Optical remote sensing image change detection method and system based on change perception and semantic guidance, storage medium and electronic equipment

Abstract

The invention provides an optical remote sensing image change detection method and system based on change perception and semantic guidance, a storage medium and electronic equipment, and the method comprises the following steps: firstly, carrying out the preprocessing of a dual-tense image, and carrying out the labeling of a change region, and generating a change label graph; and cutting the image and the label into small samples, and inputting the small samples as training data into the CASGNet model. Multi-scale features are extracted through an encoder, change perception is carried out through a space-time difference fusion module, cross-time change information is effectively captured, pseudo change noise is restrained, and reliable multi-scale difference features are obtained. Thirdly, the semantic guiding module decodes the multi-scale difference features, semantic masks are generated through the semantic aggregation module, fusion of the multi-scale features is guided, interference is eliminated, and edge details of a change area are complemented; and finally, in a detection stage, the preprocessed sample to be detected is input into the model, a complete and accurate change detection result is obtained, and the method is suitable for a detection task of a complex scene and a high-resolution image.

Description

Technical Field

[0001] The present invention relates to the field of optical remote sensing image change detection, and in particular to an optical remote sensing image change detection method, system, storage medium and electronic equipment based on change perception and semantic guidance. Background Art

[0002] The change detection task aims to identify the state changes or differences by analyzing images acquired at different times of the same geographic area. Change detection tasks can be divided into multiple types according to different subjects, and have been widely used in fields such as land cover analysis, environmental monitoring, urban planning, and post-disaster assessment. With the advancement of remote sensing technology, the widespread application of high-resolution optical remote sensing images has greatly enriched the surface detail information, bringing new opportunities for change detection. However, these high-resolution data also introduce noise and interference, increasing the complexity of detection. Therefore, how to effectively enhance the perception of change information in high-resolution optical remote sensing images has become an important challenge in the field of change detection.

[0003] In recent years, change detection methods based on deep learning have been developing continuously, bringing new breakthroughs in the field of change detection. Various deep learning architectures have been applied to such tasks. Convolutional Neural Networks (CNN) use convolution operations to automatically extract spatial features from input images, which are suitable for capturing texture and edge information in local areas. In remote sensing image change detection, CNN achieves the gradual extraction from shallow details to deep semantic features by stacking multiple convolutional layers. For example: FC-EF (Fully Convolutional Early Fusion) [1] fuses the bi-temporal images at the input end early, and then extracts the fused features through a fully convolutional network; UNet [2] and its variants obtain multi-scale features through an encoder-decoder structure, and perform outstandingly in segmentation and change detection tasks, especially for detecting details in the changed area. Although the CNN model has significant advantages in processing the spatial features of images, its local receptive field limits the capture of long-distance dependencies. In order to address the shortcomings of CNN in long-distance dependencies, the Transformer model was introduced into the change detection task [3]. Transformer uses a self-attention mechanism to model the feature relationship of the image in a global scope, thereby better capturing the association across spatial regions. Although these models show better performance than traditional methods, the traditional dual-temporal feature fusion methods still have shortcomings due to the diversity of changes and the complexity of the remote sensing imaging process. Common fusion strategies such as direct subtraction or splicing of dual-temporal features to obtain difference information, but these simple operations are often highly sensitive to imaging conditions (such as illumination changes, sensor noise, etc.), and lack the ability to perceive the real change area, which can easily cause false detection and missed detection. In addition, multi-scale feature fusion faces the problem that deep semantic features and shallow spatial detail features are difficult to effectively combine: deep features are prone to lose target and edge details, and shallow features can retain spatial details, but lack global semantic understanding and are difficult to accurately distinguish between the subject and background of the change. Existing algorithms try to combine the advantages of both by simply fusing deep and shallow features, but it is easy to cause confusion between subject and background information, especially in complex scenes and subtle change areas. At the same time, due to the lack of effective feature guidance and screening mechanisms, the fusion process fails to highlight key change areas and suppress noise in shallow features, limiting the detection accuracy and the performance of boundary details. Summary of the invention

[0004] The purpose of the present invention is to provide an optical remote sensing image change detection method, system, storage medium and electronic equipment based on change perception and semantic guidance, which can accurately and effectively identify the detail differences between dual-temporal optical remote sensing images.

[0005] The technical solution adopted by the present invention is:

[0006] A change detection method for optical remote sensing images based on change perception and semantic guidance includes the following steps:

[0007] Step A, performing orthorectification, image registration, image stretching and image numerical normalization preprocessing on the bi-temporal optical remote sensing image in sequence, so as to obtain a bi-temporal optical remote sensing image with consistent data distribution;

[0008] Step B: annotating the updated part of the preprocessed dual-temporal optical remote sensing image obtained in the previous step to obtain a corresponding change label map;

[0009] Step C, cutting the change label map obtained in step B and the bi-temporal optical remote sensing image preprocessed in step A into the same size to obtain training samples;

[0010] Step D, create a network model: Create a change detection network model based on change perception and semantic guidance, which is a dual encoder-single decoder architecture. First, use the pre-trained ResNet-18 model as the feature encoder to extract the multi-scale features of the dual-temporal images respectively; then, the proposed spatiotemporal difference fusion module performs scale-by-scale difference fusion on these multi-scale features to capture spatiotemporal change information at different levels; then, the semantic aggregation module aggregates the fused high-level difference features to generate a semantic mask to provide more accurate semantic guidance information for subsequent decoding; finally, the semantic guidance module is used as a decoder to decode the fused difference features layer by layer to generate the final change detection result;

[0011] Step E: training the network model to obtain the optimal network model, which is used to detect the remote sensing images before and after the detection, and output the detection results: input a pair of dual-temporal optical remote sensing images in the training sample into the network model created in step D, and sequentially undergo feature encoding, spatiotemporal differential fusion, semantic aggregation and semantic-guided decoding to obtain the final predicted change map; then combine the binary cross entropy loss with the Dice loss to form a hybrid loss function to calculate the loss between the predicted change map of the dual-temporal image of the training sample and the corresponding label map, minimize the hybrid loss function through the back propagation algorithm, continuously update the model parameters, and save the optimal model.

