Remote sensing image detection method based on patch channel exchange and feature double-flow interaction

Abstract

The invention discloses a remote sensing image detection method based on patch channel exchange and feature double-flow interaction. The method comprises the following steps: 1) constructing a data set; the method comprises the steps of (1) building a patch channel exchange PCE module, (2) building a patch channel exchange PCE module, (3) building a double-branch difference feature DBFD module, and (4) verifying and evaluation.According to the method, higher robustness and accuracy are achieved in the aspects of building boundary detail keeping and complex scene change recognition, and the effectiveness and universality of change detection can be improved.

Description

Technical Field

[0001] The invention relates to remote sensing image processing technology, in particular to a remote sensing image detection method based on patch channel exchange and feature dual-stream interaction. Background Art

[0002] Remote sensing image change detection technology is mainly devoted to analyzing images acquired in different periods of the same geographical area to identify information about changes in specific objects over time. As a key remote sensing analysis tool, this technology has been widely used in fields such as urban construction planning, vegetation cover monitoring, and disaster assessment, effectively promoting the rational planning and utilization of regional resources. In recent decades, change detection technology has undergone a rapid evolution from traditional methods to deep learning methods. Modern high-resolution remote sensing image change detection has significantly improved detection accuracy and efficiency with the help of convolutional neural networks (CNNs), while significantly reducing the cost of manual identification. Therefore, change detection technology plays a vital role in earth observation and resource management.

[0003] With the rapid development of deep learning, CNN-based methods, especially encoder-decoder architectures and Siamese twin networks, have brought new opportunities for change detection. Two structures are widely used in the encoder-decoder architecture: multiple encoders with a single decoder structure and dual encoders with dual decoders. Among them, the single decoder structure may be affected by irrelevant differences between dual-phase data, such as seasonal changes in vegetation, thus affecting network performance. The dual decoder uses two decoders to parse the encoding features of dual-phase images separately, effectively solving the problem of information loss caused by mixed encoding features. These methods based on Siamese twin networks can not only process dual-phase images simultaneously, but also automatically extract multi-level features to achieve efficient analysis of change areas.

[0004] Feature interaction and difference modeling technologies provide important technical support for change detection. The core of feature interaction is to make full use of the correlation information between dual-phase images to enhance the effectiveness of feature extraction; while difference modeling focuses on accurately capturing change information and improving detection accuracy through explicit difference feature expression. Traditional encoder and decoder structures often lack full use of difference information, and although the attention mechanism can improve performance, the computational overhead is large.

[0005] The difference modeling method based on feature interaction has shown significant potential in the field of remote sensing change detection. Compared with traditional methods, this method can better capture the correlation features of dual-phase images, while improving detection accuracy through accurate difference modeling. However, with the increasing demand for change detection accuracy, existing methods still face challenges in feature expression: on the one hand, higher-quality encoding features are needed to capture the subtle differences between dual-phase images; on the other hand, how to effectively model and utilize difference features remains a key challenge. Current research still faces problems such as insufficient feature interaction precision and insufficient expression of difference features. Future research needs to focus on solving the quality problems of feature interaction and the efficiency problems of difference modeling to further improve the practicality of change detection technology. Summary of the invention

[0006] The purpose of the present invention is to provide a remote sensing image detection method based on patch channel exchange and feature dual stream interaction to address the deficiencies of the prior art. This method has stronger robustness and accuracy in maintaining building boundary details and identifying changes in complex scenes, and can improve the effectiveness and universality of change detection.

