Self-adaptive edge-guided context aggregation remote sensing image change detection method

Abstract

The invention discloses a self-adaptive edge-guided context aggregation remote sensing image change detection method, which belongs to the technical field of remote sensing and comprises the following steps of: extracting dual-temporal multi-scale features of a remote sensing image to be detected based on Res2Net; performing difference information enhancement between the foreground and the background on the dual-time-phase multi-scale features by using a difference enhancement module group to obtain enhanced multi-scale features; respectively inputting the enhanced multi-scale features into a boundary extractor and a Transform module, and carrying out context information aggregation under the guidance of boundary constraint to obtain a multi-scale feature vector; and inputting the multi-scale feature vector into a classifier based on the multi-scale entropy and Dice mixed loss function to obtain a remote sensing image change detection result. According to the invention, the problem that the existing remote sensing image change detection method is not high in detection precision in a complex scene influenced by the illumination season change is solved.

Description

Technical Field

[0001] The invention belongs to the field of remote sensing technology, and in particular relates to an adaptive edge-guided context-aggregated remote sensing image change detection method. Background Art

[0002] Change detection of remote sensing images has a wide range of application needs in the fields of ecological environment monitoring, disaster management, agricultural production, environmental protection and military reconnaissance. Traditional image change detection methods are mainly based on low- and medium-resolution remote sensing images. Starting from the spectral characteristics of the image itself, methods such as algebraic calculation, image transformation and image classification are used to analyze the distribution of changes in the image. However, with the rapid development of remote sensing sensors and the continuous expansion of various high-resolution remote sensing images, traditional methods are difficult to effectively migrate and apply, which leads to reduced detection accuracy, a high degree of missed detection and false detection, and cannot meet current needs in terms of accuracy and degree of automation.

[0003] The development of deep learning has provided new technical methods for change detection. With its powerful feature extraction capability, it has become an important means of change detection in remote sensing images. Existing change detection methods based on deep learning mainly deepen the network depth by stacking convolutional layers and gradually expand the receptive field to obtain more local correlation features, or introduce attention mechanisms. By using different types of attention modules, the ability of CNN in extracting global feature vectors is further enhanced, thereby improving the effect of change detection. Although stacking convolutional blocks and using attention mechanisms have improved the extraction of global feature vectors, the ability to obtain global representation is still insufficient. In complex scenes affected by seasonal changes in light, the detection accuracy is not high. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides an adaptive edge-guided context-aggregated remote sensing image change detection method, which adopts a CNN-Transformer hybrid structure, first performs feature extraction and difference feature enhancement, then uses boundary information to guide context aggregation, and finally adopts a multi-scale output fusion strategy to fuse global features and local features, and outputs remote sensing image change detection results, thereby solving the problem of low detection accuracy of existing remote sensing image change detection methods in complex scenes affected by seasonal changes in light.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0006] The present invention provides an adaptive edge-guided context-aggregated remote sensing image change detection method, comprising the following steps:

[0007] S1. Extract dual-temporal and multi-scale features of remote sensing images to be detected based on Res2Net;

[0008] S2, using the difference enhancement module group to enhance the difference information between the foreground and the background of the dual-phase multi-scale features to obtain enhanced multi-scale features;

[0009] S3, input the enhanced multi-scale features into the boundary extractor and Transformer modules respectively, aggregate the context information under the guidance of boundary constraints, and obtain the multi-scale feature vector;

[0010] S4. Based on the multi-scale entropy and Dice hybrid loss function, the multi-scale feature vector is input into the classifier to obtain the remote sensing image change detection result.

[0011] The beneficial effects of the present invention are as follows: the present invention provides an adaptive edge-guided context-aggregated remote sensing image change detection method, which uses Res2Net for feature extraction, and replaces a single 3×3 convolution with a layered and stacked feature group in a residual block, thereby improving the ability to extract multi-scale features; the present invention adopts a difference enhancement module group, recalibrates the extracted multi-scale features, reduces noise interference, and enables the enhanced multi-scale features to focus on objects that actually change, so as to enhance the remote sensing image change detection effect; the present invention designs an adaptive boundary extractor based on deformable convolution, which can correspond to objects of different scales or deformations at different positions, and the adaptive determination of multi-scale or receptive field sizes provides a basis for fine positioning of remote sensing image deformation positions; the present invention adopts a hybrid structure of CNN-Transformer hybrid series connection, which combines the advantages of the two, is conducive to the interaction of long-distance image information, and reduces the computational complexity, and can still achieve remote sensing image change detection with high detection accuracy in complex scenes affected by seasonal changes in light.

[0012] Furthermore, the S1 comprises the following steps:

[0013] S11, obtaining three-channel remote sensing images to be detected;

[0014] S12. Extract the features of the remote sensing image to be detected based on Res2Net to obtain the dual-phase multi-scale features, where the dual-phase multi-scale features include the dual-phase second scale feature F2 outn , dual-phase third scale feature F3 outn and the dual-phase fourth scale feature F4 outn , when n=1, it is the first phase, and when n=2, it is the second phase.

[0015] The beneficial effects of adopting the above further scheme are as follows: based on Res2Net, the present invention uses a hierarchical and stacked feature group to replace the same single 3*3 convolution in a given residual block, thereby improving the network's ability to extract multi-scale features. After res2net extracts features, the present invention introduces a dilated spatial convolution pooling pyramid to expand the receptive field of the feature map and obtain more global context information, thereby providing a basis for further performing feature difference enhancement.

[0016] Furthermore, the S2 comprises the following steps:

[0017] S21, based on the difference enhancement module, performing subtraction on the dual-phase multi-scale features to obtain multi-scale initial difference features;

[0018] The calculation expression of the multi-scale initial difference feature is as follows:

[0019] D i =|Fi out1 -Fi out2 |

[0020] Among them, D i represents the multi-scale initial difference feature, |·| represents the absolute value, Fi out1 Represents the multi-scale features of the first phase, Fi out2 represents the multi-scale features of the second phase, where i=2, 3, 4, i=2 represents the second scale, i=3 represents the third scale, and i=4 represents the fourth scale;

[0021] S22, performing average pooling and maximum pooling on the multi-scale initial difference features respectively to obtain an average pooling result and a maximum pooling result;

[0022] S23, respectively extracting the features of the average pooling result and the maximum pooling result, and inputting them into a multi-layer perceptron for channel dimension feature learning, thereby extracting the average pooling channel dimension feature learning result and the maximum pooling channel dimension feature learning result;

[0023] S24, normalizing the average pooling channel dimension feature learning result and the maximum pooling channel dimension feature learning result to obtain the attention weight;

[0024] The calculation expression of the attention weight is as follows:

[0025] att i =σ(MLP(Avgpool(D i )+MLP(Maxpool(D i ))))

[0026] Among them, att irepresents the attention weight, σ(·) represents the sigmoid function, MLP(·) represents the multi-layer perceptron, Avgpool(·) represents the average pooling, and Maxpool(·) represents the maximum pooling;

[0027] S25. Based on the attention weights, construct a residual connection for the bi-phase multi-scale features to obtain enhanced multi-scale features, wherein the enhanced multi-scale features include first-scale difference enhanced features, second-scale difference enhanced features, and third-scale difference enhanced features;

[0028] The calculation expression of the enhanced multi-scale feature is as follows:

[0029] Out n =Fi outn *att i +Fi outn

[0030] Among them, Out n Represents the enhanced multi-scale features, Fi outn Represents dual-phase multi-scale features, where n=1, 2, n=1 represents the first phase, and n=2 represents the second phase.

