A method for constructing a change detection model for remote sensing image change detection

Abstract

This scheme discloses a method for constructing a change detection model for remote sensing image change detection. The model includes two twin networks, each of which includes an input layer, a network layer 1, a network layer 2, and a classification layer in sequence. The input of the input layer is image block 1 and image block 2. The detection model obtains a difference image of the two image blocks according to the output of the two classification layers. The input layer includes a convolution module, an FMBConv module, and an ASPP module. The network layer 1 includes a first dual-branch module and a second dual-branch module, one of which includes a deformable convolution. The network layer 2 includes a multi-head self-attention bottleneck module. This scheme uses a dual-branch structure including deformable convolution and a multi-head self-attention mechanism, so that the model increases its receptive field while ensuring the spatial resolution of the feature map, thereby improving the detection accuracy of the model.

Description

Technical Field

[0001] The invention belongs to the field of change detection technology, and in particular relates to a new model for remote sensing image change detection. Background Art

[0002] Remote sensing image change detection is a technology that analyzes and detects two or more images acquired at different times at the same geographical location to obtain information about changes in land objects. In remote sensing images, the spectral or texture characteristics of land objects can reflect the category and attribute information of land objects. Therefore, based on the changes in the spectral / texture characteristics of pixels at the same location in multi-temporal remote sensing images, the changes in surface land objects can be obtained, including the areas where land objects have changed and detailed information on the types of changes. Currently, remote sensing image change detection has been widely used in land cover and land use monitoring, urban development research, resource management, disaster assessment, and ecosystem monitoring.

[0003] According to the size of the basic unit used for the analysis object, the change detection methods can be divided into pixel-based change detection and object-based change detection. The pixel-based change detection method uses pixels as processing units to analyze the differences between pixels one by one, and uses the spectral characteristics of remote sensing images to detect changes in images. In the early days, such algorithms generally used direct analysis or post-classification comparison methods. The direct analysis method is to combine the image bands according to the gray value of the remote sensing image, and then perform difference, ratio, logarithmic ratio, principal component analysis and other operations to obtain the change difference map, and then set an appropriate threshold to analyze the change difference map to obtain specific change information. The post-classification comparison rule is to first classify the remote sensing image to be detected, and then compare the classification results to obtain the change detection results. The object-based change detection method adopts the method of first segmenting by object category (common methods include edge detection, threshold segmentation and region segmentation) and then detecting. According to the set scale parameters, pixels with similar characteristics such as spectrum, texture, and shape are divided into homogeneous regional objects, making full use of pixel texture and context information, and improving the detection accuracy to a certain extent. Traditional change detection methods mainly extract change information through unsupervised classification (such as rule sets, threshold segmentation, etc.) or machine learning (such as support vector machines, random forests, etc.) in a pixel-based or object-oriented framework. These methods rely heavily on manually constructed feature representations and have limited modeling capabilities for complex high-level change information. In addition, when the features of change and non-change classes overlap or their statistical distribution modeling is inaccurate, the detection results will be inaccurate.

[0004] Compared with traditional algorithms, deep learning models use convolutional neural networks to automatically, multi-level, and multi-dimensionally extract features of complex images, and the features extracted by deeper network structures are more robust to noise. The change detection network framework based on deep learning mainly includes the semantic segmentation early fusion framework and the twin network late fusion framework. The former connects images of different periods along the channel dimension as the input of the semantic segmentation network to extract change information. The latter uses a parallel twin network to extract features from two mutually aligned images, and then fuses features through splicing, subtraction, LSTM, Euclidean distance, etc. In order to improve detection accuracy, increasing the depth and width of the network is a common strategy, but this will introduce a large number of trainable network parameters and require a large storage space. In actual engineering applications, change detection models based on deep learning have problems such as high complexity, large number of parameters, and high computational overhead. In the process of model training and inference, the computing performance of the terminal device is high, which leads to problems such as difficulty in deploying the model on low-performance computing platforms such as embedded or mobile devices, or poor real-time detection after deployment.