[0012] The encoding and fusion process in step E specifically includes the following steps:

[0013] Step E1.1, first encode through the weight-sharing ResNet-18 pre-trained model, and extract multi-scale feature maps X1, X2, X3, X4, Y1, Y2, Y3, Y4 respectively; where X i Represents the multi-scale features of phase 1, Y iRepresents multi-scale features of phase 2;

[0014] Step E1.2: Input the bi-temporal feature map of the same scale obtained in the previous step into the spatiotemporal difference fusion module for bi-temporal feature fusion; first, calculate the adaptive weight of the bi-temporal feature through the channel and spatial attention mechanism, then introduce the bi-temporal difference feature to obtain the difference weight through convolution learning, and finally jointly weight the bi-temporal features to capture the change area through multiple dimensions, improve the integrity and detail description of the change area, and finally obtain four different scale difference features D1, D2, D3, D4;

[0015] Step E1.3: The semantic aggregation module is used to semantically aggregate the deep three-layer features D2, D3, and D4 to generate a semantic change mask.

[0016] The semantic guidance decoding process in step E is to perform decoding operations on D1, D2, D3, and D4 in step E1.3, and perform decoding refinement layer by layer through the semantic guidance module, specifically including the following steps:

[0017] Step E2.1, first, use semantic mask to decouple the current layer features and distinguish the foreground subject and background in the changed area;

[0018] Step E2.2, then, the features are processed at multiple scales by convolution kernels of different scales to extract the contextual information of the foreground and background;

[0019] Step E2.3, then, upsample the high-level features and align them with the current-level features, and compare them with the foreground features to eliminate irrelevant noise, so that the foreground features focus on the changing subject;

[0020] Step E2.4, after that, the filtered foreground features are compared with the background features, and the missing details in the foreground are supplemented with the background information;

[0021] Step E2.5: Finally, the mask and refined features obtained after convolution processing are passed to the next semantic guidance module, and decoded by three semantic guidance modules to generate the final predicted change map.

[0022] The bi-temporal fusion operation in step E1.2 specifically includes the following steps: First, perform maximum pooling and average pooling of space and channels on the bi-temporal features to capture information in different regions and levels:

[0023] S c =Concat(Avg(T1),Max(T1),Avg(T2),Max(T2))

[0024] S crepresents the aggregated spatial features, T1 and T2 represent the bi-temporal features, Avg(·) and Max(·) represent the global average pooling and global maximum pooling across the spatial dimensions, respectively; the aggregated spatial features are passed to two one-dimensional convolutions, which are learned to determine the bi-temporal channel weights of the input bi-temporal features; the two channel weights can be expressed as W c1 , W c2 :

[0025] W c1 =Conv1(S c )

[0026] W c2 =Conv2(S c )

[0027] Then use Softmax in the bi-temporal channel weights to make their sum equal to 1. The Softmax method can express the following process:

[0028]

[0029] Where W' c1 and W' c2 Denotes the output bi-temporal channel weight; the bi-temporal spatial weight W' is determined in the same way in the spatial branch s1 and W' s2 , thereby determining the important part of the dual-temporal features in the spatial dimension; adding the dual-temporal channel weight and the dual-temporal space weight to obtain the dual-temporal domain weight W' sc1 , W' sc2 ,

[0030] W' sc1 =W' c1 +W' s1

[0031] W' sc2 =W' c2 +W' s2

[0032] Finally, the bitemporal features are differentiated to obtain the change information, and the difference weight W in the change information is obtained through 3×3 convolution learning and Sigmoid mapping. diff :

[0033] W diff =Sigmoid(Conv 3×3 (|T1-T2|))

[0034] The final difference feature is obtained by combining the adaptation weight and the difference weight to weight the original feature

[0035] D i =Wdiff ×(W' sc1 )×T1+W diff ×(W' sc2 )×T2

[0036] Among them, i=1, 2, 3, 4; i corresponds to different scales of difference features.

[0037] The step E1.3 specifically includes the following steps: inputting the obtained deep difference feature maps D2, D3, and D4 into the semantic aggregation module, multiplying the current features with the deeper features layer by layer, which is equivalent to introducing a multi-level semantic filter, each layer further filters the noise, and only retains the meaningful information consistent with the deep features; finally, by connecting and fusing the features of each level, complementing each other, a richer and more complete semantic representation is generated; the process is as follows:

[0038] D'3=Conv(up(D4))×D3

[0039] D'2=Conv(up(D3'))×D2

[0040] D s =Conv(Concat(up(Conv((Concat(up(D4),D'3))),D'2))

[0041] mask=Sigmoid(D s )

[0042] Among them, D4, D3, and D2 represent the last three layers of multi-scale difference features, up represents double upsampling, Conv represents a convolution block with a convolution kernel of 3, Concat represents concatenation in the channel dimension, D'3 and D'2 represent feature maps after filtering in the previous layer, and D s represents the semantic features obtained by aggregation, and mask represents the obtained semantic mask.