[0031] The beneficial effects of adopting the above further scheme are as follows: the present invention provides a method for enhancing the difference information between the foreground and the background of the dual-phase multi-scale features by using the difference enhancement module group, and recalibrates the obtained dual-phase multi-scale features to reduce noise interference, focusing on the objects that have actually changed, to enhance the change detection effect. The present invention uses residual connections to strengthen the important information of the multi-scale feature map, thereby effectively overcoming the problem that the image is affected by light and season, and there are serious spatial-spectral differences between different phases.

[0032] Furthermore, S3 comprises the following steps:

[0033] S31, using a boundary extractor to extract boundary features of the enhanced multi-scale features to obtain boundary features of the multi-scale features, wherein the boundary features of the multi-scale features include boundary features of the first scale difference enhanced features, boundary features of the second scale difference enhanced features, and boundary features of the third scale difference enhanced features;

[0034] S32, inputting the third scale difference enhancement feature into the encoder of the Transformer module for encoding to obtain a first global semantic feature;

[0035] S33, using the boundary feature of the third scale difference enhancement feature as a boundary constraint guide, and using the decoder of the Transformer module to decode the first global semantic feature to obtain a first global feature vector;

[0036] S34, inputting the first global feature vector and the second scale difference enhancement feature into the encoder of the Transformer module for encoding to obtain a second global semantic feature;

[0037] S35, using the boundary feature of the second scale difference enhancement feature as a boundary constraint guide, and using the decoder of the Transformer module to decode the second global semantic feature to obtain a second global feature vector;

[0038] S36, inputting the second global feature vector and the first scale difference enhancement feature into an encoder of the Transformer module for encoding to obtain a third global semantic feature;

[0039] S37, using the boundary feature of the first scale difference enhancement feature as a boundary constraint guide, and using the decoder of the Transformer module to decode the third global semantic feature to obtain a third global feature vector;

[0040] S38, based on a multi-scale feature fusion method, concatenate and fuse the first global feature vector, the second global feature vector, and the third global feature vector to obtain a fourth global feature vector;

[0041] S39: Use the first global feature vector, the second global feature vector, the third global feature vector, and the fourth global feature vector as multi-scale feature vectors.

[0042] The beneficial effect of adopting the above further scheme is: in the context aggregation stage, the present invention utilizes the intrinsic distinguishing ability of the boundary so that pixels of the same category can obtain similar gains, and guides the Transformer module through boundary features to achieve context information aggregation, thereby obtaining a multi-scale feature vector that integrates more information.

[0043] Furthermore, the S31 includes the following steps:

[0044] S311, based on the boundary extractor, performing 3*3 convolution filtering processing on the enhanced multi-scale features to obtain a first multi-scale convolution filtering result;

[0045] The calculation expression of the first multi-scale convolution filtering result is as follows:

[0046]

[0047] Among them, y(p 0 ) represents the first multi-scale convolution filtering result, pn represents the enumeration in the convolution receptive field, ∈ represents belonging, R represents the convolution kernel receptive field, w(p n ) represents the filter window, x(p 0 +p n) represents 3*3 convolution input, p 0 Indicates the location where convolution filtering is performed in multi-scale features;

[0048] S312, based on the boundary extractor, performing deformable convolution filtering processing on the enhanced multi-scale features to obtain a second multi-scale convolution filtering result;

[0049] The calculation expression of the second multi-scale convolution filtering result is as follows:

[0050]

[0051] Among them, y(p 0 )′ represents the second multi-scale convolution filtering result, x(p 0 +p n +Δp n ) represents the deformable convolution filter input, Δp n Indicates the predicted offset;

[0052] S313, performing batch normalization processing on the first multi-scale convolution filtering result and the second multi-scale convolution filtering result respectively to obtain normalized results of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result;

[0053] S314, reconstructing the normalized results of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result to obtain the reconstructed first multi-scale convolution filtering result and the second multi-scale convolution filtering result;

[0054] The calculation expressions of the reconstructed first multi-scale convolution filtering result and the second multi-scale convolution filtering result are as follows:

[0055] y(p 0 ) rec =γy(p 0 ) Nor +β′

[0056] y(p 0 )′ rec =γy(p 0 )′ Nor +β′

[0057] Among them, y(p 0 ) rec represents the first multi-scale convolution filtering result after reconstruction, y(p 0 )′ rec represents the reconstructed second multi-scale convolution filtering result, γ represents the first reconstruction learning parameter, and β′ represents the second reconstruction learning parameter;

[0058] S315, performing nonlinear activation on the reconstructed first multi-scale convolution filtering result and the second multi-scale convolution filtering result, and summing them to obtain boundary features of multi-scale features, wherein the boundary features of the multi-scale features include boundary features of the first scale difference enhancement feature, boundary features of the second scale difference enhancement feature, and boundary features of the third scale difference enhancement feature;

[0059] The calculation expression of the boundary feature of the multi-scale feature is as follows:

[0060] Fside=max(0,y(p 0 ) rec )+max(0,y(p 0 )′ rec )

[0061] = {Fside 1 ,Fside 2 ,Fside 3}

[0062] Among them, Fside represents the boundary feature of multi-scale features, max(0,y(p 0 ) rec ) represents the nonlinear activation result of the first multi-scale convolutional filtering result after reconstruction, max(0,y(p 0 )′ rec ) represents the nonlinear activation result of the reconstructed second multi-scale convolution filtering result, Fside 1 Fside represents the boundary feature of the first scale difference enhancement feature. 2 Fside represents the boundary feature of the second scale difference enhancement feature. 3 Represents boundary features of third-scale difference-enhanced features.