[0005] In order to solve the above problems, many people in the field are currently working on the lightweight change detection model. For example, "K. Song, F. Cui, and J. Jiang, "An efficient lightweight neural network for remote sensing image change detection," Remote Sensing, vol. 13, no. 24, pp. 1-24, Dec. 2021." proposed a lightweight change detection model (3M-CDNet). This model can solve the problems of high complexity, large number of parameters, and high computational overhead of the current change detection model. However, there are still some problems, such as insufficient accuracy in small target detection and capture of boundary information, and low model interpretation accuracy. These problems will lead to inaccurate final detection results. Summary of the invention

[0006] The purpose of the present invention is to design a new network model suitable for change detection of remote sensing images in view of the above problems.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A method for constructing a change detection model for remote sensing image change detection, the model comprising two twin networks, each of which comprises an input layer, a network layer 1, a network layer 2 and a classification layer in sequence, the inputs of the two input layers are image block 1 and image block 2 respectively, the detection model obtains a difference image of the two image blocks according to the outputs of the two classification layers, the input of the network layer 1 is the output of the corresponding input layer, the input of the network layer 2 is the result of splicing the output of the corresponding input layer and the output of the network layer 1 along the channel dimension direction, and the input of the classification layer is the result of splicing the output of the corresponding input layer, the output of the network layer 1 and the output of the network layer 2 along the channel dimension direction;

[0009] The input layer includes a convolution module, an FMBConv module and an ASPP module; the network layer one includes a first dual-branch module and a second dual-branch module, both of which adopt a dual-branch structure, and one of the branches of the two dual-branch modules includes a deformable convolution to capture changing targets with different geometric features; the network layer two includes a multi-head self-attention bottleneck module.

[0010] In the above-mentioned method for constructing a change detection model for remote sensing image change detection, the FMBConv module includes a multiple convolution of a multiple of the input feature dimension of the FMBConv module and a same-dimensional convolution equal to the input feature dimension; the feature input to the FMBConv module is first input to the multiple convolution for convolution processing, the processing result of the multiple convolution is used as the input of the same-dimensional convolution, and the output of the same-dimensional convolution and the input feature of the FMBConv module are added as the output of the FMBConv module and input to the next module;

[0011] The ASSP module includes multiple branches in parallel. The features input to the APP module are respectively input to the multiple branches. Finally, the output features of the multiple branches are spliced ​​in the channel dimension and the number of output channels is adjusted using 1*1 convolution before output.

[0012] In the above-mentioned method for constructing a change detection model for remote sensing image change detection, the ASSP module includes branch one, branch two, branch three, branch four and branch five. Branch one adopts 1*1 convolution, branch two adopts 3*3 dilated convolution with a dilation rate of 6, branch three adopts 3*3 dilated convolution with a dilation rate of 12, branch four adopts 3*3 dilated convolution with a dilation rate of 18, and branch five adopts an average pooling layer for global average pooling.

[0013] In the above-mentioned method for constructing a change detection model for remote sensing image change detection, the first branch of the first dual-branch module includes a 1*1 convolution, and the second branch includes a deformable convolution. The feature map input to the first dual-branch module is respectively input to the 1*1 convolution and the deformable convolution, and the outputs of the 1*1 convolution and the deformable convolution are connected through residual connections to finally obtain a feature map with a feature dimension of 256 and a scale size of H / 4*W / 4, where H and W represent the length and width of the input feature map, respectively;

[0014] The first branch of the second dual-branch module is a direct connection structure, and the second branch includes a deformable convolution. The feature map input to the second branch module is input to the deformable convolution network. The direct connection structure directly introduces the feature map input to the second branch module, and the output of the deformable convolution network and the feature map from the direct connection structure input to the second branch module are output through a residual connection.

[0015] In the above-mentioned method for constructing a change detection model for remote sensing image change detection, the multi-head self-attention bottleneck module includes one self-attention bottleneck module 1 and two self-attention bottleneck modules 2 in sequence;

[0016] And both the self-attention bottleneck module 1 and the self-attention bottleneck module 2 adopt a dual-branch structure;

[0017] The first branch of the self-attention bottleneck module 1 includes a 1*1 convolution, and the second branch includes a multi-head self-attention module. The feature map input to the self-attention bottleneck module 1 is input to its first branch and second branch respectively, and the outputs of the first branch and the second branch are output after being residually connected;

[0018] The first branch of the second self-attention bottleneck module is a direct connection structure, and the second branch includes a multi-head self-attention module. The feature map input to the second self-attention bottleneck module is input to the second branch. The first branch directly introduces the feature map input to the second self-attention bottleneck module. The output of the second branch is connected with the feature map from the first branch and the feature map input to the second branch module through a residual connection and then output.