[0043] The semantic mask obtained in step E1.3 is input into the semantic guidance module together with the difference features of the two adjacent scales for fusion decoding. Specifically, the semantic mask mask and its inverse mask (1-mask) are first used to decouple the current layer feature current, and the foreground subject of the changed area is distinguished from the background, and F is obtained. f , F b ;

[0044]

[0045] F fn =ContextExtract(F f )

[0046] Fbn =ContextExtract(F b )

[0047] Next, the features are extracted at multiple scales using convolution kernels k of different scales, where k = 1, 3, 5, 7, respectively, to extract the context information of foreground features and background features to obtain F fn , F bn ; Then, the high-level feature high is upsampled and aligned with the current layer feature current scale, and the foreground feature F fn Subtract to eliminate irrelevant noise and ensure that the foreground feature focuses on the changing subject to obtain refine1; then refine1 is combined with the background feature F bn Add, bn+relu, i.e. batch normalization and Relu activation function, to eliminate the generated negative values; use background information to supplement the missing details of the change area in the foreground to obtain the final refined features; α and β are two self-learning parameters to control the amplitude of elimination and supplementation;

[0048] refine1=bn(relu(up(high)-α×F fn ))

[0049] refine2=bn(relu(refine1+β×F bn ))

[0050] outmap=Sigmoid(Conv(refine2))

[0051] Among them, Conv represents a convolution block with a convolution kernel of 3. The mask outmap is obtained by convolution and Sigmoid mapping the refined feature refine2, and is passed to the next semantic guidance module together with the refined feature refine2. The final predicted change map is obtained after decoding by three semantic guidance modules.

[0052] An optical remote sensing image change detection system based on change perception and semantic guidance includes a preprocessing module for standardizing an input dual-temporal optical remote sensing image; an encoder module for extracting multi-scale features of the dual-temporal image; a spatiotemporal difference fusion module for performing difference fusion on multi-scale dual-temporal features output by two encoders sharing weights; a semantic aggregation module for aggregating deep difference features to obtain a semantic mask; and a semantic guidance module for guiding fusion and decoding operations on adjacent scale difference features.

[0053] After preprocessing to ensure the alignment and consistency of the bi-temporal data, the bi-temporal remote sensing images are input into the encoder module. The encoder adopts a parameter-sharing ResNet-18 structure to extract multi-scale features from the bi-temporal images layer by layer to form the basis for difference fusion. Then, the bi-temporal features of the same scale are differentially fused scale by scale in the spatiotemporal difference fusion module to capture spatiotemporal change information and generate multi-scale difference features. Next, the semantic aggregation module aggregates the high-level difference features to extract global semantic information and generate semantic masks to provide semantic guidance for the subsequent semantic guidance module. Finally, the semantic guidance module guides the fusion and decoding of the multi-scale difference features layer by layer during the decoding process and outputs the final change detection result map.

[0054] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the device where the computer-readable storage medium is located executes any one of the optical remote sensing image change detection methods based on change perception and semantic guidance.

[0055] An electronic device comprises: a memory and a processor, wherein the memory stores a program that can be run on the processor, and when the processor executes the program, the optical remote sensing image change detection method based on change perception and semantic guidance is implemented.

[0056] The present invention takes optical remote sensing image change detection as the application background. Aiming at the problem that the model trained by the existing change detection method is incomplete for the extracted change area and the edge of the change target is irregular, a new change detection method is proposed, which improves the recognition of the overall features of the change area and the detail features. Specifically, the present invention combines two methods: spatiotemporal differential fusion of dual-temporal features and multi-scale feature fusion guided by semantics. The former focuses on the integration of cross-temporal features and perceives the change area through multiple dimensions, while the latter guides the fusion of features of different scales through deep semantic information, and the added refinement strategy makes the expression of the change area clearer and more complete. The organic combination of these two methods can significantly improve the performance of change detection, making the model more accurate and robust when processing complex scenes and high-resolution images. Compared with traditional algorithms, the scheme of the present invention fully mines the differences between dual-temporal images in multiple dimensions and effectively reduces the interference of pseudo-changes. At the same time, the semantic aggregation module is used to generate a deep semantic change mask rich in semantic information as prior information to guide the fusion of multi-scale features. This strategy significantly improves the expression ability and integrity of the model for change features. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0058] Figure 1 The figure is a flow chart of the method of the present invention.

[0059] Figure 2 This is the CASGNet backbone network structure diagram of the present invention.

[0060] Figure 3 This is a structural diagram of the TDFM of the present invention.

[0061] Figure 4 It is the SAM structure diagram of the present invention.

[0062] Figure 5 FIG. 4 is a structural diagram of the SGM of the present invention. DETAILED DESCRIPTION

[0063] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0064] like Figure 1 As shown, the present invention comprises the following steps:

[0065] Step A, performing orthorectification, image registration, image stretching and image numerical normalization preprocessing on the bi-temporal optical remote sensing image in sequence, so as to obtain a bi-temporal optical remote sensing image with consistent data distribution;

[0066] Step B: annotating the updated part of the preprocessed dual-temporal optical remote sensing image obtained in the previous step to obtain a corresponding change label map; wherein the changed areas in the annotated remote sensing image include newly built urban buildings, suburban expansion, and vegetation changes;

[0067] Step C, cutting the change label map obtained in step B and the bi-temporal optical remote sensing image preprocessed in step A into the same size to obtain training samples;

[0068] Step D, create a network model: Create a change detection network model based on change perception and semantic guidance, which is a dual encoder-single decoder architecture. First, use the pre-trained ResNet-18 model as the feature encoder to extract the multi-scale features of the dual-temporal images respectively; then, the proposed spatiotemporal difference fusion module performs scale-by-scale difference fusion on these multi-scale features to capture spatiotemporal change information at different levels; then, the semantic aggregation module aggregates the fused high-level difference features to generate a semantic mask to provide more accurate semantic guidance information for subsequent decoding; finally, the semantic guidance module is used as a decoder to decode the fused difference features layer by layer to generate the final change detection result;

[0069] Step E: training the network model to obtain the optimal network model, which is used to detect the remote sensing images before and after the detection, and output the detection results: input a pair of dual-temporal optical remote sensing images in the training sample into the network model created in step D, and sequentially undergo feature encoding, spatiotemporal differential fusion, semantic aggregation and semantic-guided decoding to obtain the final predicted change map; then combine the binary cross entropy loss with the Dice loss to form a hybrid loss function to calculate the loss between the predicted change map of the dual-temporal image of the training sample and the corresponding label map, minimize the hybrid loss function through the back propagation algorithm, continuously update the model parameters, and save the optimal model.