[0007] The technical solution for achieving the purpose of the present invention is:

[0008] A remote sensing image detection method based on patch channel exchange and feature dual-stream interaction includes the following steps:

[0009] 1) Constructing datasets: Three representative public datasets, SYSU, LEVIR-CD, and WHU, are used to verify the effectiveness of the method. The SYSU dataset is released by Sun Yat-sen University and contains dual-phase image pairs taken by aerial remote sensing of Guangzhou city. It mainly records the change information in the process of urban construction. The image resolution is 0.5 meters per pixel and contains changes in various types of objects including buildings, roads, and vegetation. The LEVIR-CD dataset is a large-scale remote sensing building change detection dataset released by Wuhan University. It contains satellite images from multiple cities in Texas, with a time span from 2002 to 2018 and a spatial resolution of 0.5 meters per pixel. It mainly focuses on the changes in buildings. The WHU dataset is released by Wuhan University and is a high-resolution building change detection dataset. It contains orthophotos obtained from Google Earth with a spatial resolution of 0.2 meters per pixel, covering the changes in buildings in urban areas with precise pixel-level annotations. These three datasets have different geographical features and change types, which can comprehensively evaluate the performance of change detection algorithms.

[0010] 2) Constructing the Patch Channel Exchange (PCE) module: The PCE module performs feature exchange and fusion in the channel dimension in the form of patches, which not only prevents the redundancy of excessive feature separation, but also improves the efficiency of feature interaction between the two time phases. First, the two-phase image features are divided into several patches in the channel dimension. 1 and T 2 The obtained features are denoted as F 1 and F 2 In the channel mask processing, global average pooling Avgpool is used to obtain the response value of each channel, and the difference between the channel and the feature is calculated to obtain the mask matrix, as shown in formula (1). This mask indicates which channels in the current phase contain more important feature information:

[0011] CF m =CF i -Avgpool(CF i ),i∈(1,2) (1),

[0012] Among them, CF m The mask matrix representing the channel features, CF i represents the channel features of phase i, Avgpool represents the average pooling operation, and the patch-level channel exchange operation is used to promote the interaction of dual-phase features, including the specific implementation of channel mask generation and feature exchange;

[0013] After obtaining the channel mask, T 1 The channels with large mask values ​​in the feature are swapped to T 2 The corresponding channel position of the feature, and vice versa, so that information can be exchanged between the channels of the dual-phase image features to obtain a feature expression containing more complementary information between the phases, as shown in formula (2). 1 Sort the channels of F in descending order by mask size and replace F with a portion of them 2 The corresponding channels in are exchanged, and the feature after exchange is expressed as F pce :

[0014] F pce =Exchange(F 1 ,F 2 ,CF 1 ,CF 2 ), (2),

[0015] Among them, Fpce represents the feature after patch channel exchange, Exchange(·) represents the channel exchange operation based on the mask size;

[0016] 3) Constructing a dual branch feature difference DBFD (Dual Branches Feature Difference, DBFD for short) module: The DBFD module decodes the encoding features of the dual branches and generates difference features to achieve more targeted and robust change area judgment. Before entering the DBFD module, the features output by the PCE encoder are first input into two parallel decoding branches. Each branch contains several D-blocks for feature upsampling and multi-scale fusion, which not only maintains the consistency of resolution but also prepares for subsequent difference feature modeling. In order to further highlight the dual-phase difference at the feature level, it is necessary to explicitly calculate the difference of the dual branch output features, as shown in formula (3):

[0017] F s =|T 1 -T 2 ∣, (3),

[0018] Among them, F s Represents the constructed explicit difference features, which are then input into the difference mask submodule. The difference mask submodule generates a mask matrix by comparing with the global average pooling value, indicating the channels with important difference information, thereby filtering the interference of background or pseudo-change areas. This process is shown in formula (4):

[0019] F m =F li -Avgpool(F li ),i∈(1,2),m∈(1,5), (4),

[0020] Among them, Fm represents the feature mask matrix, F li Represents the l-th layer feature of phase i, and generates a differential feature F containing differential expressions that are more sensitive to the changed area by masking the non-differential information. d As shown in formula (5), the difference features of each decoder layer are fused to obtain the final change feature F c :

[0021]

[0022] Where l represents the number of decoder layers, F dk represents the difference feature of the k-th decoder layer;

