Remote sensing image change detection method based on multi-scale semantic labeling transformer network

Abstract

The application relates to a remote sensing image change detection method based on a multi-scale semantic label Transformer network, and relates to a remote sensing image change detection method. In order to solve the problem that the existing Transformer-based method neglects spatial information of a change region to a certain extent, thereby affecting the remote sensing image change detection capability, the application first extracts a feature map by using a feature encoder-decoder, converts feature maps of different scales from the feature encoder into semantic labels of different lengths by using a multi-scale semantic label encoder, and then inputs the semantic labels into a Transformer to obtain global semantic information. Then, the multi-scale semantic label decoder is used to combine the semantic labels of different scales containing rich semantic information and multi-scale feature maps containing rich spatial information to obtain semantic spatial joint features, and finally, the multi-scale semantic spatial joint features are aggregated in a feature decoder in a skip connection mode, and the final change result map is obtained through a classifier.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing technology, and specifically relates to a method for detecting changes in remote sensing images. Background Technology

[0002] The task of remote sensing image change detection is to detect change information in remote sensing images of the same area at different time phases. Due to the impact of human activities and ecosystems, the Earth's surface is constantly changing. Timely and accurate acquisition of surface change information can provide technical support for various decision-making processes, such as environmental protection, disaster detection, and natural resource management. With the development of remote sensing technology, the spatial resolution of acquired remote sensing images has continuously improved. However, while high-resolution images bring rich details of ground features, they also create difficulties in detecting changed areas: in complex scenes, irrelevant interference is significant, making accurate detection of changed areas impossible.

[0003] Currently, change detection methods based on convolutional neural networks (CNNs) perform well in extracting local information, but struggle to extract global information from high-resolution images. Transformers, which effectively extract long-range dependencies in inputs, were initially applied in Natural Language Processing (NLP). With the advent of Vitality Interaction (ViT), Transformers entered the field of computer vision and have achieved widespread application. However, while Transformer-based methods can effectively extract global semantic information between pixels, they to some extent neglect spatial information in changing regions. Therefore, this invention considers introducing multi-scale information to reduce information loss in Transformers and designs a multi-scale semantic tag decoder to aggregate spatial and semantic information, obtaining joint semantic-spatial information for better detection of changing regions and filtering out unnecessary interference. Summary of the Invention

[0004] This invention addresses the problem that existing Transformer-based methods, to some extent, ignore spatial information of changing regions, thus affecting the change detection capability of remote sensing images. Therefore, it proposes a remote sensing image change detection method based on a multi-scale semantically labeled Transformer.

[0005] 1. A remote sensing image change detection method based on a multi-scale semantic label Transformer network, characterized in that: firstly, a dual-temporal image of the remote sensing image is acquired, and then input into a remote sensing image change detection network for detection to obtain remote sensing image change information; the remote sensing image change detection network includes two encoding networks, a concat layer, a Transformer module, two multi-scale semantic label decoders, a feature decoder, and a classifier;

[0006] Two encoding networks are respectively input to one phase image of the dual-temporal image. The two encoding networks have the same structure, each including a feature encoder and a multi-scale semantic label encoder.

[0007] The feature encoder comprises four sequentially connected convolutional neural network units; the output of each feature encoder is input to its corresponding multi-scale semantic tag encoder and multi-scale semantic tag decoder.

[0008] A multi-scale semantic tag encoder consists of four parallel semantic tag encoders. The feature encoder comprises four sequentially connected convolutional neural network units, each with its own output serving as the input to one of the four parallel semantic tag encoders. The processing procedure of the semantic tag encoder is as follows:

[0009] Let the input feature map be F. First, the number of channels in the feature map is set to a fixed value N through a pixel-level convolutional layer, i.e. Where H and W are the width and height of the feature map; then the feature map F1 is passed through a convolutional layer with a kernel of 1 and a softmax activation layer to obtain the attention map. Finally, multiplying F1 and A yields the semantic tag.