[0070] The encoding and fusion process in step E specifically includes the following steps:

[0071] Step E1.1: First, encode through the weight-sharing ResNet-18 pre-trained model to extract multi-scale feature maps X1, X2, X3, X4, Y1, Y2, Y3, Y4 respectively; where X i Represents the multi-scale features of phase 1, Y i Represents multi-scale features of phase 2;

[0072] In actual use, the dual-temporal optical remote sensing images in the training samples are input into the model created in step D; first, the input dual-temporal remote sensing images are sent to the ResNet-18 model with shared parameters for encoding. ResNet-18 is a deep residual network that solves the gradient vanishing problem in deep network training by introducing residual connections, so that features can be effectively extracted. Specifically, the ResNet-18 network consists of multiple stages, each of which contains multiple residual blocks. Each residual block extracts local features through two (or three) 3×3 convolutional layers, combined with BatchNorm and ReLU activation functions. Through skip connections, the input and output of the residual block are added, alleviating the gradient vanishing problem. After the last residual block in each stage, the convolutional layer with a stride of 2 is downsampled to gradually reduce the spatial size of the feature map while increasing the number of channels, thereby extracting features X of different scales of the dual-temporal (phase one and phase two) i and Y i (i=1,2,3,4). Finally, we obtain multi-size feature maps X1, X2, X3, X4, Y1, Y2, Y3, Y4, which are equivalent to 1 / 2 width and height (number of channels 64), 1 / 4 width and height (number of channels 128), 1 / 8 width and height (number of channels 256), and 1 / 16 width and height (number of channels 512) of the original image.

[0073] Step E1.2: The bi-phase feature map of the same scale obtained in the previous step (such as X i and Y i ) is input into the spatiotemporal difference fusion module for dual-temporal feature fusion; first, the adaptive weights of the dual-temporal features are calculated through the channel and spatial attention mechanism, and then the dual-temporal difference features are introduced to obtain the difference weights through convolution learning, and finally the dual-temporal features are weighted and fused together to capture the change area through multiple dimensions, thereby improving the integrity and detail description of the change area, and finally obtaining four different scale difference features D1, D2, D3, and D4;

[0074] The bi-temporal fusion operation in step E1.2 specifically includes the following steps: First, the bi-temporal feature graphs X1, X2, X3, X4, Y1, Y2, Y3, Y4 obtained in step F are fused, and one of the fusion operations is taken as an example. Figure 3 As shown in Figure 1, first, the bi-temporal features are subjected to maximum and average pooling in space and channels respectively to capture information in different regions and levels.

[0075] S c =Concat(Avg(T1),Max(T1),Avg(T2),Max(T2))

[0076] S crepresents the aggregated spatial features, T1 and T2 represent the bi-temporal features, Avg(·) and Max(·) represent the global average pooling and global maximum pooling across the spatial dimension, respectively. Concat represents the channel dimension concatenation. The aggregated spatial features are passed to two 1D convolutions, which are learned to determine the bi-temporal channel weights W of the input bi-temporal features. c1 and W c2 The weights of the two channels are expressed as:

[0077] W c1 =Conv1(S c )

[0078] W c2 =Conv2(S c )

[0079] Then use Softmax in the bi-temporal channel weights to make their sum equal to 1. The Softmax method can represent the following process;

[0080]

[0081] This means comparing the bi-temporal weights to determine the higher value among them, thereby determining the important part between the bi-temporal features in the channel dimension. c1 and W' c2 Denotes the output bi-temporal channel weight. The bi-temporal spatial weight W' is determined in the same way in the spatial branch s1 and W' s2 , thereby determining the important part of the dual-temporal features in the spatial dimension. Add the dual-temporal channel weight and the dual-temporal space weight to obtain the dual-temporal weight W' sc1 , W' sc2 :

[0082] W' sc1 =W' c1 +W' s1

[0083] W' sc2 =W' c2 +W' s2

[0084] Finally, the bitemporal features are differentially operated to extract the change information, and the change pattern is learned through 3×3 convolution, combined with Sigmoid mapping to obtain the difference weight W in the change information. diff :

[0085] W diff =Sigmoid(Conv 3×3 (|T1-T2|))

[0086] The final difference feature is obtained by combining the dual time domain weight and the difference weight to weight the original feature.

[0087] D i =W diff ×(W' sc1 )×T1+W diff ×(W' sc2 )×T2

[0088] Among them, D i (i=1, 2, 3, 4; i corresponds to different scales of difference features).

[0089] Step E1.3, semantically aggregate the deep three-layer features D2, D3, and D4 through the semantic aggregation module to generate a semantic change mask. The step E1.3 specifically includes the following steps: input the obtained deep difference feature maps D2, D3, and D4 into the semantic aggregation module, such as Figure 4 As shown in the figure, by multiplying the current features with the deeper features layer by layer, it is equivalent to introducing a multi-level semantic filter. Each layer further filters the noise and only retains the meaningful information consistent with the deep features. Finally, by connecting and fusing the features of each level, they complement each other and generate a richer and more complete semantic representation. The process is as follows:

[0090] D'3=Conv(up(D4))×D3

[0091] D'2=Conv(up(D'3))×D2

[0092] D s =Conv(Concat(up(Conv((Concat(up(D4),D'3))),D'2))

[0093] mask=Sigmoid(D s )

[0094] Among them, D4, D3, and D2 represent the last three layers of multi-scale difference features, up represents double upsampling, Conv represents a convolution block with a convolution kernel of 3, Concat represents channel dimension concatenation, D'3 and D'2 represent feature maps after filtering in the previous layer, and D s represents the semantic features obtained by aggregation, and mask represents the obtained semantic mask.