[0023] 4) Verification and evaluation: Four standard indicators, namely precision, recall, F1 score and intersection-over-union, are used to comprehensively evaluate the change detection performance: Precision P (P for short) measures the accuracy of the detection results, indicating the proportion of pixels that have actually changed among all pixels predicted by the model as changed areas. The higher the precision, the fewer false positives the model generates, i.e., the model mistakenly identifies unchanged areas as changed areas; Recall R (R for short) measures the completeness of the detection results, indicating the proportion of pixels correctly detected by the model among all pixels that have actually changed. The higher the recall, the more the model misses, i.e., fails to identify actual changes. The smaller the area; the F1 score is the harmonic mean of precision and recall, providing a balanced evaluation of the overall performance of the model. The higher the F1 score, the better the overall detection ability of the model while maintaining high precision. The intersection over union (IoU) evaluates the detection accuracy by calculating the overlap between the predicted change area and the actual change area. It is the intersection of the predicted area and the actual area divided by their union. The higher the IoU value, the higher the overlap between the change area predicted by the model and the actual change area, and the more accurate the detection result. These four indicators complement and support each other, and together provide a comprehensive evaluation of the model detection performance.

[0024] The patch channel exchange PCE module in this technical solution realizes feature exchange at the patch level, ensuring that the bi-phase features can interact in a more fine-grained manner, which not only improves the accuracy of feature extraction, but also effectively maintains the unique information of the bi-phase images, thereby enhancing the model's ability to identify the changed area;

[0025] The dual-branch feature difference DBFD module in this technical solution adopts a dual-branch decoding architecture and enhances the ability to extract change information through an explicit difference modeling strategy. The DBFD module achieves efficient difference feature expression through multi-level feature fusion, while avoiding the computational burden brought by the traditional attention mechanism, thereby improving the accuracy and efficiency of change detection.

[0026] In the joint and collaborative optimization of the patch channel exchange PCE module and the dual-branch feature difference DBFD module, PCE provides high-quality dual-phase feature interaction, while DBFD further enhances the ability to extract change information on this basis. The combination of the two significantly improves the detection ability of complex scenes and subtle changes. Through the organic combination of feature interaction and difference modeling, a more efficient and robust change detection model is constructed.

[0027] This method is more robust and accurate in preserving building boundary details and identifying changes in complex scenes, and can improve the effectiveness and universality of change detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the overall structure of the invention in the embodiment;

[0029] Figure 2 An example diagram of a PCE module in an embodiment;

[0030] Figure 3 Schematic diagram of a DBFD module in an embodiment;

[0031] Figure 4 This is a visualization diagram of an ablation experiment in the embodiment;

[0032] Figure 5 It is a visual schematic diagram of the comparative experiment in the embodiment. DETAILED DESCRIPTION

[0033] The content of the present invention is further described below in conjunction with the drawings and embodiments, but the present invention is not limited thereto.

[0034] Example:

[0035] Reference Figure 1 , a remote sensing image detection method based on patch channel exchange and feature dual-stream interaction, comprising the following steps:

[0036] 1) Constructing datasets: Three representative public datasets, SYSU, LEVIR-CD, and WHU, are used to verify the effectiveness of the method. The SYSU dataset is released by Sun Yat-sen University and contains dual-phase image pairs taken by aerial remote sensing of Guangzhou city. It mainly records the change information in the process of urban construction. The image resolution is 0.5 meters per pixel and contains changes in various types of objects including buildings, roads, and vegetation. The LEVIR-CD dataset is a large-scale remote sensing building change detection dataset released by Wuhan University. It contains satellite images from multiple cities in Texas, with a time span from 2002 to 2018 and a spatial resolution of 0.5 meters per pixel. It mainly focuses on the changes in buildings. The WHU dataset is released by Wuhan University and is a high-resolution building change detection dataset. It contains orthophotos obtained from Google Earth with a spatial resolution of 0.2 meters per pixel, covering the changes in buildings in urban areas with precise pixel-level annotations. These three datasets have different geographical features and change types, which can comprehensively evaluate the performance of change detection algorithms.