[0010] The concat layer superimposes the multi-scale semantics of the multi-scale semantic tag encoder along the dimension of the semantic tag length to obtain new semantic tags. Where L refers to the sum of the lengths of semantic tags at different scales, and C is the number of channels for semantic tags;

[0011] The Transformer module is used for new semantic tags. Processing is performed to obtain global semantic information T out Then, the global semantic information T out Input two multi-scale semantic tag decoders respectively;

[0012] The two multi-scale semantic tag decoders have the same structure; each multi-scale semantic tag decoder includes four semantic tag decoders at different scales. Within each semantic tag decoder, a feature map is processed... Semantic tags First, the number of channels in feature map F is set to N through a pixel-level convolutional layer, i.e. Then the global semantic information T out Attention weights are obtained through a softmax activation function layer and multiplied by a feature map F2 with N channels to obtain the semantic feature map. Next, semantic feature map F sem Aggregation is performed using a convolutional layer with a kernel of 1 and combined with the feature map. Adding them together yields the joint features of the semantic space. Finally, a convolutional layer with a kernel of 1 is used to set the number of channels of the joint feature to C; each semantic tag decoder at different scales obtains a semantic space joint feature map, and the two multi-scale semantic tag decoders obtain a bi-temporal semantic space joint feature map accordingly;

[0013] The feature decoder consists of three deconvolutional layers connected in sequence. First, the bi-temporal images are passed through the feature encoder to obtain two outputs. and The input to the feature decoder is obtained by subtracting the two outputs and taking the absolute value. Then, the joint feature maps of semantic space from semantic label decoders of different scales in the multi-scale semantic label decoder are subtracted and their absolute values ​​are superimposed on their respective inputs and fed into the deconvolution layer.

[0014] The feature decoder obtains the final output. Then, the classifier classifies the output results to obtain the final change result map.

[0015] Furthermore, the remote sensing image change detection network is pre-trained, and the loss function used in training the remote sensing image change detection network is as follows:

[0016] Loss = αLoss ce +(1-α)Loss dice

[0017] Where α is the adjustment parameter, Loss ce The cross-entropy loss function, Loss dice Here is the Dice loss function; the Dice loss function is as follows:

[0018]

[0019] Where X refers to the prediction result map and Y refers to the ground truth map.

[0020] Preferably, α is set to 0.9.

[0021] Furthermore, the input to the feature decoder:

[0022]

[0023] Here, abs(·) means taking the absolute value of the content within the parentheses.

[0024] Furthermore, the process of subtracting the absolute values ​​of the joint feature maps of semantic spaces from semantic tag decoders of different scales in the multi-scale semantic tag decoder and then superimposing them with their respective inputs before feeding them into the deconvolution layer is represented as follows:

[0025]

[0026] in, and This represents the multi-scale semantic space joint feature map of a dual-temporal image, where different values ​​of 'i' indicate different feature map scales. This represents the processed output corresponding to scale i-1; concat(·) represents the concatenation operation, and TransConv(·) represents the deconvolution operation.

[0027] Furthermore, the first convolutional neural network unit of the feature encoder adopts a ResNet18 with pre-trained weights, and the last two layers of the ResNet18 are replaced with convolutional layers with a kernel of 1.

[0028] Furthermore, the last three convolutional neural network units of the feature encoder have the same structure, each including four convolutional layers, an average pooling layer between the third and fourth convolutional layers, and a maximum pooling layer between the first and second convolutional layers.

[0029] Preferably, for the last three convolutional neural network units of the feature encoder, the kernel sizes of the four convolutional layers in each convolutional neural network unit are 7*7, 3*3, 3*3, and 3*3, respectively.

[0030] Furthermore, the Transformer module processing procedure is as follows:

[0031] For its input First, a positional encoding is added to account for the position of each semantic tag, and then a multi-head self-attention mechanism is used for processing. This is then fed into a Norm layer, followed by an MLP layer to obtain the global semantic information T. out ;

[0032] The MLP layer consists of a fully connected linear layer and a Gaussian activation function layer placed between the two fully connected linear layers, as shown in the formula:

[0033] MLP(T′)=GELU(T′W1)W2

[0034] in, It is a learnable linear parameter matrix; GELU(·) represents a Gaussian activation function layer; T′ represents the input.

[0035] Furthermore, before inputting the dual-temporal images into the remote sensing image change detection network, the dual-temporal images need to be segmented into 256×256 image blocks, and then input into the remote sensing image change detection network for detection.