[0063] The beneficial effect of adopting the above further scheme is as follows: the present invention introduces deformable convolution when extracting boundary features, and uses deformable convolution to adaptively and finely locate feature objects of different scales, thereby realizing adaptive perception of boundary features. The present invention adds the two-branch features obtained by deformable convolution, normalization, and activation function to obtain edge features, which provides a basis for decoding by the Transformer module.

[0064] Furthermore, the S313 includes the following steps:

[0065] S3131, respectively calculating the mean of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result;

[0066] The calculation expression of the mean of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result is as follows:

[0067]

[0068] Among them, μ β represents the mean of the first multi-scale convolutional filtering result, m represents the size of the training sample Batch, i′ represents the i′th Batch, μ β ′ represents the mean value of the second multi-scale convolution filtering result;

[0069] S3132. Calculate the standard deviation of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result based on the mean of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result;

[0070] The calculation expression of the standard deviation of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result is as follows:

[0071]

[0072] Among them, σ β represents the standard deviation of the first multi-scale convolution filtering result, σ β ′ represents the standard deviation of the second multi-scale convolution filtering result;

[0073] S3133. Based on the mean of the first multiscale convolution filtering result and the second multiscale convolution filtering result, and the standard deviation of the first multiscale convolution filtering result and the second multiscale convolution filtering result, respectively calculate the normalized results of the first multiscale convolution filtering result and the second multiscale convolution filtering result;

[0074] The calculation expression of the normalized result of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result is as follows:

[0075]

[0076] Among them, y(p 0 ) Nor represents the normalized result of the first multi-scale convolution filtering result, y(p 0 )′ Nor represents the normalized result of the second multi-scale convolution filtering result, and ε represents a constant close to 0.

[0077] Furthermore, the calculation expression of the Transformer module is as follows:

[0078]

[0079] Among them, Attention(Q,K,V) represents the attention mechanism in the Transformer module, Q represents the query vector, K represents the key vector, T represents the transpose, V represents the value vector, and dk Represents the length of the key vector; the encoding stage of the Transformer module divides the encoder input into several fixed-size image blocks, and adds position encoding to each image block. The correlation between different image blocks is calculated through the multi-head attention mechanism to obtain global semantic features; the decoding stage of the Transformer module inputs multi-scale boundary features into the decoder, guides the fusion of global semantic features and spatial features of different scales, and completes context information aggregation.

[0080] The beneficial effect of adopting the above further scheme is: the present invention adds position encoding through the Transformer module, strengthens the spatial connection between image blocks, and then calculates the correlation between different image blocks based on the multi-head attention mechanism, so as to obtain rich global semantic features.

[0081] Furthermore, the calculation expression of the multi-scale entropy and Dice mixed loss function is as follows:

[0082] Loss = L BCE +L Dice

[0083] L BCE =-[y j logp j +(1-y j )log(1-p j )]

[0084]

[0085] Among them, LOSS represents the multi-scale entropy and Dice mixed loss function, L BCE represents the multi-scale entropy loss function, L Dice represents the Dice loss function, y j represents the true label value of pixel j in the remote sensing image to be detected, p j Represents the predicted probability of a change in pixel j in the remote sensing image to be detected.

[0086] The beneficial effect of adopting the above further scheme is: the binary cross entropy loss function adopted by the present invention can be applicable to the binary classification task of remote sensing image change detection, and at the same time, combined with the Dice loss function, it can improve the segmentation performance and reduce the impact of sample category imbalance to measure the similarity between the predicted value and the true value.

[0087] Other advantages of the present invention will be analyzed in more detail in subsequent embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0089] Figure 1 The present invention is a flowchart of the steps of an adaptive edge-guided context-aggregated remote sensing image change detection method in an embodiment of the present invention.

[0090] Figure 2 Schematic diagram of the network structure corresponding to the remote sensing image change detection method of adaptive edge-guided context aggregation in an embodiment of the present invention.

[0091] Figure 3 Schematic diagram of the structure of the difference enhancement module in the embodiment of the present invention.

[0092] Figure 4 This is an image sample in the CLCD farmland change detection dataset in an embodiment of the present invention.

[0093] Figure 5 This is an image sample in the RSCD farmland change detection dataset in an embodiment of the present invention. DETAILED DESCRIPTION

[0094] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present invention.

[0095] like Figure 1 As shown, in one embodiment of the present invention, the present invention provides an adaptive edge-guided context-aggregated remote sensing image change detection method, comprising the following steps:

[0096] S1. Extract dual-temporal and multi-scale features of remote sensing images to be detected based on Res2Net;

[0097] The S1 comprises the following steps:

[0098] S11, obtaining three-channel remote sensing images to be detected;

[0099] S12. Extract the features of the remote sensing image to be detected based on Res2Net to obtain the dual-phase multi-scale features, where the dual-phase multi-scale features include the dual-phase second scale feature F2 outn , dual-phase third scale feature F3 outn and the dual-phase fourth scale feature F4 outn , when n=1, it is the first phase, and when n=2, it is the second phase.

[0100] In this embodiment, the dual-phase second scale feature F2 outn Including the first phase second scale feature F2 out1 and the second scale feature F2 of the second phase out2 , dual-phase third scale feature F3 outn Including the first phase third scale feature F3 out1 and the third scale feature F3 of the second phase out2 , dual-phase fourth scale feature F4 outn Including the fourth scale feature F4 of the first phase out1 and the third scale feature F4 of the second phase out2 ;

[0101] like Figure 2 As shown, in this embodiment, Res2Net includes a first sub-network, a second sub-network, and a first difference enhancement module, a second difference enhancement module, and a third difference enhancement module respectively connected to the first sub-network and the second sub-network, wherein the first sub-network and the second sub-network constitute a twin network; the first sub-network and the second sub-network each include a first convolutional layer, a first residual block, a second residual block, a third residual block, and a fourth residual block connected in sequence;