[0019] In the above-mentioned method for constructing a change detection model for remote sensing image change detection, the multi-head self-attention modules each include four self-attention layers and one relative position encoding module.

[0020] In the above-mentioned method for constructing a change detection model for remote sensing image change detection, the input layer includes a convolution module, two FMBConv modules, a maximum pooling layer and an ASPP module in sequence;

[0021] The network layer 1 includes a first dual-branch module and two second dual-branch modules in sequence;

[0022] The network layer 2 includes a maximum pooling layer, a multi-head self-attention bottleneck module and an upsampling module in sequence.

[0023] In the above-mentioned method for constructing a change detection model for remote sensing image change detection, the classification layer sequentially includes a first convolutional layer, a first upsampling module, a second convolutional layer, a channel attention module, a third convolutional layer, a fourth convolutional layer and a second upsampling module.

[0024] In the above-mentioned method for constructing a change detection model for remote sensing image change detection, the model is trained by the following deep supervision method:

[0025] S1. Subtract the feature maps of image block 1 and image block 2 processed by the input layer, network layer 1, network layer 2 and classification layer to form a new feature map containing change information;

[0026] S2. Perform 1*1 convolution on the new feature map so that its channel dimension is the number of classifications (1);

[0027] S3. Upsample the feature map processed by step S2 to the original input image size, and finally calculate its loss function in sequence and perform multiple back propagations.

[0028] In the above-mentioned method for constructing a change detection model for remote sensing image change detection, in step S3, formula (1) is used to calculate the loss of the prediction graphs obtained at the input layer, network layer 1, network layer 2 and classification layer, and the gradient is updated by back propagation. After completing the specified number of rounds of model training, the parameters of the model are saved and deployed to other devices to carry out prediction work:

[0029] L=L dice +L focal (1)

[0030] L focal (p t )=-α t (1-p t ) γ log(p t ) (2)

[0031]

[0032] L dice represents Dice Loss, L focal is Focal Loss, p t is the predicted probability, α and γ are adjustment parameters, ε is the smoothing factor, y is the predicted value, and t is the true value / label.

[0033] The advantages of the present invention are: using a dual-branch structure including deformable convolution and a multi-head self-attention mechanism, the model increases its receptive field while ensuring the spatial resolution of the feature map, which is conducive to the detection of small targets and narrow objects and the accurate capture of the boundary information of the changing objects, generating a prediction map with more distinct semantic features, and improving the detection accuracy of the model; using deep supervision to alleviate the gradient vanishing phenomenon, by using the ASSP module and the FMBConv module in the input layer to improve the ability of the network to handle changes of different scales, on this basis, the input layer, network layer one, network layer two and the classification layer are connected by a skip connection method, which can better fuse the features of different layers to capture the changing targets of different sizes, and alleviate the interference of the imaging angle and geometric distortion of the image on the performance of the model. Through the use of the above-mentioned modules, the interpretation accuracy of the model can be improved, the gradient disappearance and overfitting can be prevented, while ensuring the prediction accuracy and reasoning speed of the model, thereby improving the change detection accuracy achieved by the detection model. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 The overall workflow diagram of the method for constructing a change detection model for remote sensing image change detection of the present invention;

[0035] Figure 2 It is a structural diagram of the input layer in the method for constructing a change detection model for remote sensing image change detection of the present invention;

[0036] Figure 3 It is a structural diagram of the FMBConv module in the method for constructing a change detection model for remote sensing image change detection of the present invention;

[0037] Figure 4 It is a structural diagram of the ASPP module in the method for constructing a change detection model for remote sensing image change detection of the present invention;

[0038] Figure 5 It is a structural diagram of the network layer 1 in the method for constructing a change detection model for remote sensing image change detection of the present invention;