[0036] Beneficial effects:

[0037] This invention proposes a remote sensing image change detection algorithm based on a multi-scale semantic labeling Transformer. Addressing the issue that most Transformer-based methods only consider global semantic information while neglecting spatial information, this invention employs a structure combining Transformer and convolutional neural networks to extract global and local features from high-resolution remote sensing images. The proposed multi-scale semantic labeling encoder transforms feature maps at different scales into semantic labels, reducing the computational cost of the Transformer and minimizing information loss during the transformation process. Furthermore, to more effectively identify changed regions and filter out irrelevant interference, this invention designs a multi-scale semantic labeling decoder that combines feature maps and semantic labels to obtain joint semantic-spatial features. Experimental results on two datasets show that the overall evaluation index of this invention is higher than that of current state-of-the-art change detection methods, thus validating the effectiveness of the invention. Attached Figure Description

[0038] Figure 1 This is a network framework diagram of the method proposed in this invention.

[0039] Figure 2 It is a semantic tag encoder.

[0040] Figure 3 This is a schematic diagram of the Transformer structure.

[0041] Figure 4 The following are partial detection results of the present invention on two high-resolution remote sensing images, wherein the first column is the image before the change, the second column is the image after the change, the third column is the true value map of the change, and the fourth column is the detection result of the present invention. Detailed Implementation

[0042] Specific implementation method one: Combining Figure 1 This implementation method is described below.

[0043] This implementation method is a remote sensing image change detection method based on a multi-scale semantically labeled Transformer network.

[0044] 1. The original image is cropped into image patches of uniform size, and divided into training set, validation set and test set according to the number of patches, and then fed into the network.

[0045] In this embodiment, the biphase image is segmented into 256×256 image blocks, and experiments were conducted on two datasets.

[0046] 2. An encoder-decoder structure is employed, combining Transformer and Convolutional Neural Networks. This not only extracts long-range dependency information from images but also introduces inductive bias, maximizing the advantages of both Transformer and CNN. The feature encoder and decoder each utilize Convolutional Neural Networks, with the Transformer acting as a converter connecting them. The bi-temporal input images are encoded by the feature encoder, and the absolute value of the output difference is obtained at the end of the feature decoder. This difference is then used by the decoder to finally reconstruct the transformation result.

[0047] 3. Design a multi-scale token encoder module to transform feature maps of different scales from the feature encoder into semantic tokens of different lengths. This not only reduces the computational cost of the self-attention mechanism in the Transformer, but also reduces information loss during the transformation process.

[0048] 4. Input the multi-scale semantic tags obtained in step 3 into Transformer to calculate global semantics and obtain global semantic information.

[0049] 5. Design a multi-scale semantic token decoder module. This decoder combines semantic tokens with rich semantic information at different scales with multi-scale feature maps with rich spatial information to obtain joint semantic space features, which helps to identify changing regions and filter out irrelevant interference. Finally, the joint semantic space features at multiple scales are aggregated in the feature decoder using skip connections.

[0050] Because the imbalance between positive and negative samples poses a significant challenge to network learning, this invention employs a hybrid loss function that combines the basic cross-entropy loss function with the Dice loss function.

[0051] The cross-entropy loss function is a pixel-level loss function applied to each pixel value of the prediction result. Its specific calculation formula is as follows:

[0052]

[0053] However, in change detection tasks, the large disparity between positive and negative samples makes it easier for the model to learn features from negative samples but difficult to learn features from positive samples. Therefore, this invention employs the Dice loss function. The Dice loss function addresses the imbalance between positive and negative samples by considering the IoU between the predicted result and the ground truth graph. It is an object-level loss function, and its specific formula is shown below:

[0054]

[0055] Here, X refers to the predicted result image, and Y refers to the ground truth image. The two loss functions are combined by a weighted average, as shown in the following formula:

[0056] Loss = αLoss ce +(1-α)Loss dice

[0057] Experiments showed that the best results were achieved when α was set to 0.9, therefore, this invention sets α to 0.9.

[0058] This implementation uses an encoder-decoder as the overall framework and employs supervised training. More specifically, it uses a convolutional neural network as the encoder-decoder and a Transformer as the converter connecting the encoder and decoder. Figure 1 As shown, the feature encoder consists of four units;

[0059] The first unit of the feature encoder uses a ResNet18 with pre-trained weights. The original ResNet18 has 5 layers. This invention follows the method in BIT, selects the first 3 layers, and replaces the last two layers with convolutional layers with a kernel of 1.

[0060] The last three units of the feature encoder have the same structure, consisting of four convolutional layers, one average pooling layer, and one max pooling layer. Specific parameter settings are shown in Table 1.

[0061] Table 1. Parameter settings for the CNN in the feature encoder.