[0102] The first convolution layer in the first sub-network is used as the image input layer of the first sub-network to input the features of the remote sensing image to be detected; the features of the remote sensing image to be detected are sequentially extracted by the first convolution layer, the first residual block, the second residual block, the third residual block and the fourth residual block in the first sub-network, and the first phase first scale features F1 are obtained from the outputs of the first residual block, the second residual block, the third residual block and the fourth residual block. out1 , the first phase second scale feature F2 out1 , the third scale feature of the first phase F3 out1 and the fourth scale feature F4 of the first phase out1 Similarly, after the remote sensing image to be detected is input into the first convolutional layer of the second sub-network, after multi-scale feature extraction, four levels of multi-scale features can be obtained, namely, the first scale feature F1 of the second phase out2 , the second phase second scale feature F2 out2 , the third scale feature of the second phase F3out2 and the fourth scale feature F4 of the second phase out2 ; The first phase first scale feature F1 out1 and the first scale feature F1 of the second phase out2 The size of the remote sensing image to be detected is 1 / 2 of the size, the number of channels is 32, and the first phase second scale feature F2 out1 and the second scale feature F2 of the second phase out2 The size of the remote sensing image to be detected is 1 / 4 of the size, the number of channels is 32, and the first phase third scale feature F3 out1 and the third scale feature F3 of the second phase out2 The size of the remote sensing image to be detected is 1 / 8 of the size, the number of channels is 32, and the first phase fourth scale feature F4 out1 and the fourth scale feature F4 of the second phase out2 The size of is 1 / 16 of the size of the remote sensing image to be detected, and the number of channels is 32 dimensions. Since the feature size of the shallow position is large, although the spatial information is rich but the semantic information is weak, and the feature size of the deep position of the network is small, the spatial detail information is weak but the semantic information is rich, so the dual-phase second scale feature F2 is preferred. outn , dual-phase third scale feature F3 outn and the dual-phase fourth scale feature F4 outn For feature aggregation and change prediction. In this embodiment, in order to facilitate subsequent feature fusion, the extracted features of different levels are reduced in dimension through 1*1 convolution, and the different channel numbers are unified into 32 dimensions, which can reduce feature redundancy and reduce the amount of calculation to a certain extent.

[0103] S2, using the difference enhancement module group to enhance the difference information between the foreground and the background of the dual-phase multi-scale features to obtain enhanced multi-scale features;

[0104] In this embodiment, the enhanced multi-scale features include first scale difference enhancement features, second scale difference enhancement features, and third scale difference enhancement features. The first scale difference enhancement features include first time phase first scale difference enhancement features and second time phase first scale difference enhancement features, the second scale difference enhancement features include first time phase second scale difference enhancement features and second time phase second scale difference enhancement features, and the third scale difference enhancement features include first time phase third scale difference enhancement features and second time phase third scale difference enhancement features.

[0105] like Figure 3 As shown, in this embodiment, the difference enhancement module group includes a first difference enhancement module, a second difference enhancement module and a third difference enhancement module; the first difference enhancement module is used to enhance the second scale feature F2 of the dual-phase outnThe second difference enhancement module is used to enhance the difference information. outn To enhance the difference information, the third difference enhancement module is used to enhance the fourth scale feature F4 of the dual time phase outn Perform difference information enhancement;

[0106] The first difference enhancement module, the second difference enhancement module and the third difference enhancement module each include a first time phase input submodule, a second time phase input submodule, a difference submodule, an average pooling submodule, a maximum pooling submodule, four second convolutional layers, two multi-layer perceptrons, an attention submodule, a first residual connection submodule, a second residual connection submodule, a first feature output submodule and a second feature output submodule;

[0107] The first time phase input submodule and the second time phase input submodule are used to input dual-time phase multi-scale features, wherein the first time phase second scale feature F2 out1 , the third scale feature of the first phase F3 out1 and the dual-phase fourth scale feature F4 outn From the first phase input submodule to the corresponding difference enhancement module, the second phase second scale feature F2 out2 , the third scale feature of the second phase F3 out2 and the third scale feature F4 of the second phase out2Input from the second time phase input submodule to the corresponding difference enhancement module; the difference submodule is used to perform subtraction difference on the input dual-phase multi-scale features to obtain multi-scale initial difference features; the average pooling submodule is used to perform average pooling on the multi-scale initial difference features to obtain average pooling results; the maximum pooling submodule is used to perform maximum pooling on the multi-scale initial difference features to obtain maximum pooling results; the first second convolutional layer is used to extract the features of the average pooling result and input it into the first multi-layer perceptron, the second second convolutional layer is used to extract the features of the maximum pooling result and input it into the second multi-layer perceptron, the first multi-layer perceptron is used to learn the channel dimension features of the average pooling result, the second multi-layer perceptron is used to learn the channel dimension features of the maximum pooling result, the third second convolutional layer is used to extract the average pooling channel dimension feature learning result output by the first multi-layer perceptron, and the fourth second convolutional layer is used to extract the maximum pooling channel dimension feature learning result output by the second multi-layer perceptron; the attention submodule is used to perform average pooling on the average pooling channel. The channel dimension feature learning results and the maximum pooling channel dimension feature learning results are normalized to obtain the attention weight; the first residual connection submodule is used to multiply the first phase multi-scale feature of the input dual-phase multi-scale feature by the attention weight, and then add its corresponding first phase multi-scale feature to obtain the first phase difference enhancement feature; the second residual connection submodule is used to multiply the second phase multi-scale feature of the input dual-phase multi-scale feature by the attention weight, and then add its corresponding second phase multi-scale feature to obtain the second phase difference enhancement feature; the first feature output submodule is used to output the first phase difference enhancement feature; the second feature output submodule is used to output the second phase difference enhancement feature; the first phase difference enhancement feature includes the first phase first scale difference enhancement feature, the first phase second scale difference enhancement feature and the first phase third scale difference enhancement feature; the second phase difference enhancement feature includes the second phase first scale difference enhancement feature, the second phase second scale difference enhancement feature and the second phase third scale difference enhancement feature.

[0108] The S2 comprises the following steps:

[0109] S21, based on the difference enhancement module, performing subtraction on the dual-phase multi-scale features to obtain multi-scale initial difference features;

[0110] The calculation expression of the multi-scale initial difference feature is as follows:

[0111] D i =|Fi out1 -Fi out2 |

[0112] Among them, D i represents the multi-scale initial difference feature, |·| represents the absolute value, Fiout1 Represents the multi-scale features of the first phase, Fi out2 represents the multi-scale features of the second phase, where i=2, 3, 4, i=2 represents the second scale, i=3 represents the third scale, and i=4 represents the fourth scale;

[0113] S22, performing average pooling and maximum pooling on the multi-scale initial difference features respectively to obtain an average pooling result and a maximum pooling result;

[0114] S23, respectively extracting the features of the average pooling result and the maximum pooling result, and inputting them into a multi-layer perceptron for channel dimension feature learning, thereby extracting the average pooling channel dimension feature learning result and the maximum pooling channel dimension feature learning result;

[0115] S24, normalizing the average pooling channel dimension feature learning result and the maximum pooling channel dimension feature learning result to obtain the attention weight;

[0116] The calculation expression of the attention weight is as follows:

[0117] att i =σ(MLP(Avgpool(D i )+MLP(Maxpool(D i ))))

[0118] Among them, att i represents the attention weight, σ(·) represents the sigmoid function, MLP(·) represents the multi-layer perceptron, Avgpool(·) represents the average pooling, and Maxpool(·) represents the maximum pooling;

[0119] S25. Based on the attention weights, construct a residual connection for the bi-phase multi-scale features to obtain enhanced multi-scale features, wherein the enhanced multi-scale features include first-scale difference enhanced features, second-scale difference enhanced features, and third-scale difference enhanced features;

[0120] The calculation expression of the enhanced multi-scale feature is as follows:

[0121] Out n =Fi outn *att i +Fi outn

[0122] Among them, Out n Represents the enhanced multi-scale features, Fi outn Represents dual-phase multi-scale features, where n=1, 2, n=1 represents the first phase, and n=2 represents the second phase.

