a deconvolution and sigmoid operation are performed on UM4(Out). The loss is calculated based on the change probability map and the change labels in the real image. The network parameters are continuously updated through backpropagation. After multiple iterations, the model with the optimal relative parameters is obtained, and a change detection result map is generated. The learning rate is set to 10. -5 The model is trained iteratively using the Adam optimization function. The iteration stops when the model parameters reach a relative optimum, and then a change detection result map is generated. The loss function (LOSS) is determined by the following formula:

[0074] LOSS = L bce +λL dice

[0075] Where L bce L represents the weighted binary cross-entropy loss function. dice Let λ represent the Dice loss function, and λ represent the coefficient that balances the two loss functions. In this embodiment of the invention, let λ = 0.5.

[0076]

[0077] Where β represents the proportion of unchanged pixels in the actual change map, Y + and Y - Let represent the sets of changed and unchanged pixels in the actual change map, respectively. Pr(.) is the output class probability of pixel j after passing through the Sigmoid layer, and y j Represents the category label of pixel j, y j =1 indicates that pixel j belongs to the change class, y j =0 indicates that pixel j belongs to the unchanged class.

[0078]

[0079] Where Y and These represent the change detection map predicted by the model and the actual change map of the two periods of imagery, respectively.

[0080] This invention proposes a remote sensing image change detection technique that combines feature fusion and enhanced difference analysis. It extracts features from images from two different time periods using two downsampling channels, demonstrating strong feature extraction capabilities and capturing richer information. An auxiliary upsampling channel constructed using depthwise separable convolution compensates for the feature loss of the main upsampling channel, resulting in more accurate change detection results. Simultaneously, the proposed technique considers the enhanced difference features at various scales between the two channels during downsampling, making the extracted information richer and more detailed. Through these measures, this invention achieves more accurate change detection results.

[0081] To verify the change detection performance of this invention, it is compared with two advanced deep learning change detection techniques. The two comparison techniques are: the Unet network with a standard encoder-decoder structure and the Unet++ network with multiple outputs. Four widely used quantitative metrics will be used to evaluate the performance of different change detection techniques: accuracy, precision, recall, and F1 score. Figures 5-7 Change detection maps of example images obtained using experimental data from Unet++, Unet, and embodiments of the present invention are presented. Table 3 provides quantitative evaluation indicators for different change detection techniques.

[0082] Table 3 Statistical Indicators of Change Detection Chart

[0083]

[0084] from Figures 5-7 As can be seen, compared with the two existing advanced deep learning change detection techniques, the present invention obtains a superior change detection map, and the change detection map of the present invention is closest to the real change map. Table 3 shows that all three quantitative evaluation indicators of the present invention reach their optimal values. For example, the F1 value of the present invention is 0.9063, which is 1.88%-4.24% higher than the F1 values ​​of other methods.

Claims

1. A method for detecting changes in remote sensing images that combines feature fusion and enhanced difference analysis, characterized in that, Includes the following steps: Step 1: First, preprocess the two remote sensing images, then input the two images into two downsampling channels with the same parameters to extract image features; and during the downsampling process, obtain the different scale enhancement difference features of the two channels. Step 2: Concatenate the features obtained from downsampling of the two channels and the enhancement difference features obtained from the last downsampling. Use the concatenated image features as the input of the two upsampling channels. During the upsampling process, concatenate the shallow features at each scale, the enhancement difference features at each scale, and the features of the auxiliary upsampling channel to the same scale feature layer through short connections. Step 3: After the upsampling operation is completed, the output of the last upsampling layer is deconvolved and Sigmoided to obtain the image change probability map. The loss is calculated based on the real image change label using a hybrid loss function based on binary cross-entropy and Dice. The network parameters are continuously updated through the backpropagation algorithm. After multiple iterations, the model with the optimal relative parameters is obtained and the final change detection result map is generated. In step 1, features are extracted using two downsampling channels, taking into account the differences between the two downsampling channels, as detailed below: 1) Let the five downsampling convolutional modules in channel 1 be denoted as T1 i , i = 1, 2, 3, 4, 5, no. i Convolutional modules T1 i The input and output are denoted as respectively. T1 i ( In )and T1 i ( Out The first convolutional module T1 The input for 1 is X 1, that is T1 1( In ) = X 1. Subsequent convolutional modules T1 i Input T1 i ( In )for MAX ( T1 i-1 ( Out )) MAX () indicates a max pooling operation; 2) Let the five downsampling convolutional modules in channel 2 be denoted as... T2 i , i = 1, 2, 3, 4, 5, no. i Convolutional modules T2 i The input and output are denoted as respectively. T2 i ( In )and T2 i ( Out The first convolutional module T2 The input for 1 is X 2, that is T2 1( In ) = X 2. Subsequent convolutional modules T2 i Input T2 i ( In )for MAX ( T2 i-1 ( Out )), MAX () indicates a max pooling operation; 3) While downsampling the two channels, generate enhanced differential features from the features at each scale of the two channels. STCA i It is determined by the following formula: ； ； ； in i = 1, 2, 3, 4, 5, T1 i ( Out ) is channel 1 i The output of each downsampled convolutional module T2 i ( Out ) is channel 2 i The output of each downsampled convolutional module P i1 Indicates channel 1 i The output feature maps of each convolutional module P i2 Indicates channel 2 i The feature maps output by each convolutional module STCA i This represents the first convolution channel generated by two convolution channels. i A enhanced differential feature map.