[0062]

[0063] This invention also designs a multi-scale semantic tag encoder, which transforms feature maps of different scales from the feature encoder into semantic tags of different lengths. This not only reduces the computational cost of the self-attention mechanism in the Transformer but also reduces information loss during the transformation process. A multi-scale semantic tag encoder consists of four semantic tag encoders. Figure 2 This is a schematic diagram of a semantic tag encoder, with the input feature map being... Where H, W, and C represent the width, height, and number of channels of the feature map. Firstly, to facilitate the concatenation of semantic labels at different scales, the number of channels in the feature map is set to a fixed value N through a pixel-level convolutional layer, i.e. In this invention, the fixed value N is set to 32. Then, the feature map F1 is passed through a convolutional layer with a kernel of 1 and a softmax activation layer to obtain the attention map. Finally, multiplying F1 and A yields the semantic tag. Depending on the scale of the feature map, the length of the semantic tag will also be different. In this invention, the length of the semantic tags generated by the multi-scale semantic tag encoder is set to 4, 8, 16 and 32 respectively.

[0064] The multi-scale semantics of the obtained dual-temporal images are superimposed along the dimension of semantic tag length to obtain new semantic tags. Where L represents the sum of the lengths of the semantic tags at different scales, and C represents the number of channels for the semantic tags. The superimposed tags are then fed into a Transformer to calculate the global semantics, obtaining the global semantic information T. out (The output of Transformer).

[0065] The structure of Transformer is as follows: Figure 3 As shown, the Transformer employs a self-attention mechanism for its input. First, a positional encoding is added to consider the position of each semantic tag. Then, a multi-head self-attention mechanism is used to process it. For each tag, it is multiplied by three different matrix weights to obtain the sequence Q, the key K, and the value V, as shown in the formula below:

[0066] Q = TW q K = TW k V = TW v

[0067] in, These are three learnable, linearly varying parameters, where d is the number of channels after transformation. After obtaining the semantic tags Q, K, and V, the self-attention weights can be obtained through self-attention calculation, as shown in the following formula:

[0068]

[0069] To extract richer feature information, the Transformer employs a multi-head self-attention mechanism, the core of which is the parallel computation of multiple attention layers, as shown in the following formula:

[0070] Multihead(Q,K,V)=Concat(head1,head2,...head h W O head i =Attention(TW i q TW i k TW i v )

[0071] Where h refers to the number of attention heads. and These are linearly learnable matrix parameters.

[0072] Then it is fed into a Norm layer, and then into an MLP layer to obtain global semantic information T. out The MLP layer consists of a fully connected linear layer and a Gaussian activation function layer placed between the two fully connected linear layers, as shown in the formula:

[0073] MLP(T′)=GELU(T′W1)W2

[0074] in, It is a learnable linear parameter matrix; GELU(·) represents a Gaussian activation function layer; T′ represents the input.

[0075] This invention designs a multi-scale semantic tag decoder. Although semantic tags of different scales obtain rich semantic information through the Transformer, these semantic tags cannot be combined with feature maps containing rich spatial information. Therefore, to aggregate spatial and semantic information, this invention proposes a multi-scale semantic tag decoder that combines semantic tags of different scales containing rich semantic information with multi-scale feature maps containing rich spatial information to obtain joint semantic-space features, which helps to identify changing regions and filter out irrelevant interference. Finally, the multi-scale joint semantic-space features are aggregated in the feature decoder in a skip connection manner. The multi-scale semantic tag decoder includes four semantic tag decoders of different scales. In the semantic tag decoder, for the feature map... Semantic tags First, to aggregate these semantic tags, the number of channels in the feature map F is set to N through a pixel-level convolutional layer, i.e. Then the global semantic information T out Attention weights are obtained through a softmax activation function layer and multiplied by a feature map F2 with N channels to obtain the semantic feature map. Next, semantic feature map F sem Aggregation is performed using a convolutional layer with a kernel of 1 and combined with the feature map. Adding them together yields the joint features of the semantic space. Finally, in order to facilitate the feeding of the joint semantic space features into the feature decoder via skip connections, a convolutional layer with a kernel of 1 is used to set the number of channels of the joint features to C.