[0039] Figure 6a It is a structural diagram of a network layer DConvBottleneck-A in the method for constructing a change detection model for remote sensing image change detection of the present invention;

[0040] Figure 6b It is a structural diagram of the DConvBottleneck-B module of the network layer 1 in the method for constructing a change detection model for remote sensing image change detection of the present invention;

[0041] Figure 7 It is a structural diagram of the network layer 2 in the method for constructing a change detection model for remote sensing image change detection of the present invention;

[0042] Figure 8a It is a structural diagram of the MHSABottleneck-1 module of the network layer 2 in the method for constructing a change detection model for remote sensing image change detection of the present invention;

[0043] Figure 8b It is a structural diagram of the MHSABottleneck-2 module of the network layer 2 in the method for constructing a change detection model for remote sensing image change detection of the present invention;

[0044] Fig. 9 It is a structural diagram of the MHSA module in the method for constructing a change detection model for remote sensing image change detection of the present invention;

[0045] Fig.10 It is a structural diagram of the classification layer in the method for constructing a change detection model for remote sensing image change detection of the present invention;

[0046] Fig.11 This is a structural diagram of the channel attention mechanism in the method for constructing a change detection model for remote sensing image change detection in the present invention. DETAILED DESCRIPTION

[0047] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] This proposal proposes a change detection model for remote sensing image change detection and its construction method. Figure 1 As shown, the model includes two twin networks, each of which includes an input layer, a network layer 1, a network layer 2 and a classification layer in sequence. The inputs of the two input layers are image block 1 of the first phase and image block 2 of the second phase, respectively. The detection model obtains the depth feature map of the two image blocks according to the output of the two classification layers, and generates a difference image accordingly. The specific method of obtaining the difference image in this embodiment is: subtracting the output results of the two classification layers, and then performing a 1*1 convolution on the subtraction result and upsampling it to the size of the original input image to obtain a difference image of the original size.

[0049] In particular, the input of network layer 1 is the output of the corresponding twin network input layer, the input of network layer 2 is the output of the corresponding twin network input layer and the output of network layer 1 concatenated along the channel dimension, and the input of the classification layer is the output of the corresponding twin network input layer, the output of network layer 1, and the output of network layer 2 concatenated along the channel dimension. The change detection problem is regarded as a special semantic segmentation problem. The image block 1 and the image block 2 are processed by the input layer, network layer 1, network layer 2, and classification layer respectively. The results of the two images at different stages are then fused, and then the binary classification prediction of the change information is performed.

[0050] Specifically, Figure 2 As shown in the figure, the input layer includes a convolution module, two FMBConv modules, a maximum pooling layer and an ASPP module in sequence. Assuming that image blocks 1 and 2 of size H*W*3 are respectively input into the input layer of the twin network for related processing, the image is first input into a convolution layer with a convolution kernel of 3*3 and a stride of 2, then passes through two FMBConv modules, and finally passes through a maximum pooling layer with a window size of 3*3 and an ASPP layer. Finally, the input module resamples the input image into a feature map of size H / 4*W / 4*128.

[0051] Specifically, the convolution module includes a convolution layer with a convolution kernel of 3*3 and a step size of 2; the window size of the maximum pooling layer is 3*3.

[0052] Among them, Figure 3 As shown, H, W, C are the length, width and channel dimensions of the feature map, respectively. The FMBConv module includes a multiple convolution that is a multiple of the input feature dimension of the FMBConv module and a same-dimensional convolution that is equal to the input feature dimension; the features input to the FMBConv module are first input to the multiple convolution for convolution processing, and the processing result of the multiple convolution is used as the input of the same-dimensional convolution, and the output of the same-dimensional convolution is added to the input feature of the FMBConv module as the output of the FMBConv module and input to the next module. The introduction of the FMBConv module can enhance the information extraction capability of the model. The dimension of the multiple convolution in this embodiment is 4 times the input dimension, which enables the model to capture more fine-grained features and is also easier to train.