[0037] 2) Build the patch channel switching PCE module: Figure 2 As shown in Figure 1, the PCE module performs feature exchange and fusion in the channel dimension in the form of patches, which not only prevents the redundancy of excessive feature separation, but also improves the efficiency of feature interaction between the two time phases. First, the two-phase image features are divided into several patches in the channel dimension.1 and T 2 The obtained features are denoted as F 1 and F 2 In the channel mask processing, global average pooling Avgpool is used to obtain the response value of each channel, and the difference between the channel and the feature is calculated to obtain the mask matrix, as shown in formula (1). This mask indicates which channels in the current phase contain more important feature information:

[0038] CF m =CF i -Avgpool(CF i ),i∈(1,2) (1),

[0039] Among them, CF m The mask matrix representing the channel features, CF i represents the channel features of phase i, Avgpool represents the average pooling operation, and the patch-level channel exchange operation is used to promote the interaction of dual-phase features, including the specific implementation of channel mask generation and feature exchange;

[0040] After obtaining the channel mask, T 1 The channels with large mask values ​​in the feature are swapped to T 2 The corresponding channel position of the feature, and vice versa, so that information can be exchanged between the channels of the dual-phase image features to obtain a feature expression containing more complementary information between the phases, as shown in formula (2). 1 Sort the channels of F in descending order by mask size and replace F with a portion of them 2 The corresponding channels in are exchanged, and the feature after exchange is expressed as F pce :

[0041] F pce =Exchange(F 1 ,F 2 ,CF 1 ,CF 2 ), (2),

[0042] Among them, Fpce represents the feature after patch channel exchange, Exchange(·) represents the channel exchange operation based on the mask size;

[0043] 3) Construct a dual-branch difference feature DBFD module: Figure 3As shown in the figure, the DBFD module decodes the encoded features of the dual branches and generates difference features to achieve more targeted and robust change area judgment. Before entering the DBFD module, the features output by the PCE encoder are first input into two parallel decoding branches. Each branch contains several D-blocks for feature upsampling and multi-scale fusion, which not only maintains the consistency of resolution but also prepares for subsequent difference feature modeling. In order to further highlight the dual-phase difference at the feature level, it is necessary to explicitly calculate the difference of the dual-branch output features, as shown in formula (3):

[0044] F s =|T 1 -T 2 ∣, (3),

[0045] Among them, F s Represents the constructed explicit difference features, which are then input into the difference mask submodule. The difference mask submodule generates a mask matrix by comparing with the global average pooling value, indicating the channels with important difference information, thereby filtering the interference of background or pseudo-change areas. This process is shown in formula (4):

[0046] F m =F li -Avgpool(F li ),i∈(1,2),m∈(1,5), (4),

[0047] Among them, Fm represents the feature mask matrix, F li Represents the l-th layer feature of phase i, and generates a differential feature F containing differential expressions that are more sensitive to the changed area by masking the non-differential information. d As shown in formula (5), the difference features of each decoder layer are fused to obtain the final change feature F c :

[0048]

[0049] Where l represents the number of decoder layers, F dk represents the difference feature of the k-th decoder layer;

[0050] 4) Verification and evaluation: Four standard indicators, namely precision, recall, F1 score and intersection-over-union, are used to comprehensively evaluate the change detection performance: Precision P (P for short) measures the accuracy of the detection results, indicating the proportion of pixels that have actually changed among all pixels predicted by the model as changed areas. The higher the precision, the fewer false positives the model generates, i.e., the model mistakenly identifies unchanged areas as changed areas; Recall R (R for short) measures the completeness of the detection results, indicating the proportion of pixels correctly detected by the model among all pixels that have actually changed. The higher the recall, the more the model misses, i.e., fails to identify actual changes. The smaller the area; the F1 score is the harmonic mean of precision and recall, providing a balanced evaluation of the overall performance of the model. The higher the F1 score, the better the overall detection ability of the model while maintaining high precision. The intersection over union (IoU) evaluates the detection accuracy by calculating the overlap between the predicted change area and the actual change area. It is the intersection of the predicted area and the actual area divided by their union. The higher the IoU value, the higher the overlap between the change area predicted by the model and the actual change area, and the more accurate the detection result. These four indicators complement and support each other, and together provide a comprehensive evaluation of the model detection performance.