[0076] This invention designs a feature decoder, which mainly consists of three deconvolutional layers for upsampling. Each deconvolutional layer has the same parameter settings: kernel_size is set to 3, stride to 2, padding to 1, and output_padding to 1. First, the bi-temporal image is passed through the feature encoder to obtain two outputs. and The input to the feature decoder is obtained by subtracting the two outputs and taking the absolute value, as shown in the formula:

[0077]

[0078] Among them, F in This represents the input of the first upsampling layer, and abs(·) represents taking the absolute value of the content within the parentheses.

[0079] Subsequently, the absolute value of the subtraction between the bi-temporal semantic space joint feature maps from the multi-scale semantic tag decoder is superimposed on the input and fed into the deconvolution layer, as shown in the formula:

[0080]

[0081] in, and This represents the multi-scale semantic space joint feature map of the dual-temporal image. Different values ​​of i indicate different feature map scales. concat(·) represents the concatenation operation, and TransConv(·) represents the deconvolution operation.

[0082] The final output of the feature decoder was obtained. Subsequently, this invention designs a simple classifier to perform binary classification on the output results, thereby obtaining the final change result image. This classifier contains two convolutional layers, with a regularization layer and an activation function layer placed between them. The parameters of the two convolutional layers are identical: kernel_size is set to 3, stride is set to 1, and padding is set to 1. The specific implementation process is shown in the formula:

[0083]

[0084] Example

[0085] The simulation was conducted according to Specific Implementation Method 1. This invention processed the CDD dataset and the WHU-CD dataset respectively. The CDD dataset, taken from Google Earth, contains seven pairs of real remote sensing images that vary with the seasons, with a spatial resolution ranging from 2 to 100 cm. In the simulation, the seven pairs of images were cut into 256×256 image patches with a step size of 128. The WHU-CD dataset mainly focuses on changes in buildings, with a size of 32507×15345 and a spatial resolution of 0.2 m. In this simulation, this dataset was cut into 256×256 image patches in a non-overlapping manner. The processing results are as follows... Figure 4 As shown, in the partial detection results on two high-resolution remote sensing images, the first and second rows are the processing results of the CDD dataset, and the third and fourth rows are the processing results of the WHU-CD dataset; the first column is the image before the change, the second column is the image after the change, the third column is the ground truth map of the change, and the fourth column is the detection result of this invention. The results from the images show that the changed areas can be detected almost completely. Furthermore, because this invention considers the semantic space joint features of dual-temporal images, it can effectively handle noise such as shadows and illumination, and more accurately identify the changed areas.

[0086] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A remote sensing image change detection method based on a multi-scale semantically labeled Transformer network, characterized in that, First, a dual-temporal image of the remote sensing image is acquired, and then input into a remote sensing image change detection network for detection to obtain remote sensing image change information. The remote sensing image change detection network includes two encoding networks, a concat layer, a Transformer module, two multi-scale semantic label decoders, a feature decoder, and a classifier. Two encoding networks are respectively input to one phase image of the dual-temporal image. The two encoding networks have the same structure, each including a feature encoder and a multi-scale semantic label encoder. The feature encoder comprises four sequentially connected convolutional neural network units; the output of each feature encoder is input to its corresponding multi-scale semantic tag encoder and multi-scale semantic tag decoder. A multi-scale semantic tag encoder consists of four parallel semantic tag encoders. The feature encoder comprises four sequentially connected convolutional neural network units, each with its own output serving as the input to one of the four parallel semantic tag encoders. The processing procedure of the semantic tag encoder is as follows: Let the input feature map be F. First, the number of channels in the feature map is set to a fixed value N through a pixel-level convolutional layer, i.e. Where H and W are the width and height of the feature map; then the feature map F1 is passed through a convolutional layer with a kernel of 1 and a softmax activation layer to obtain the attention map. Finally, multiplying F1 and A yields the semantic tag. The concat layer superimposes the multi-scale semantics of the multi-scale semantic tag encoder along the dimension of the semantic tag length to obtain new semantic tags. Where L refers to the sum of the lengths of semantic tags at different scales, and C is the number of channels for semantic tags; The Transformer module is used for new semantic tags. Processing is performed to obtain global semantic information T out Then, the global semantic information T out Input two multi-scale semantic tag decoders respectively; The two multi-scale semantic tag decoders have the same structure; each multi-scale semantic tag decoder includes four semantic tag decoders at different scales. Within each semantic tag decoder, a feature map is processed... Semantic tags First, the number of channels in feature map F is set to N through a pixel-level convolutional layer, i.e. Then the global semantic information T out Attention weights are obtained through a softmax activation function layer and multiplied by a feature map F2 with N channels to obtain the semantic feature map. Next, semantic feature map F sem Aggregation is performed using a convolutional layer with a kernel of 1 and combined with the feature map. Adding them together yields the joint features of the semantic space. Finally, a convolutional layer with a kernel of 1 is used to set the number of channels of the joint feature to C; each semantic tag decoder at different scales obtains a semantic space joint feature map, and the two multi-scale semantic tag decoders obtain a bi-temporal semantic space joint feature map accordingly; The feature decoder consists of three deconvolutional layers connected in sequence. First, the bi-temporal images are passed through the feature encoder to obtain two outputs. and The input to the feature decoder is obtained by subtracting the two outputs and taking the absolute value. Then, the joint feature maps of semantic space from semantic label decoders of different scales in the multi-scale semantic label decoder are subtracted and their absolute values ​​are superimposed on their respective inputs and fed into the deconvolution layer. The feature decoder obtains the final output. Then, the classifier classifies the output results to obtain the final change result map.