[0053] like Figure 4 As shown, x is the input feature map, The output after being processed by the ASSP module, the ASSP module includes multiple branches in parallel, including branch one, branch two, branch three, branch four and branch five, branch one uses 1*1 convolution, branch two uses 3*3 dilated convolution with a dilation rate of 6, branch three uses 3*3 dilated convolution with a dilation rate of 12, branch four uses 3*3 dilated convolution with a dilation rate of 18, and branch five uses an average pooling layer for global average pooling. The features input to the APP module are respectively input to the multiple branches, and finally the output features of the multiple branches are spliced ​​in the channel dimension and the output channels are adjusted using 1*1 convolution before output. This scheme adds ASPP to the input layer, which can obtain multiple receptive fields of different sizes, improve the network's ability to handle changes of different scales, and connect the input layer, network layer one, network layer two and classification layer in a skip connection manner, so as to obtain multiple receptive fields of different sizes, improve the network's ability to handle changes of different scales, better integrate the features of different layers to capture targets of different sizes, and alleviate the interference of imaging angles and geometric distortions on the performance of the model.

[0054] like Figure 5 As shown, the network layer 1 includes a first dual-branch module and two second dual-branch modules in sequence; the first dual-branch module DConvBottleneck-A and the second dual-branch module DConvBottleneck-B both adopt a dual-branch structure, and one of the branches of the two dual-branch modules includes a deformable convolution to capture changing targets with different geometric features.

[0055] Specifically, Figure 6a As shown, the first branch (i.e., the left branch in the figure) of the first dual-branch module DConvBottleneck-A includes 1*1 convolution, and the second branch (i.e., the right branch in the figure) includes deformable convolution. The feature map input to the first dual-branch module is input to the 1*1 convolution and the deformable convolution respectively. The 1*1 convolution changes the number of channels of the input feature map from 128 to 256, and then normalizes it with the BN (Batch Normalization) layer. The right branch is used to capture the changing targets of different geometric features and increase the receptive field of the feature map. The outputs of the 1*1 convolution and the deformable convolution are output through residual connections, and finally a feature map with a feature dimension of 256 and a scale of H / 4*W / 4 is obtained, where H and W represent the length and width of the input feature map, respectively.

[0056] like Figure 6bAs shown, similarly, the second dual-branch module DConvBottleneck-B is similar to the first branch module, except that the left branch of the second branch module adopts a direct connection structure, which directly introduces the feature map input to the second branch module. The second branch (right branch) also includes a deformable convolution, and the feature map input to the second branch module is input to the deformable convolution network. The output of the deformable convolution network is connected to the feature map input to the second branch module from the direct connection structure and then output through a residual connection. Finally, a feature map with a feature dimension of 256 and a scale of H / 4*W / 4 is obtained. The introduction of deformable convolution can increase its receptive field while ensuring the spatial resolution of the feature map, which is beneficial to the detection of small targets and the capture of edge information of changing objects.

[0057] like Figure 7 As shown, the network layer 2 includes a maximum pooling layer, a multi-head self-attention bottleneck module and an upsampling module in sequence. The multi-head self-attention bottleneck module includes a self-attention bottleneck module 1 MHSABottleneck-1 and two self-attention bottleneck modules 2 MHSABottleneck-2 in sequence. Both the self-attention bottleneck module 1 and the self-attention bottleneck module 2 adopt a dual-branch structure.

[0058] like Figure 8a As shown, the first branch of the self-attention bottleneck module MHSABottleneck-1 includes a 1*1 convolution, and the second branch includes a multi-head self-attention module. The feature map input to the self-attention bottleneck module 1 is input to its first branch and second branch respectively, and the outputs of the first branch and the second branch are output after being connected through residuals;

[0059] like Figure 8b As shown, the first branch of the self-attention bottleneck module 2 MHSABottleneck-2 is a direct connection structure, the second branch includes a multi-head self-attention module, the feature map input to the self-attention bottleneck module 2 is input to the second branch, the first branch directly introduces the feature map input to the self-attention bottleneck module 2, and the output of the second branch is connected with the feature map from the first branch and the feature map input to the second branch module through a residual connection and then output.

[0060] Figure 8a and Figure 8b In it, x is the input feature map, is the output after processing by the MHSABottleneck module, and MHSA is the multi-head self-attention module.