[0123] S3, input the enhanced multi-scale features into the boundary extractor and Transformer modules respectively, aggregate the context information under the guidance of boundary constraints, and obtain the multi-scale feature vector;

[0124] The S3 comprises the following steps:

[0125] S31, using a boundary extractor to extract boundary features of the enhanced multi-scale features to obtain boundary features of the multi-scale features, wherein the boundary features of the multi-scale features include boundary features of the first scale difference enhanced features, boundary features of the second scale difference enhanced features, and boundary features of the third scale difference enhanced features;

[0126] The S31 comprises the following steps:

[0127] S311, based on the boundary extractor, performing 3*3 convolution filtering processing on the enhanced multi-scale features to obtain a first multi-scale convolution filtering result;

[0128] The calculation expression of the first multi-scale convolution filtering result is as follows:

[0129]

[0130] Among them, y(p 0 ) represents the first multi-scale convolution filtering result, p n represents the enumeration in the convolution receptive field, ∈ represents belonging, R represents the convolution kernel receptive field, w(p n ) represents the filter window, x(p 0 +p n ) represents 3*3 convolution input, p 0 Indicates the location where convolution filtering is performed in multi-scale features;

[0131] S312, based on the boundary extractor, performing deformable convolution filtering processing on the enhanced multi-scale features to obtain a second multi-scale convolution filtering result;

[0132] The calculation expression of the second multi-scale convolution filtering result is as follows:

[0133]

[0134] Among them, y(p 0 )′ represents the second multi-scale convolution filtering result, x(p 0 +p n +Δp n ) represents the deformable convolution filter input, Δp n Indicates the predicted offset;

[0135] S313, performing batch normalization processing on the first multi-scale convolution filtering result and the second multi-scale convolution filtering result respectively to obtain normalized results of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result;

[0136] The S313 comprises the following steps:

[0137] S3131, respectively calculating the mean of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result;

[0138] The calculation expression of the mean of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result is as follows:

[0139]

[0140]

[0141] Among them, μ β represents the mean of the first multi-scale convolutional filtering result, m represents the size of the training sample Batch, i′ represents the i′th Batch, μ β ′ represents the mean value of the second multi-scale convolution filtering result;

[0142] S3132. Calculate the standard deviation of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result based on the mean of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result;

[0143] The calculation expression of the standard deviation of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result is as follows:

[0144]

[0145] Among them, σ β represents the standard deviation of the first multi-scale convolution filtering result, σ β ′ represents the standard deviation of the second multi-scale convolution filtering result;

[0146] S3133. Based on the mean of the first multiscale convolution filtering result and the second multiscale convolution filtering result, and the standard deviation of the first multiscale convolution filtering result and the second multiscale convolution filtering result, respectively calculate the normalized results of the first multiscale convolution filtering result and the second multiscale convolution filtering result;

[0147] The calculation expression of the normalized result of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result is as follows:

[0148]

[0149] Among them, y(p0 ) Nor represents the normalized result of the first multi-scale convolution filtering result, y(p 0 )′ Nor represents the normalized result of the second multi-scale convolution filtering result, and ε represents a constant close to 0. In this embodiment, ε is a constant close to 0 set to avoid the denominator being 0, and can be set to one ten-thousandth.

[0150] S314, reconstructing the normalized results of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result to obtain the reconstructed first multi-scale convolution filtering result and the second multi-scale convolution filtering result;

[0151] The calculation expressions of the reconstructed first multi-scale convolution filtering result and the second multi-scale convolution filtering result are as follows:

[0152] y(p 0 ) rec =γy(p 0 ) Nor +β′

[0153] y(p 0 )′ rec =γy(p 0 )′ Nor +β′

[0154] Among them, y(p 0 ) rec represents the first multi-scale convolution filtering result after reconstruction, y(p 0 )′ rec represents the reconstructed second multi-scale convolution filtering result, γ represents the first reconstruction learning parameter, and β′ represents the second reconstruction learning parameter;

[0155] S315, performing nonlinear activation on the reconstructed first multi-scale convolution filtering result and the second multi-scale convolution filtering result, and summing them to obtain boundary features of the multi-scale features, wherein the boundary features of the multi-scale features include boundary features C1 of the first scale difference enhancement feature, boundary features C2 of the second scale difference enhancement feature, and boundary features C3 of the third scale difference enhancement feature;

[0156] The calculation expression of the boundary feature of the multi-scale feature is as follows:

[0157] Fside=max(0,y(p 0 ) rec )+max(0,y(p 0 )′ rec )

[0158] = {Fside 1 ,Fside2 ,Fside 3}

[0159] Among them, Fside represents the boundary feature of multi-scale features, max(0,y(p 0 ) rec ) represents the nonlinear activation result of the first multi-scale convolutional filtering result after reconstruction, max(0,y(p 0 )′ rec ) represents the nonlinear activation result of the reconstructed second multi-scale convolution filtering result, Fside 1 Fside represents the boundary feature of the first scale difference enhancement feature. 2 Fside represents the boundary feature of the second scale difference enhancement feature. 3 Represents boundary features of third-scale difference-enhanced features.

[0160] S32, inputting the third scale difference enhancement feature C3 into the encoder of the Transformer module for encoding to obtain a first global semantic feature;

[0161] S33, using the boundary feature of the third scale difference enhancement feature C3 as a boundary constraint guide, and using the decoder of the Transformer module to decode the first global semantic feature to obtain a first global feature vector V1;

[0162] S34, inputting the first global feature vector V1 and the second scale difference enhanced feature C2 into the encoder of the Transformer module for encoding to obtain a second global semantic feature;

[0163] S35, using the boundary feature of the second scale difference enhancement feature C2 as a boundary constraint guide, and using the decoder of the Transformer module to decode the second global semantic feature to obtain a second global feature vector V2;

[0164] S36, inputting the second global feature vector V2 and the first scale difference enhanced feature C1 into the encoder of the Transformer module for encoding to obtain a third global semantic feature;

[0165] S37, using the boundary feature of the first scale difference enhancement feature C1 as a boundary constraint guide, and using the decoder of the Transformer module to decode the third global semantic feature to obtain a third global feature vector V3;

[0166] The calculation expression of the Transformer module is as follows:

[0167]

[0168] Among them, Attention(Q,K,V) represents the attention mechanism in the Transformer module, Q represents the query vector, K represents the key vector, T represents the transpose, V represents the value vector, and d k Represents the length of the key vector; the encoding stage of the Transformer module divides the encoder input into several fixed-size image blocks, and adds position encoding to each image block. The correlation between different image blocks is calculated through the multi-head attention mechanism to obtain global semantic features; the decoding stage of the Transformer module inputs multi-scale boundary features into the decoder, guides the fusion of global semantic features and spatial features of different scales, and completes context information aggregation.