[0051] In this example, PyTorch is used to conduct experimental verification on a NVIDIA GTX 4090 GPU (24GB video memory), using Python 3.10 programming language and PyTorch 2.0.1+cu117 deep learning framework. During the training process, data enhancement is performed by random flipping, transposition, translation, scaling and rotation. The loss function uses a combination of binary cross entropy loss and Dice coefficient loss. In this example, the AdamW optimizer is used, the initial learning rate is 0.001, and the weight decay is 0.001. In addition, if the F1 score of the validation set does not improve for 12 consecutive epochs, the learning rate is reduced to 0.1 times the original, the batch size is set to 32, and the network training is performed for a total of 300 epochs. Comparisons and ablation experiments are performed on three different datasets: SYSU, LEVIR-CD and WHU to fully verify the effectiveness of the method:

[0052] As shown in Table 1, the basic EDED backbone network is used as the baseline, and the PCE module and DBFD module are gradually added for comparative analysis. The experimental results show that on the SYSU dataset, after adding the PCE module, the F1 score of the model is increased from 78.85% to 81.80%, and the IoU is increased from 65.08% to 69.20%. This confirms that the feature exchange of the PCE module in the channel dimension can effectively enhance the complementarity of the bi-phase features. After further adding the DBFD module, the method (OURS) in this example has achieved the best performance in all indicators, with the F1 score and IoU reaching 82.22% and 69.81% respectively; similar trends are also observed on the LEVIR-CD and WHU datasets. This method is superior to the baseline model and the PCE module alone in all evaluation indicators. These results fully confirm that the synergy of the PCE module and the DBFD module can significantly improve the performance of change detection:

[0053] Table 1:

[0054]

[0055] In order to comprehensively evaluate the performance of this method, comparative experiments with multiple latest or most classic change detection methods are analyzed on three benchmark datasets, including FC-EF, FC-Sima-diff, FC-Siam-conc, AMTNet-50 and SGSLN. The experimental results on SYSU and LEVIR-CD datasets are shown in Table 2:

[0056] Table 2:

[0057]

[0058]

[0059] In Table 2, on the SYSU dataset, this method achieved an F1 score of 82.22% and an IoU of 69.81%, which are 4.95 and 6.85 percentage points higher than the classic method FC-EF, respectively. Compared with the second-best method SGSLN, it has better balanced the precision and recall while surpassing the F1 score and IoU, which shows that the feature exchange mechanism of this method can better handle complex urban change scenarios. On the LEVIR-CD dataset, this method also achieved the highest F1 score and IoU, which fully confirms the effectiveness of this method.

[0060] As shown in Table 3, on the WHU dataset, the proposed method achieved the best performance in all evaluation indicators, with F1 score and IoU reaching 93.11% and 87.11% respectively. Compared with AMTNet-50 and SGSLN with better performance, the proposed method still achieved an F1 score improvement of 0.84-1.08 percentage points and an IoU improvement of 1.47-1.87 percentage points. This result fully demonstrates the superiority of the proposed method in the task of building change detection:

[0061] Table 3:

[0062]

[0063] The experimental results of the three datasets show that this method has shown stable and superior performance in different types of change detection scenarios through effective feature interaction and difference modeling strategies. In particular, when dealing with complex urban changes and building changes, the method in this study has obvious advantages over existing methods. These results verify the effectiveness and universality of the PCE module and DBFD module of this method in improving change detection performance.