[0095] The semantic mask obtained in step E1.3 is input into the semantic guidance module together with the difference features of two adjacent scales for fusion decoding, as shown in Figure 5Specifically, we first use the semantic mask and its reverse mask (1-mask) to decouple the current layer feature current, distinguish the foreground subject from the background in the changed area, and obtain F f , F b .

[0096] F f =mask×current

[0097] F b =(1-mask)×current

[0098] F fn =ContextExtract(F f )

[0099] F bn =ContextExtract(F b )

[0100] Among them, ContextExtract means using convolution kernels of different scales (k = 1, 3, 5, 7) to perform multi-scale extraction of features, respectively extracting the context information of foreground features and background features to obtain F fn , F bn Then, the high-level feature high is upsampled and aligned with the current layer feature current scale, and the foreground feature F fn Subtract to eliminate irrelevant noise and ensure that the foreground feature focuses on the changing subject to obtain refine1; then refine1 is combined with the background feature F bn Add, bn+relu (batch normalization and Relu activation function) to eliminate the negative values. Use the background information to supplement the missing details of the change area in the foreground to obtain the final refined features (refine2), where α and β are two self-learning parameters to control the amplitude of elimination and supplementation.

[0101] refine1=bn(relu(up(high)-α×F fn ))

[0102] refine2=bn(relu(refine1+β×F bn ))

[0103] outmap=Sigmoid(Conv(refine2))

[0104] Conv represents a convolution block with a convolution kernel of 3. The mask (outmap) is obtained by convolution and Sigmoid mapping of the refined features (refine2), and is passed to the next semantic guidance module together with the refined features (refine2). After decoding by three semantic guidance modules, the final predicted change map is obtained.

[0105] The semantic guidance decoding process in step E is to perform decoding operations on D1, D2, D3, and D4 in step E1.3, and perform decoding refinement layer by layer through the semantic guidance module, specifically including the following steps:

[0106] Step E2.1, first, use semantic mask to decouple the current layer features and distinguish the foreground subject and background in the changed area;

[0107] Step E2.2, then, the features are processed at multiple scales by convolution kernels of different scales to extract the contextual information of the foreground and background;

[0108] Step E2.3, then, upsample the high-level features and align them with the current-level features, and compare them with the foreground features to eliminate irrelevant noise, so that the foreground features focus on the changing subject;

[0109] Step E2.4, then, compare the filtered foreground features with the background features, and use the background information to supplement the missing details in the foreground; the use of the loss function, this application uses the cross entropy loss commonly used in binary classification tasks and a loss that can alleviate background imbalance in samples Combined to form a mixed loss L = L bce +L dice To calculate the loss between the predicted change map of the bi-temporal image of the training sample obtained in step H and the corresponding label map, where y i,j Indicates the probability that the pixel (i, j) in the corresponding label image is a changed pixel, It indicates the probability that the pixel (i, j) in the predicted change map is a changed pixel, and n and m represent the width and height of the image pixel level respectively;

[0110] Step E2.5: Finally, the mask and refined features obtained after convolution processing are passed to the next semantic guidance module, and decoded by three semantic guidance modules to generate the final predicted change map.

[0111] After the operation of step AE, the present invention saves all parameter information of the trained model after the training is completed, and the establishment of the change detection model can be realized. Subsequently, the remote sensing images of the before and after phases to be detected can be sequentially subjected to orthorectification, image registration, image stretching and image numerical normalization preprocessing, and then cut into the same size to obtain the samples to be detected; the samples to be tested are then input into the obtained change detection model, and the detection result graph is calculated and output.

[0112] In order to solve the problem that the model trained by the existing change detection method extracts incomplete change areas and irregular edges of change targets, this patent invention combines two methods: dual-temporal feature spatiotemporal differential fusion and semantic-guided refined multi-scale feature fusion. The spatiotemporal differential fusion module focuses on the integration of cross-temporal features, combines channel-space attention, and performs multi-dimensional perception of change areas to help the model improve its perception of changes. The latter guides the combination of features of different scales through deep semantic information, and the added refinement strategy makes the expression of change areas clearer and more complete. The semantic guidance module helps to enhance the information expression of change areas and suppress redundant information in the background by introducing semantic information in the feature fusion process. This method not only combines the multi-level spatial and semantic information contained in features of different scales, but also realizes deep-level synergy between features through semantic guidance, effectively improving the network model's perception of change areas.

[0113] Most change detection algorithms do not fully extract the overall features of the changed area and lack attention to the edge details of the changed target. To solve this problem, the present invention establishes a network architecture based on change perception and semantic guidance refinement (such as Figure 2 ). It is an encoder-decoder architecture, which consists of an encoder, a spatiotemporal differential fusion module (TDFM), a semantic aggregation module (SAM), and a semantic guidance module (SGM). The encoder is a ResNet-18 pre-trained model. The spatiotemporal differential fusion module performs dual-temporal feature fusion on the dual-temporal features output by the encoder module and outputs differential features (such as Figure 3 ); the semantic aggregation module aggregates deep features to obtain semantic masks (such as Figure 4 ); use it as prior information to guide the fusion of multi-scale features, and use the semantic guidance module to perform semantic guidance fusion on the difference features step by step (such as Figure 5 ), and obtain the final predicted change graph.