[0061] The multi-head self-attention module MHSA (Multi-Head Self-Attention) is composed of 4 self-attention layers and a relative position encoding module. Its network structure is as follows: Fig. 9 As shown in the figure, this module uses 4 Self-Attention Layers. For simplicity, not all of them are shown in the figure. Among them, q, k, v, and r represent the query vector, the correlation vector between the queried information and other information, the vector and position encoding of the queried information, respectively. x is the input feature map of size H×W×d, z is the output feature map, and 1×1 represents 1×1 convolution.

[0062] At the same time, deformable convolution and multi-head self-attention mechanism are introduced to increase the receptive field while ensuring the spatial resolution of the feature map, which is conducive to the detection of small targets and the accurate capture of the boundary information of changing objects.

[0063] like Fig.10 As shown in Figure 1, the classification layer includes the first convolution layer, the first upsampling module, the second convolution layer, the channel attention module, the third convolution layer, the fourth convolution layer and the second upsampling module. The structure of the channel attention module is shown in Figure 1. Fig.11 As shown, in the figure, x is the input feature map, is the output after processing by the Channel Attention module, FC is the fully connected layer, and ReLU is the activation function.

[0064] Preferably, the model is trained by the following deep supervision method, so that the intermediate network layers (input layer, network layer 1, network layer 2) can also be effectively trained, and the network weights can be better updated, thereby avoiding gradient disappearance and improving the performance of the model:

[0065] S1. Subtract the feature maps of image block 1 and image block 2 processed by the input layer, network layer 1, network layer 2 and classification layer to form a new feature map containing change information;

[0066] S2. Perform 1*1 convolution on the new feature map so that its channel dimension is the number of classifications (1);

[0067] S3. Upsample the feature map processed by step S2 to the original input image size, and finally calculate its loss function in sequence and perform multiple back propagations.

[0068] Furthermore, in step S3, the loss is calculated using formula (1) for the prediction graphs obtained at the input layer, network layer 1, network layer 2, and classification layer, and the gradient is updated through back propagation. After completing the specified number of rounds of model training, the parameters of the model are saved and deployed to other devices to carry out prediction work:

[0069] L=L dice +L focal (1)

[0070] L focal (p t )=-α t (1-p t ) γ log(p t ) (2)

[0071]

[0072] L dice represents Dice Loss, L focal is Focal Loss, p t is the predicted probability, α and γ are adjustment parameters, ε is the smoothing factor, y is the predicted value, and t is the true value / label.

[0073] Focal Loss adds adjustment parameters (α and γ) to the cross entropy, so that the model can reduce the weights of easy negative samples and pay more attention to the learning of difficult and misclassified samples. Dice Loss focuses on mining positive samples and can effectively alleviate the problem of sample imbalance.

[0074] The model implemented by the above method has a small number of parameters and can be easily ported to other mobile devices. In addition, the model can be quickly trained with a larger batch size and has good generalization. These improvements are conducive to practical engineering applications and rapid deployment, and improve the efficiency and accuracy of the model in remote sensing image change detection.

[0075] The focus of this solution is to propose a change detection model that can extract high-precision difference images. The subsequent steps of change detection, such as processing the difference images to obtain change detection results, are consistent with the prior art and will not be repeated here.

[0076] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

[0077] Although this article uses more terms such as input layer, network layer 1, network layer 2, classification layer, convolution module, FMBConv module, ASPP module, first dual-branch module, second dual-branch module, deformable convolution, multi-head self-attention bottleneck module, etc., it does not exclude the possibility of using other terms. These terms are used only to more conveniently describe and explain the essence of the present invention; interpreting them as any additional limitations is contrary to the spirit of the present invention.