[0169] S38, based on a multi-scale feature fusion method, concatenate and fuse the first global feature vector V1, the second global feature vector V2 and the third global feature vector V3 to obtain a fourth global feature vector V4;

[0170] S39, taking the first global feature vector V1, the second global feature vector V2, the third global feature vector V3 and the fourth global feature vector V4 as multi-scale feature vectors.

[0171] S4. Based on the multi-scale entropy and Dice hybrid loss function, the multi-scale feature vector is input into the classifier to obtain the remote sensing image change detection result.

[0172] The calculation expression of the multi-scale entropy and Dice mixed loss function is as follows:

[0173] Loss = L BCE +L Dice

[0174] L BCE =-[y j logp j +(1-y j )log(1-p j )]

[0175]

[0176] Among them, LOSS represents the multi-scale entropy and Dice mixed loss function, L BCE represents the multi-scale entropy loss function, L Dice represents the Dice loss function, y j represents the true label value of pixel j in the remote sensing image to be detected, p j Represents the predicted probability of a change in pixel j in the remote sensing image to be detected.

[0177] In a practical example of the present invention, an adaptive edge-guided context-aggregated remote sensing image change detection method proposed in the present invention was experimentally evaluated on two land change datasets, and the two land change datasets were the CLCD cultivated land change detection dataset and the RSCD cultivated land change detection dataset; the CLCD cultivated land change detection dataset consists of 600 pairs of cultivated land change sample images, of which 360 pairs are used for training, 120 pairs are used for verification, and 120 pairs are used for testing. The dual-phase images in the CLCD cultivated land change detection dataset were collected by the Gaofen-2 satellite in 2017 and 2019, respectively, with a spatial resolution ranging from 0.5 to 2 meters. Each group of samples consists of two 512×512 images and corresponding cultivated land change binary labels. Figure 4 As shown in the figure, the types of changes marked in the CLCD farmland change detection dataset mainly include buildings, roads, lakes, and bare land. The RSCD farmland change detection dataset is a public dataset provided by the 2022 Aerospace Hongtu Cup Remote Sensing Image Intelligent Processing Algorithm Competition. It consists of 6,000 images from 1-meter-2-meter resolution GF-1 and GF-2 images, of which 2,052 pairs are used for training, 638 pairs are used for verification, and 203 pairs are used for testing. Each group of samples consists of two 512×512 images and corresponding binary labels for farmland changes. Figure 5 As shown in Figure 2, the change types annotated in the RSCD farmland change detection dataset mainly include roads, lakes, and bare land.

[0178] In order to test the performance of the proposed network, four commonly used evaluation indicators are selected: precision, recall rate, F1 score and intersection-over-union ratio for evaluation; in order to prove the effectiveness of the adaptive edge-guided context-aggregated remote sensing image change detection method proposed in the present invention, four methods are selected in this embodiment for comparison, including STANet model, EGRNN model, BIT model and MSCANet model; STANet model is a spatiotemporal attention change detection model based on twin structure, which uses spatiotemporal attention mechanism to capture spatiotemporal dependencies at different scales. EGRNN model uses recurrent neural network for change detection, uses LSTM analysis to establish temporal correlation between images, and enhances the robustness of the network; BIT model is the first model to introduce Transformer encoder-decoder into the change detection task, realizing long-distance modeling of dual-phase images; MSCANet model is a change detection model based on CNN and Transformer hybrid structure, using CNN to obtain local features and Transformer to aggregate context information.

[0179] When conducting comparative verification experiments based on the CLCD cultivated land change detection dataset, the CLCD dataset is mainly based on land changes, and the image data is greatly affected by seasonal differences and brightness and tone. These characteristics bring huge challenges to the network's identification of changed objects. The STANet model is greatly affected by changes in illumination and brightness, and is prone to missed detection for some small targets. It can detect objects with large target changes. This is because the STANet receptive field is limited, resulting in insufficient feature extraction capabilities. EGRNN has good detection effects on large-scale change objects, with tight internal changes and continuous boundaries, but false detections may occur for small target changes, which reflects the high recall rate and low precision of the EGRNN model. The BIT model has loose internal changes in large target change detection and missed detections on small targets. This is because the BIT model loses some local detail information of the change area during direct upsampling, and has a weak ability to detect changes in small objects. MSCANet has false detections, because the local features provided by CNN to Transformer still lose some detail information during the acquisition process. The method proposed in the present invention has achieved relatively good results in the recognition of changes in small and large targets. Edge guidance and context aggregation can complement local detail features and global semantic information, thereby improving the detection effect. The comparative experimental evaluation results based on the CLCD cultivated land change detection dataset are shown in Table 1:

[0180] Table 1

[0181] Accuracy Recall F1 score Intersection and Union STANet model 65.13％ 69.84％ 67.34％ 72.71％ EGRNN model 59.36％ 73.27％ 65.58％ 71.37％ BIT Model 68.98％ 55.11％ 61.27％ 69.38％ MSCANet Model 71.96％ 71.46％ 71.71％ 75.73％ Method of the present invention 73.52％ 72.14％ 72.82％ 76.51％

[0182] As can be seen from Table 1, the method of the present invention is superior to other comparison models in terms of accuracy, F1 score and intersection over union ratio. Among them, the accuracy, recall rate and F1 score of the STANet model and the BIT model are all low. Although the STANet model can obtain time-space dependencies of different scales, its receptive field is limited and it is easily affected by noise such as light tone. The BIT model has a large receptive field and can capture global features, but it is weak in local detail positioning. The EGRNN model has a higher recall rate and lower accuracy, which is consistent with the visualization results. The overall result of the MSCANet model is better than the first three models in the table, but its accuracy is lower than that of the method of the present invention. Although the MSCANet model adopts a hybrid structure of CNN and Transformer, in the process of extracting local features using CNN, some detail information will still be lost due to the characteristics of CNN itself.