2. The method according to claim 1, characterized in that, In step 1, specifically, the two-stage image input network → T1 1 and T2 1 convolutional layer → First STCA module → First convolutional output MAX ( T1 1( Out )) MAX ( T2 1( Out )) → T1 2 and T2 2 convolutional layers → second STCA module → second convolutional output MAX ( T1 2( Out )) MAX ( T2 2( Out )) → T1 3 and T2 3 convolutional layers → third STCA module → third layer convolutional output MAX ( T1 3( Out )) MAX ( T2 3( Out )) → T1 4 and T2 4 convolutional layers → fourth STCA module → fourth layer convolutional output MAX ( T1 4( Out )) MAX ( T2 4( Out )) → T1 5 and T2 5 convolutional layers → 5th STCA module → 5th layer convolutional output MAX ( T1 5( Out )) MAX ( T2 5( Out )).

3. The method according to claim 1, characterized in that: In step 2, the highest-level features obtained by downsampling channel 1 and channel 2 are used. T1 5( Out ), T2 5( Out ), and the enhanced difference features of the highest layer features of the two channels. STCA 5. Perform cascading operations to obtain cascading features. FUSE ,Right now ;Will FUSE The upsampling operation is performed as the input to the main upsampling channel and the auxiliary upsampling channel, specifically including the following sub-steps: Step 2.1: In the auxiliary upsampling channel, denoted as 4 deconvolution modules UA i , i = 1,2,3,4,th i One deconvolution module UA i The input and output are denoted as respectively. UA i ( In )and UA i ( Out The image is gradually restored to the input size by passing through four deconvolution modules that are different from the main channel. The output feature map size of each layer is the same as the scale of the corresponding layer in the main channel, so that it can be cascaded into the main channel to supplement the features lost by the main channel deconvolution. Step 2.2: In the main upsampling channel, the image is also sequentially passed through four deconvolution modules to restore the image scale to the input size; these four deconvolution modules are denoted as... UM i , i = 1, 2, 3, 4, no. i One deconvolution module UM i The input and output are denoted as respectively. UM i ( In )and UM i ( Out ).

4. The method according to claim 3, characterized in that, In step 2.1, each deconvolution module UA i Where i = 1, 2, 3, 4, the following operations are performed: Step 2.1.1: Input UA i ( In Perform a deconvolution operation, denoted as F. Then, deconvolution of the auxiliary upsampling channel is denoted as F1. The output after the deconvolution operation is F1( UA i ( In )); Step 2.1.2: Convert the output F1 after the deconvolution operation... UA i ( In Perform depthwise separable convolution operations; Step 2.1.3: Receive the output from the previous step and perform the same operation; After the above three steps, we obtain the deconvolution modules for each layer of the auxiliary upsampling channel. UA i Output UA i ( Out ).

5. The method according to claim 3, characterized in that, In step 2.2, each deconvolution module UM i Where i = 1, 2, 3, 4, the following operations are performed: Step 2.2.1: Input UM i ( In Perform a deconvolution operation, denoted as F. Then, the deconvolution operation of the main upsampling channel is denoted as F2. The output after the deconvolution operation is F2(...). UM i ( In )); Step 2.2.2: Downsample channel 1 and channel 2's 5th – i The output of each downsampling convolution module T1 5- i ( Out ), T2 5- i ( Out Enhanced differential features between downsampling channel 1 and channel 2 STCA 5- i The deconvolution output F2 of step 2.2.1 in the main channel ( UM i ( In ), and the output of the corresponding scale auxiliary upsampling channel. UA i ( Out Cascading operations are performed to obtain cascading features. FUSE i ,Right now Where i=2,3,4, when i=1, it receives the deconvolutional features after downsampling and fusion of the last layer and the fused features of the corresponding layer. ; Step 2.2.3: Combine the above cascaded features FUSE i The process is sequentially handled by two convolutional units, each consisting of a convolutional layer, a batch normalization layer, and a ReLU activation layer. The convolutional units used here are the same as those used in downsampling channel 1 and channel 2. The main upsampling channel is processed through steps 2.2.1 to 2.2.3 to obtain the deconvolution module. UM i Output UM i ( Out ).

6. The method according to any one of claims 3 to 5, characterized in that, The specific process of step 2 is as follows: FUSE → F( FUSE → F1 UA 1 ( In F2 and F2 UM 1( In Deconvolutional layer → UA 1( O ut) MAX ( T1 4( Out )) MAX ( T2 4( Out F2 UM 1( In )) STCA 4 fusion to UM 1. In the main deconvolution channel → UM 1( Out → F1 UA 2 ( In F2 and F2 UM 2( In Deconvolutional layer → UA 2( Out ), MAX ( T1 3( Out )) MAX ( T2 3( Out F2 UM 2( In )) STCA 3 fusion to UM 2. In the main deconvolution channel → UM 2( Out → F1 UA 3 ( In F2 and F2 UM 3( In Deconvolutional layer → UA 3( Out ), MAX ( T1 2( Out )) MAX ( T2 2( Out F2 UM 3( In )) STCA 2 fused to UM 3. In the main deconvolution channel → UM 3( Out ) → F1 ( UA 4 ( In F2 and F2 UM 4( In Deconvolutional layer → UA 4( Out ), MAX ( T1 1( Out )) MAX ( T2 1( Out F2 UM 4( In )) STCA 1. Merge into UM 4. In the main deconvolution channel → UM 4( Out ).

7. The method according to claim 1, characterized in that: In step 3, the upsampling output feature map is obtained after the upsampling operation is completed. UM 4( Out After a deconvolution and sigmoid activation step, a change detection feature map is obtained. A hybrid loss function based on binary cross-entropy and Dice is used to calculate the loss based on the change labels of the real image. The network parameters are continuously updated through backpropagation. After multiple iterations, the model with the optimal relative parameters is obtained and the final change detection result map is generated.

Images