Claims

1. A method for constructing a change detection model for remote sensing image change detection, the model comprising two twin networks, each of which comprises an input layer, a network layer 1, a network layer 2 and a classification layer in sequence, the inputs of the two input layers are image blocks 1 and image blocks 2 respectively, and the detection model obtains a difference image of the two image blocks according to the outputs of the two classification layers, characterized in that: The input of network layer 1 is the output of the corresponding input layer, the input of network layer 2 is the result of concatenating the output of the corresponding input layer and the output of network layer 1 along the channel dimension, and the input of the classification layer is the result of concatenating the output of the corresponding input layer, the output of network layer 1, and the output of network layer 2 along the channel dimension. The input layer includes a convolution module, an FMBConv module and an ASPP module; the network layer 1 includes a first dual-branch module and a second dual-branch module, both of which adopt a dual-branch structure, and one of the branches of the two dual-branch modules includes a deformable convolution; the network layer 2 includes a multi-head self-attention bottleneck module; The FMBConv module includes a multiple convolution of a multiple of the input feature dimension of the FMBConv module and a same-dimensional convolution of the same dimension as the input feature dimension; the feature input to the FMBConv module is first input to the multiple convolution for convolution processing, the processing result of the multiple convolution is used as the input of the same-dimensional convolution, and the output of the same-dimensional convolution and the input feature of the FMBConv module are added as the output of the FMBConv module and input to the ASPP module; The ASPP module includes branch one, branch two, branch three, branch four and branch five, wherein branch one adopts 1*1 convolution, branch two adopts 3*3 dilated convolution with a dilation rate of 6, branch three adopts 3*3 dilated convolution with a dilation rate of 12, branch four adopts 3*3 dilated convolution with a dilation rate of 18, and branch five adopts an average pooling layer for global average pooling. The features input to the ASPP module are respectively input to the multiple branches, and finally the output features of the multiple branches are spliced ​​in the channel dimension and the number of output channels is adjusted using 1*1 convolution before output; The first branch of the first dual-branch module includes a 1*1 convolution, and the second branch includes a deformable convolution. The feature map input to the first dual-branch module is input to the 1*1 convolution and the deformable convolution respectively, and the outputs of the 1*1 convolution and the deformable convolution are connected through a residual network and then output; The first branch of the second dual-branch module is a direct connection structure, the second branch includes a deformable convolution, the feature map input to the second branch module is input to the deformable convolution network, the direct connection structure directly introduces the feature map input to the second branch module, and the output of the deformable convolution network is connected with the feature map from the direct connection structure through a residual network and then output; The multi-head self-attention bottleneck module includes one self-attention bottleneck module 1 and two self-attention bottleneck modules 2 in sequence; And both the self-attention bottleneck module 1 and the self-attention bottleneck module 2 adopt a dual-branch structure; The first branch of the self-attention bottleneck module 1 includes a 1*1 convolution, and the second branch includes a multi-head self-attention module. The feature map input to the self-attention bottleneck module 1 is input to its first branch and second branch respectively, and the outputs of the first branch and the second branch are connected through a residual network and then output; The first branch of the self-attention bottleneck module 2 is a direct connection structure, and the second branch includes a multi-head self-attention module. The feature map input to the self-attention bottleneck module 2 is input to the second branch. The first branch directly introduces the feature map input to the self-attention bottleneck module 2, and the output of the second branch is connected to the feature map from the first branch through a residual network and then output.

2. The method for constructing a change detection model for remote sensing image change detection according to claim 1, characterized in that: The multi-head self-attention modules each include four self-attention layers and a relative position encoding module.

3. The method for constructing a change detection model for remote sensing image change detection according to claim 1 or 2, characterized in that: The input layer includes a convolution module, two FMBConv modules, a maximum pooling layer and an ASPP module in sequence; The network layer 1 includes a first dual-branch module and two second dual-branch modules in sequence; The network layer 2 includes a maximum pooling layer, a multi-head self-attention bottleneck module and an upsampling module in sequence.

4. The method for constructing a change detection model for remote sensing image change detection according to claim 1 or 2, characterized in that: The classification layer includes a first convolutional layer, a first upsampling module, a second convolutional layer, a channel attention module, a third convolutional layer, a fourth convolutional layer and a second upsampling module in sequence.

5. The method for constructing a change detection model for remote sensing image change detection according to claim 1 or 2, characterized in that: The model is trained using the following deep supervision: S1. Subtract the feature maps of image block 1 and image block 2 processed by the input layer, network layer 1, network layer 2 and classification layer respectively to form a new feature map containing change information; S2. Perform 1*1 convolution on the new feature map so that its channel dimension is the number of classifications; S3. Upsample the feature map processed by step S2 to the original input image size, and finally calculate its loss function in sequence and perform multiple back propagations.

Images