[0114] Specifically, the optical remote sensing image change detection system based on change perception and semantic guidance described in the present application includes a preprocessing module for standardizing the input dual-temporal optical remote sensing image; an encoder module for extracting multi-scale features of the dual-temporal image; a spatiotemporal difference fusion module for performing difference fusion on the multi-scale dual-temporal features output by the encoder; a semantic aggregation module for aggregating deep difference features to obtain a semantic mask; and a semantic guidance module for guiding the fusion and decoding operations on adjacent scale difference features.

[0115] After preprocessing to ensure the alignment and consistency of the bi-temporal data, the bi-temporal remote sensing images are input into the encoder module. The encoder adopts a parameter-sharing ResNet-18 structure to extract multi-scale features from the bi-temporal images layer by layer to form the basis for difference fusion. Then, the bi-temporal features of the same scale are differentially fused scale by scale in the spatiotemporal difference fusion module to capture spatiotemporal change information and generate multi-scale difference features. Next, the semantic aggregation module aggregates the high-level difference features to extract global semantic information and generate semantic masks to provide semantic guidance for the subsequent semantic guidance module. Finally, the semantic guidance module guides the fusion and decoding of the multi-scale difference features layer by layer during the decoding process and outputs the final change detection result map.

[0116] The embodiments of the present application provide an optical remote sensing image change detection method based on change perception and semantic guidance, a system, a storage medium, and an electronic device that can implement the steps and processes of any of the above-mentioned optical remote sensing image change detection method embodiments based on change perception and semantic guidance, and achieve the same technical effects, which will not be described one by one here.

[0117] A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the device where the computer-readable storage medium is located executes the optical remote sensing image change detection method based on change perception and semantic guidance as described above. The computer program includes computer program code, which may be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory and other memory.

[0118] An electronic device comprises: a memory and a processor, wherein the memory stores a program that can be run on the processor, and when the processor executes the program, the optical remote sensing image change detection method based on change perception and semantic guidance as described above is implemented.

[0119] If the module / unit integrated in the electronic device described in this application is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware devices through a computer program. The computer program can be stored in a computer-readable storage medium, and when the computer program is executed by the processor, the steps of each of the above-mentioned method embodiments can be implemented.

[0120] Furthermore, the computer-readable storage medium may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function, etc.; the data storage area may store data created according to the use of the blockchain node, etc.

[0121] The present invention designs three groups of experiments: the first group of experiments is conducted on the building change detection dataset LEVIR-CD dataset [4]; the second group of experiments is conducted on the WHU-CD dataset [5], which mainly contains changes in building classes; the third group of experiments is conducted on the SYSU-CD dataset [6], which contains multiple types of changes. In order to verify the effectiveness of the proposed CASGNet, the following eight state-of-the-art remote sensing image change detection methods are selected for comparison with the present invention method, and they are briefly introduced.

[0122] Fully-Convolutional-Early-Fusion (FC-EF): FC-EF[1] is proposed based on the U-Net architecture, in which the bi-temporal images are cascaded into multi-band images for input, and skip connections are used to gradually transfer multi-scale features from the encoder to the decoder to recover spatial information. Fully Convolutional-Siamese-Concatenation (FC-Siam-conc)[1]: As a variant of the FC-EF model, FC-Siam-conc uses a Siamese encoder to extract features of the bi-temporal images, and then connects the features of the same level from the encoder to the decoder. Fully Convolutional-Siamese-Difference (FC-Siam-diff) is different from FC-Siam-conc. The skip connection of FC-Siam-diff[1] is another type of FC-EF model, which transmits the absolute difference between the bi-temporal features. BIT[7] represents the bi-temporal image as a number of tokens, and then through the self-attention mechanism, the model is able to learn the correlation between different regions in the image. Changeformer[8] uses weight-sharing Segformer-B4 to extract multi-layer features and applies differential and convolution operations at each layer to achieve the fusion of bi-temporal features. The differential features from each layer are concatenated and input into a decoder consisting of fully connected layers to complete the change detection task. SiameseNestedUNet-Change Detection (SNUNet-CD): SNUNet-CD[9] combines the Siamese network with the UNet++ network. First, the bi-temporal image is input into the two branches of the Siamese network, then the feature maps obtained from the two branches are concatenated and input into UNet++, and finally the Ensemble Channel Attention Module (ECAM) is used to fuse the feature maps obtained from the backbone network at multiple semantic levels, thereby suppressing positioning errors and semantic gaps. CGNet

[10] guides hierarchical multi-scale feature fusion through deep prior feature maps and uses a self-attention mechanism to capture the long-distance dependencies between pixels, effectively enhancing the expressive power of change features. DMINet

[11] integrates self-attention and cross-attention mechanisms through the JointAtt module to solve the problem of category imbalance and interference. The model uses differential features and multi-level aggregation to improve performance while simplifying the architecture.

[0123] Table I is a comparative experiment conducted on the LEVIR-CD dataset, Table II is a comparative experiment conducted on the WHU-CD dataset, and Table III is a comparative experiment conducted on the SYSU-CD dataset. Precision (Precision, P), recall (Recall, R), F1 score (F1 Score, F1), and intersection over union (Intersection over Union, IoU) are used to quantitatively evaluate the performance of the methods involved. The calculation of these indicators is as follows:

[0124]

[0125] Among them, true positive (TP) represents the number of correctly detected non-changed pixels, false positive (FP) represents the number of non-changed pixels that were not predicted, and false negative (FN) represents the number of changed pixels that were not predicted. Precision represents the probability that all detected pixels have changed. Recall represents the probability that all changed pixels have been correctly detected. F1 is the harmonic mean of recall and recall, which balances the conflict by considering both recall and recall at the same time. IoU is the overlap area between predicted changed pixels and changed pixels divided by the union area between them.