[0064] In order to further verify the effectiveness of the PCE module and the DBFD module, this example conducts a detailed visual comparative analysis of the ablation experiment, such as Figure 4 As shown, in Figure 4 The following figure shows the detection results of the backbone network EDED, after adding the PCE module, and after further adding the DBFD module:

[0065] From the visualization results of the SYSU dataset, it can be observed that when detecting object change areas, the performance of the model is gradually improved with the gradual addition of modules, especially in the complete model, which overcomes the interference of the object color being similar to the color of the ocean in the second phase, and the detection results of object changes are more complete and accurate. In the experiment of the LEVIR-CD dataset, the gradual improvement of the detection effect can be clearly seen in the evolution from EDED to EDED+PCE and then to the complete model. The complete model not only inherits the advantages of the PCE module in building edge detection, but the introduction of the DBFD module further improves the model's grasp of details, making the detection results more accurate, and the cases of false detection and missed detection are reduced. The complete model shows the best performance in the WHU dataset with complex building structures. The detection of building outlines is clearer and more accurate, and the boundaries are more complete. This proves the important role of the DBFD module in detail feature extraction and boundary optimization. These visualization results fully demonstrate that the synergy of the PCE module and the DBFD module can significantly improve the overall performance of the model. By better utilizing the feature information of multi-temporal images and optimizing boundary details, the accuracy and reliability of change detection are greatly improved. The experimental results not only verify the importance of the two modules in improving detection accuracy, but also confirm their adaptability and stability in dealing with different scenarios and complex change patterns.

[0066] In order to comprehensively evaluate the performance of this method, the model of this method is visually compared with the current mainstream change detection methods AMTNet and SGSLN. Figure 5 The detection results on three benchmark datasets are shown:

[0067] On the SYSU dataset, compared with AMTNet and SGSLN, this method shows significant advantages in dealing with building changes in complex scenes, especially in the building area above the image, where AMTNet has more missed detections, and although SGSLN has improved, it still has the problem of incomplete detection. This method can capture the changed area more accurately, and the detection result is closer to the true label (GT). In the arc-shaped building scene of the LEVIR-CD dataset, AMTNet and SGSLN have different degrees of false detection and missed detection at the edge of the building. Taking the building in the upper left corner of the picture as an example, this method shows stronger boundary perception ability, which can not only It can accurately identify the changed area and maintain the integrity and continuity of the building. On the WHU dataset with regular building structures, this method shows obvious advantages in the refined detection of building contours. Compared with the rough detection results of AMTNet at the building boundaries and the local false detection problems of SGSLN, this method can better maintain the geometric features of the building and produce more accurate change detection results. These visual comparisons fully demonstrate that this method has stronger robustness and accuracy when dealing with change detection tasks in different scenes and scales, especially in maintaining the details of building boundaries and identifying changes in complex scenes.

[0068] This method significantly improves the accuracy and robustness of remote sensing image change detection through the collaborative design of PCE and DBFD modules. The experimental results on the three benchmark datasets of SYSU, LEVIR-CD and WHU show that this method achieves F1 scores of 82.22%, 92.50% and 93.11% respectively, which has obvious advantages over existing methods, especially when dealing with complex urban changes and building changes, it shows stronger feature extraction ability and boundary perception ability. This method has broad application prospects and can be applied to: urban construction planning and monitoring, dynamic assessment of vegetation cover, disaster assessment and emergency response, ecological environment protection and land use change analysis. With the continuous growth of remote sensing data and the deepening of its application in various fields, this method shows great potential in improving the efficiency and accuracy of change detection, which will provide important support for the further development of remote sensing technology.