[0183] When conducting comparative verification experiments based on the RSCD cultivated land change detection dataset, 4 pairs of multi-temporal images in the RSCD dataset were selected, mainly focusing on road changes and cultivated land changes. The STANet model can detect changes in roads and cultivated land, but there will be discontinuous boundaries; the EGRNN model and the BIT model will have false detection problems in the process of detecting cultivated land changes due to the influence of light and season, and in the process of road detection, they will be affected by changes in road structure and cause missed detection; the MSCANet model can detect changes in roads and cultivated land, but there will be problems with blurred boundaries and loose internal structures. The change map obtained by the method of the present invention is closer to the actual results, and can detect the changed area completely and accurately. The comparative experimental evaluation results based on the RSCD cultivated land change detection dataset are shown in Table 2:

[0184] Table 2

[0185] Accuracy Recall F1 score Intersection and Union STANet model 47.40％ 56.51％ 51.62％ 65.73％ EGRNN model 53.26％ 58.04％ 55.54％ 67.74％ BIT Model 61.97％ 55.88％ 58.77％ 69.55％ MSCANet Model 61.37％ 53.44％ 57.13％ 68.71％ Method of the present invention 63.29％ 55.19％ 58.96％ 69.67％

[0186] As can be seen from Table 2, the precision, recall rate, F1 score and intersection-over-union ratio of the STANet model are the lowest among all methods; the EGRNN model has a high recall rate, but its precision is the lowest among all model methods, because the precision and recall rate are negatively correlated, and it is difficult for both to obtain high scores at the same time; the experimental precision of the BIT model is second only to the method proposed in this invention; the overall experimental precision of the MSCANet model is lower than that of the BIT model. Except for the recall rate, which is not the optimal value, the other indicators of the method of the present invention have reached the leading position, which is consistent with the visualization results.

[0187] In order to further verify the effectiveness of each module of the present invention, this embodiment conducts an ablation experiment on the CLCD farmland change detection dataset. The specific operation is to conduct experiments to verify the effectiveness of different components by controlling the presence or absence of the module. The experimental results are in the table below, and the best results in the table are indicated in bold.

[0188] After the difference enhancement module group was introduced on the basic model, the experimental results showed that the recall rate increased by 3.72%, the F1 score increased by 0.27%, and the intersection-over-union ratio increased by 0.08%, but the precision decreased by 2.92%. The introduction of the difference enhancement module group is conducive to reducing the false detection phenomenon in the change detection task; after adding the edge-guided context aggregation proposed by the present invention, the precision and recall rate increased by 1.07% and 0.35% respectively, the F1 score increased by 0.71%, and the intersection-over-union ratio increased by 0.48%; the difference enhancement module group and the edge-guided context information aggregation were combined, and the accuracy and recall rate reached a balance state, maintaining the advantages of the two modules, and the accuracy of each indicator was improved to a certain extent. Compared with the basic model, the accuracy and recall rate increased by 1.56% and 0.68% respectively, the F1 score increased by 1.11%, and the intersection-over-union ratio increased by 0.78%. The present invention obtains better semantic representation by effectively encoding and aggregating context information, and can obtain better detection effects.

[0189] Based on the comprehensive experimental results, the combination of difference enhancement module group and edge-guided context aggregation is effective and improves the detection results.

[0190] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An adaptive edge-guided context-aggregated remote sensing image change detection method, characterized in that: The steps include: S1. Extract dual-temporal and multi-scale features of remote sensing images to be detected based on Res2Net; S2, using the difference enhancement module group to enhance the difference information between the foreground and the background of the dual-phase multi-scale features to obtain enhanced multi-scale features; S3, input the enhanced multi-scale features into the boundary extractor and Transformer modules respectively, aggregate the context information under the guidance of boundary constraints, and obtain the multi-scale feature vector; S4. Based on the multi-scale entropy and Dice hybrid loss function, the multi-scale feature vector is input into the classifier to obtain the remote sensing image change detection result.

2. The adaptive edge-guided context-aggregated remote sensing image change detection method according to claim 1, characterized in that: The S1 comprises the following steps: S11, obtaining three-channel remote sensing images to be detected; S12. Extract the features of the remote sensing image to be detected based on Res2Net to obtain the dual-phase multi-scale features, where the dual-phase multi-scale features include the dual-phase second scale feature F2 outn , dual-phase third scale feature F3 outn and the dual-phase fourth scale feature F4 outn , when n=1, it is the first phase, and when n=2, it is the second phase.

3. The adaptive edge-guided context-aggregated remote sensing image change detection method according to claim 2, characterized in that: The S2 comprises the following steps: S21, based on the difference enhancement module, performing subtraction on the dual-phase multi-scale features to obtain multi-scale initial difference features; The calculation expression of the multi-scale initial difference feature is as follows: D i =|Be out1 -Be out2 | Among them, D i represents the multi-scale initial difference feature, |·| represents the absolute value, Fi out1 Represents the multi-scale features of the first phase, Fi out2 represents the multi-scale features of the second phase, where i=2, 3, 4, i=2 represents the second scale, i=3 represents the third scale, and i=4 represents the fourth scale; S22, performing average pooling and maximum pooling on the multi-scale initial difference features respectively to obtain an average pooling result and a maximum pooling result; S23, respectively extracting the features of the average pooling result and the maximum pooling result, and inputting them into a multi-layer perceptron for channel dimension feature learning, thereby extracting the average pooling channel dimension feature learning result and the maximum pooling channel dimension feature learning result; S24, normalizing the average pooling channel dimension feature learning result and the maximum pooling channel dimension feature learning result to obtain the attention weight; The calculation expression of the attention weight is as follows: to i =σ(MLP(Avgpool(D i )+MLP(Maxpool(D i )))) Among them, att i represents the attention weight, σ(·) represents the sigmoid function, MLP(·) represents the multi-layer perceptron, Avgpool(·) represents the average pooling, and Maxpool(·) represents the maximum pooling; S25. Based on the attention weights, construct a residual connection for the bi-phase multi-scale features to obtain enhanced multi-scale features, wherein the enhanced multi-scale features include first-scale difference enhanced features, second-scale difference enhanced features, and third-scale difference enhanced features; The calculation expression of the enhanced multi-scale feature is as follows: Out n =Fi outn *to i +Fi outn Among them, Out n Represents the enhanced multi-scale features, Fi outn Represents dual-phase multi-scale features, where n=1, 2, n=1 represents the first phase, and n=2 represents the second phase.