2. The remote sensing image change detection method based on a multi-scale semantic labeling Transformer network according to claim 1, characterized in that, The remote sensing image change detection network is pre-trained, and the loss function used in training the network is as follows: Loss=αLoss ce +(1-α)Loss dice Where α is the adjustment parameter, and Loss ce Let cross-entropy be the loss function, and Loss be the loss function. dice Here is the Dice loss function; the Dice loss function is as follows: Where X refers to the prediction result map and Y refers to the ground truth map.

3. The remote sensing image change detection method based on a multi-scale semantic tagging Transformer network according to claim 2, characterized in that, α is set to 0.

9.

4. The remote sensing image change detection method based on a multi-scale semantic tagging Transformer network according to claim 3, characterized in that, Input to the feature decoder: Here, abs(·) means taking the absolute value of the content within the parentheses.

5. The remote sensing image change detection method based on a multi-scale semantic labeling Transformer network according to claim 4, characterized in that, The process of subtracting the joint feature maps of semantic spaces from semantic tag decoders at different scales in a multi-scale semantic tag decoder, taking the absolute value, and then superimposing it with their respective inputs before feeding it into a deconvolution layer is represented as follows: in, and This represents the multi-scale semantic space joint feature map of a dual-temporal image, where different values ​​of 'i' indicate different feature map scales. This represents the processed output corresponding to scale i-1; concat(·) represents the concatenation operation, and TransConv(·) represents the deconvolution operation.

6. The remote sensing image change detection method based on a multi-scale semantic labeling Transformer network according to claim 5, characterized in that, The first convolutional neural network unit of the feature encoder uses a ResNet18 with pre-trained weights, and the last two layers of the ResNet18 are replaced with convolutional layers with a kernel of 1.

7. The remote sensing image change detection method based on a multi-scale semantic tagging Transformer network according to claim 6, characterized in that, The last three convolutional neural network units of the feature encoder have the same structure, each including four convolutional layers, an average pooling layer between the third and fourth convolutional layers, and a maximum pooling layer between the first and second convolutional layers.

8. The remote sensing image change detection method based on a multi-scale semantic labeling Transformer network according to claim 7, characterized in that, For the last three convolutional neural network units of the feature encoder, the kernel sizes of the four convolutional layers in each convolutional neural network unit are 7*7, 3*3, 3*3, and 3*3, respectively.

9. The remote sensing image change detection method based on a multi-scale semantic tagging Transformer network according to claim 7, characterized in that, The Transformer module processing procedure is as follows: For its input First, a positional encoding is added to account for the position of each semantic tag, and then a multi-head self-attention mechanism is used for processing. The result is then fed into a Norm layer, followed by an MLP layer to obtain the global semantic information T. out ; The MLP layer consists of a fully connected linear layer and a Gaussian activation function layer placed between the two fully connected linear layers, as shown in the formula: MLP(T′)=GELU(T′W1)W2 in, It is a learnable linear parameter matrix; GELU(·) represents a Gaussian activation function layer; T′ represents the input.

10. The remote sensing image change detection method based on a multi-scale semantic tagging Transformer network according to any one of claims 1 to 9, characterized in that, Before inputting the dual-temporal images into the remote sensing image change detection network, the dual-temporal images need to be divided into 256×256 image blocks, and then input into the remote sensing image change detection network for detection.

Images