Claims

1. A remote sensing image detection method based on patch channel exchange and feature dual-stream interaction, characterized in that: The steps include: 1) Constructing datasets: Three public datasets, SYSU, LEVIR-CD and WHU, are used to verify the effectiveness of the method. The SYSU dataset is released by Sun Yat-sen University and contains dual-phase image pairs taken by aerial remote sensing of Guangzhou city. The image resolution is 0.5 m / pixel and contains changes in various types of objects including buildings, roads and vegetation. The LEVIR-CD dataset is a large-scale remote sensing building change detection dataset released by Wuhan University. It contains satellite images from multiple cities in Texas, with a time span from 2002 to 2018 and a spatial resolution of 0.5 m / pixel. It mainly focuses on the changes in buildings. The WHU dataset is released by Wuhan University and is a high-resolution building change detection dataset. It contains orthophotos obtained from Google Earth with a spatial resolution of 0.2 m / pixel. 2) Construct the patch channel exchange PCE module: First, the dual-phase image features are divided into several patches in the channel dimension. The features obtained from phases T1 and T2 are represented as F1 and F2 respectively. In the channel mask processing, the global average pooling Avgpool is used to obtain the response value of each channel, and the difference with the feature is calculated by channel to obtain the mask matrix, as shown in formula (1): CF m =CF i -Avgpool(CF i ),i∈(1,2) (1), Among them, CF m The mask matrix representing the channel features, CF i represents the channel feature of phase i, Avgpool represents the average pooling operation. After obtaining the channel mask, the channel with a large mask value in the T1 feature is exchanged to the corresponding channel position of the T2 feature, and vice versa. In this way, information is exchanged between the channels of the dual-phase image features to obtain a feature expression containing more complementary information between phases, as shown in formula (2). The channels of F1 are sorted in descending order according to the mask size, and a part of them is used to replace the corresponding channels in F2 to complete the exchange. The exchanged feature is represented as F pce : F pce =Exchange(F1,F2,CF1,CF2), (2), Among them, Fpce represents the feature after patch channel exchange, Exchange(·) represents the channel exchange operation based on the mask size; 3) Constructing a dual-branch difference feature DBFD module: Before entering the DBFD module, the features output by the PCE encoder are first input into two parallel decoding branches. Each branch contains several D-blocks for feature upsampling and multi-scale fusion. The difference of the dual-branch output features is explicitly calculated, as shown in formula (3): F s =∣T1-T2∣, (3), Among them, F s Represents the constructed explicit difference features, which are then input into the difference mask submodule. The difference mask submodule generates a mask matrix by comparing with the global average pooling value, indicating the channels with important difference information and filtering the interference of background or pseudo-change areas. The process is shown in formula (4): F m =F li -Avgpool(F li ),i∈(1,2),m∈(1,5), (4), Among them, Fm represents the feature mask matrix, F li Represents the l-th layer feature of phase i, and generates a differential feature F containing differential expressions that are more sensitive to the changed area by masking the non-differential information. d As shown in formula (5), the difference features of each decoder layer are fused to obtain the final change feature F c : Where l represents the number of decoder layers, F dk represents the difference feature of the k-th decoder layer; 4) Verification and evaluation: Four standard indicators, namely precision, recall, F1 score and intersection over union, are used to comprehensively evaluate the change detection performance: Precision P measures the accuracy of the detection results, indicating the proportion of pixels that have actually changed among all pixels predicted by the model as changed areas. A higher precision indicates that the model generates fewer false detections, i.e., incorrectly identifies unchanged areas as changed areas; Recall R measures the completeness of the detection results, indicating the proportion of pixels that have actually changed and are correctly detected by the model among all pixels that have actually changed. A higher recall indicates that the model misses fewer areas that have actually changed; F1-score is the harmonic mean of precision and recall, providing a balanced evaluation of the overall performance of the model. A higher F1-score indicates that the model maintains a high recall while maintaining a high precision, showing better comprehensive detection capabilities; Intersection over union (IoU) evaluates the accuracy of the detection by calculating the degree of overlap between the predicted change area and the actual change area. It is the intersection of the predicted area and the actual area divided by their union. A higher IoU value indicates that the model predicts a higher degree of overlap between the change area and the actual change area, and the more accurate the detection result.

Images