4. The adaptive edge-guided context-aggregated remote sensing image change detection method according to claim 3, characterized in that: The S3 comprises the following steps: S31, using a boundary extractor to extract boundary features of the enhanced multi-scale features to obtain boundary features of the multi-scale features, wherein the boundary features of the multi-scale features include boundary features of the first scale difference enhanced features, boundary features of the second scale difference enhanced features, and boundary features of the third scale difference enhanced features; S32, inputting the third scale difference enhancement feature into the encoder of the Transformer module for encoding to obtain a first global semantic feature; S33, using the boundary feature of the third scale difference enhancement feature as a boundary constraint guide, and using the decoder of the Transformer module to decode the first global semantic feature to obtain a first global feature vector; S34, inputting the first global feature vector and the second scale difference enhancement feature into the encoder of the Transformer module for encoding to obtain a second global semantic feature; S35, using the boundary feature of the second scale difference enhancement feature as a boundary constraint guide, and using the decoder of the Transformer module to decode the second global semantic feature to obtain a second global feature vector; S36, inputting the second global feature vector and the first scale difference enhancement feature into an encoder of the Transformer module for encoding to obtain a third global semantic feature; S37, using the boundary feature of the first scale difference enhancement feature as a boundary constraint guide, and using the decoder of the Transformer module to decode the third global semantic feature to obtain a third global feature vector; S38, based on a multi-scale feature fusion method, concatenate and fuse the first global feature vector, the second global feature vector, and the third global feature vector to obtain a fourth global feature vector; S39: Use the first global feature vector, the second global feature vector, the third global feature vector, and the fourth global feature vector as multi-scale feature vectors.

5. The adaptive edge-guided context-aggregated remote sensing image change detection method according to claim 4, characterized in that: The S31 comprises the following steps: S311, based on the boundary extractor, performing 3*3 convolution filtering processing on the enhanced multi-scale features to obtain a first multi-scale convolution filtering result; The calculation expression of the first multi-scale convolution filtering result is as follows: Among them, y(p0) represents the first multi-scale convolution filtering result, p n represents the enumeration in the convolution receptive field, ∈ represents belonging, R represents the convolution kernel receptive field, w(p n ) represents the filter window, x(p0+p n ) represents the 3*3 convolution input, and p0 represents the position of convolution filtering in the multi-scale feature; S312, based on the boundary extractor, performing deformable convolution filtering processing on the enhanced multi-scale features to obtain a second multi-scale convolution filtering result; The calculation expression of the second multi-scale convolution filtering result is as follows: Among them, y(p0)′ represents the second multi-scale convolution filtering result, x(p0+p n +Δp n ) represents the deformable convolution filter input, Δp n Indicates the predicted offset; S313, performing batch normalization processing on the first multi-scale convolution filtering result and the second multi-scale convolution filtering result respectively to obtain normalized results of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result; S314, reconstructing the normalized results of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result to obtain the reconstructed first multi-scale convolution filtering result and the second multi-scale convolution filtering result; The calculation expressions of the reconstructed first multi-scale convolution filtering result and the second multi-scale convolution filtering result are as follows: y(p0) rec =γy(p0) Nor +b′ y(p0)′ rec =γy(p0)′ Nor +b′ Among them, y(p0) rec Represents the reconstructed first multi-scale convolution filtering result, y(p0)′ rec represents the reconstructed second multi-scale convolution filtering result, γ represents the first reconstruction learning parameter, and β′ represents the second reconstruction learning parameter; S315, performing nonlinear activation on the reconstructed first multi-scale convolution filtering result and the second multi-scale convolution filtering result, and summing them to obtain boundary features of multi-scale features, wherein the boundary features of the multi-scale features include boundary features of the first scale difference enhancement feature, boundary features of the second scale difference enhancement feature, and boundary features of the third scale difference enhancement feature; The calculation expression of the boundary feature of the multi-scale feature is as follows: Fside=max(0,y(p0) rec )+max(0,y(p0)′ rec ) ={Fside1,Fside2,Fside3} Among them, Fside represents the boundary feature of multi-scale features, max(0,y(p0) rec ) represents the nonlinear activation result of the reconstructed first multi-scale convolution filter result, max(0,y(p0)′ rec ) represents the nonlinear activation result of the reconstructed second multi-scale convolution filtering result, Fside1 represents the boundary feature of the first scale difference enhancement feature, Fside2 represents the boundary feature of the second scale difference enhancement feature, and Fside3 represents the boundary feature of the third scale difference enhancement feature.

6. The adaptive edge-guided context-aggregated remote sensing image change detection method according to claim 5, characterized in that: The S313 comprises the following steps: S3131, respectively calculating the mean of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result; The calculation expression of the mean of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result is as follows: Among them, μ β represents the mean of the first multi-scale convolutional filtering result, m represents the size of the training sample Batch, i′ represents the i′th Batch, μ β ′ represents the mean value of the second multi-scale convolution filtering result; S3132. Calculate the standard deviation of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result based on the mean of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result; The calculation expression of the standard deviation of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result is as follows: Among them, σ β represents the standard deviation of the first multi-scale convolution filtering result, σ β ′ represents the standard deviation of the second multi-scale convolution filtering result; S3133. Based on the mean of the first multiscale convolution filtering result and the second multiscale convolution filtering result, and the standard deviation of the first multiscale convolution filtering result and the second multiscale convolution filtering result, respectively calculate the normalized results of the first multiscale convolution filtering result and the second multiscale convolution filtering result; The calculation expression of the normalized result of the first multi-scale convolution filtering result and the second multi-scale convolution filtering result is as follows: Among them, y(p0) Nor Represents the normalized result of the first multi-scale convolution filtering result, y(p0)′ Nor represents the normalized result of the second multi-scale convolution filtering result, and ε represents a constant close to 0.

7. The adaptive edge-guided context-aggregated remote sensing image change detection method according to claim 4, characterized in that: The calculation expression of the Transformer module is as follows: Among them, Attention(Q,K,V) represents the attention mechanism in the Transformer module, Q represents the query vector, K represents the key vector, T represents the transpose, V represents the value vector, and d k Represents the length of the key vector; the encoding stage of the Transformer module divides the encoder input into several fixed-size image blocks, and adds position encoding to each image block. The correlation between different image blocks is calculated through the multi-head attention mechanism to obtain global semantic features; the decoding stage of the Transformer module inputs multi-scale boundary features into the decoder, guides the fusion of global semantic features and spatial features of different scales, and completes context information aggregation.

8. The adaptive edge-guided context-aggregated remote sensing image change detection method according to claim 1, characterized in that: The calculation expression of the multi-scale entropy and Dice mixed loss function is as follows: Loss=L BCE +L Dice L BCE =-[y j logp j +(1-y j )log(1-p j )] Among them, LOSS represents the multi-scale entropy and Dice mixed loss function, L BCE represents the multi-scale entropy loss function, L Dice represents the Dice loss function, y j represents the true label value of pixel j in the remote sensing image to be detected, p j Represents the predicted probability of a change in pixel j in the remote sensing image to be detected.

Images