[0126] It can be seen from the data in Tables I to III that the scheme of the present invention improves at least 1.12% F1 and 1.07% IoU, 1.52% F1 and 2.17% IoU, and 1.86% F1 and 2.05% IoU on the LEVIR-CD dataset, WHU-CD dataset, and SYSU-CD dataset, respectively, compared with other existing remote sensing image change detection methods. The scheme of the present invention has achieved the best performance on the three public datasets and can more completely identify the changed areas. The change detection method based on change perception and semantic guidance proposed in this patent innovatively uses channel-spatial attention for comprehensive perception of cross-temporal features through the multi-scale feature fusion that combines spatiotemporal difference fusion with semantic guidance, thereby improving the detection ability of the changed area. At the same time, the collaborative fusion of features of different scales is guided by deep semantic information, and a refinement strategy is introduced to make the change expression clearer and more complete, and suppress background redundant information. This method effectively integrates multi-level spatial and semantic information, and enhances the model's perception and expression ability of the changed area.

[0127] Table I Comparative experiments on the LEVIR-CD dataset

[0128]

[0129] Table II Comparative experiments on the WHU-CD dataset

[0130]

[0131] Table III Comparative experiments on the SYSU-CD dataset

[0132]

[0133] The references in the invention patent are as follows:

[0134] [1]Daudt RC, Le Saux B, Boulch A.Fully convolutional siamese networks for change detection[C] / / 2018 25th IEEE International Conference onImageProcessing(ICIP),2018:4063-4067.

[0135] [2] Ronneberger O, Fischer P, Brox TU-net: Convolutional networks for biomedical image segmentation[C] / / 2015 Medical Image Computing and Computer-Assisted Intervention(MICCAI), 2015.

[0136] [3]Li Q, Zhong R, Du

[0137] [4]Chen H, Shi ZA spatial-temporal attention-aased method and a newdataset for remote sensing image Change Detection[J].Remote Sensing, 2020,12(10):1662.

[0138] [5]Ji S,Wei S,Lu M.Fully convolutional networks for multisourcebuilding extraction from an open aerial and satellite imagery dataset[J].IEEETransactions on Geoscience and Remote Sensing,2019,57(1):574-586.

[0139] [6]Shi Q,Liu M,Li S,et al.A deeply supervised attention metric-basednetwork and an open aerial image dataset for remote sensing change detection[J].IEEE Transactions on Geoscience and Remote Sensing,2021,60:1-16.

[0140] [7]Chen H,Qi Z,Shi Z.Remote sensing image change detection withtransformers[J].IEEE Transactions on Geoscience and Remote Sensing,2021,60:1-14.

[0141] [8]Bandara W G C,Patel VM.A transformer-based siamese network forchange detection[C] / / 2022 IEEE International Geoscience and Remote SensingSymposium(IGARSS),2022.

[0142] [9]Fang S,Li K,Shao J,et al.SNUNet-CD:A densely connected siamesenetwork for change detection of VHR images[J].IEEE Geoscience and RemoteSensing Letters,2021,19:1-5.

[0143]

[10] Han C, Wu C, Guo H, et al. Change guiding network: Incorporating change prior to guide change detection in remote sensing imagery [J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2023, 8395-8407.

[0144]

[11] Feng Y,Jiang J,Xu H,et al.Change detection on remote sensingimages using dual-branch multilevel intertemporal network[J].IEEETransactions on Geoscience and Remote Sensing,2023,61:1-15.

[0145] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only illustrative, for example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation.

[0146] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0147] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be construed as limiting the scope of the patent application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent application shall be subject to the attached claims.

[0148] In the description of the present invention, it should be noted that directional words, such as the terms "center", "lateral", "longitudinal", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like, indicating directions and positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and cannot be understood as limiting the specific protection scope of the present invention.

[0149] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchanged in appropriate circumstances, so that the embodiments of the present application described here. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, the process, method, system, product or equipment comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or equipment.

[0150] Note that the above are only preferred embodiments of the present invention and the principles of the application technology. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention is described in more detail through the above embodiments, the present invention is not limited to the specific embodiments described herein, and may include more other effective embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A change detection method for optical remote sensing images based on change perception and semantic guidance, characterized in that: It includes the following steps: Step A, performing orthorectification, image registration, image stretching and image numerical normalization preprocessing on the bi-temporal optical remote sensing image in sequence, so as to obtain a bi-temporal optical remote sensing image with consistent data distribution; Step B: annotating the updated part of the preprocessed dual-temporal optical remote sensing image obtained in the previous step to obtain a corresponding change label map; Step C, cutting the change label map obtained in step B and the bi-temporal optical remote sensing image preprocessed in step A into the same size to obtain training samples; Step D, create a network model: Create a change detection network model based on change perception and semantic guidance, which is a dual encoder-single decoder architecture. First, use the pre-trained ResNet-18 model as the feature encoder to extract the multi-scale features of the dual-temporal images respectively; then, the proposed spatiotemporal difference fusion module performs scale-by-scale difference fusion on these multi-scale features to capture spatiotemporal change information at different levels; then, the semantic aggregation module aggregates the fused high-level difference features to generate a semantic mask to provide more accurate semantic guidance information for subsequent decoding; finally, the semantic guidance module is used as a decoder to decode the fused difference features layer by layer to generate the final change detection result; Step E: training the network model to obtain the optimal network model, which is used to detect the remote sensing images before and after the detection, and output the detection results: input a pair of dual-temporal optical remote sensing images in the training sample into the network model created in step D, and sequentially undergo feature encoding, spatiotemporal differential fusion, semantic aggregation and semantic-guided decoding to obtain the final predicted change map; then combine the binary cross entropy loss with the Dice loss to form a hybrid loss function to calculate the loss between the predicted change map of the dual-temporal image of the training sample and the corresponding label map, minimize the hybrid loss function through the back propagation algorithm, continuously update the model parameters, and save the optimal model.

2. The optical remote sensing image change detection method based on change perception and semantic guidance according to claim 1 is characterized in that: The encoding and fusion process in step E specifically includes the following steps: Step E1.1, first encode through the weight-sharing ResNet-18 pre-trained model, and extract multi-scale feature maps X1, X2, X3, X4, Y1, Y2, Y3, Y4 respectively; where X i Represents the multi-scale features of phase 1, Y i Represents multi-scale features of phase 2; Step E1.2: Input the bi-temporal feature map of the same scale obtained in the previous step into the spatiotemporal difference fusion module for bi-temporal feature fusion; first, calculate the adaptive weight of the bi-temporal feature through the channel and spatial attention mechanism, then introduce the bi-temporal difference feature to obtain the difference weight through convolution learning, and finally jointly weight the bi-temporal features to capture the change area through multiple dimensions, improve the integrity and detail description of the change area, and finally obtain four different scale difference features D1, D2, D3, D4; Step E1.3: The semantic aggregation module is used to semantically aggregate the deep three-layer features D2, D3, and D4 to generate a semantic change mask.

3. The optical remote sensing image change detection method based on change perception and semantic guidance according to claim 1 is characterized in that: The semantic guidance decoding process in step E is to perform decoding operations on D1, D2, D3, and D4 in step E1.3, and perform decoding refinement layer by layer through the semantic guidance module, specifically including the following steps: Step E2.1, first, use semantic mask to decouple the current layer features and distinguish the foreground subject and background in the changed area; Step E2.2, then, the features are processed at multiple scales by convolution kernels of different scales to extract the contextual information of the foreground and background; Step E2.3, then, upsample the high-level features and align them with the current-level features, and compare them with the foreground features to eliminate irrelevant noise, so that the foreground features focus on the changing subject; Step E2.4, after that, the filtered foreground features are compared with the background features, and the missing details in the foreground are supplemented with the background information; Step E2.5: Finally, the mask and refined features obtained after convolution processing are passed to the next semantic guidance module, and decoded by three semantic guidance modules to generate the final predicted change map.

4. The optical remote sensing image change detection method based on change guidance and refinement according to claim 1, characterized in that: The bi-temporal fusion operation in step E1.2 specifically includes the following steps: First, perform maximum pooling and average pooling of space and channels on the bi-temporal features to capture information in different regions and levels: S c =Concat(Avg(T1),Max(T1),Avg(T2),Max(T2)) S c represents the aggregated spatial features, T1 and T2 represent the bi-temporal features, Avg(·) and Max(·) represent the global average pooling and global maximum pooling across the spatial dimensions, respectively. The aggregated spatial features are passed to two one-dimensional convolutions, which are learned to determine the bi-temporal channel weights of the input bi-temporal features. The two channel weights are represented as W c1 , W c2 : W c1 =Conv1(S c ) W c2 =Conv2(S c ) Then use Softmax in the bi-temporal channel weights to make their sum equal to 1. The Softmax method can express the following process: Where W' c1 and W' c2 Denotes the output bi-temporal channel weight; the bi-temporal spatial weight W' is determined in the same way in the spatial branch s1 and W' s2 , thereby determining the important part of the dual-temporal features in the spatial dimension; adding the dual-temporal channel weight and the dual-temporal space weight to obtain the dual-temporal weight W' sc1 , W' sc2 , IN' sc1 =W' c1 +W' s1 IN' sc2 =W' c2 +W' s2 Finally, the bitemporal features are differentiated to obtain the change information, and the difference weight W in the change information is obtained through 3×3 convolution learning and Sigmoid mapping. diff : W diff =Sigmoid(Conv 3×3 (|T1-T2|)) The final difference feature is obtained by combining the adaptation weight and the difference weight to weight the original feature D i =W diff ×(W' sc1 )×T1+W diff ×(W' sc2 )×T2 Among them, i=1, 2, 3, 4; i corresponds to different scales of difference features.

5. The optical remote sensing image change detection method based on change perception and semantic guidance according to claim 1, characterized in that: The step E1.3 specifically includes the following steps: inputting the obtained deep difference feature maps D2, D3, and D4 into the semantic aggregation module, multiplying the current features with the deeper features layer by layer, which is equivalent to introducing a multi-level semantic filter, each layer further filters the noise, and only retains the meaningful information consistent with the deep features; finally, by connecting and fusing the features of each level, complementing each other, a richer and more complete semantic representation is generated; the process is as follows: D'3=Conv(up(D4))×D3 D'2=Conv(up(D3'))×D2 D s =Conv(Concat(up(Conv((Concat(up(D4),D'3))),D'2)) mask=Sigmoid(D s ) Among them, D4, D3, and D2 represent the last three layers of multi-scale difference features, up represents double upsampling, Conv represents a convolution block with a convolution kernel of 3, Concat represents concatenation in the channel dimension, D'3 and D'2 represent feature maps after filtering in the previous layer, and D s represents the semantic features obtained by aggregation, and mask represents the obtained semantic mask.

6. The optical remote sensing image change detection method based on change perception and semantic guidance according to claim 1, characterized in that: The semantic mask obtained in step E1.3 is input into the semantic guidance module together with the difference features of the two adjacent scales for fusion decoding. Specifically, the semantic mask mask and its inverse mask (1-mask) are first used to decouple the current layer feature current, and the foreground subject of the changed area is distinguished from the background, and F is obtained. f , F b ; F fn =ContextExtract(F f ) F bn =ContextExtract(F b ) Next, the features are extracted at multiple scales using convolution kernels k of different scales, where k = 1, 3, 5, 7, respectively, to extract the context information of foreground features and background features to obtain F fn , F bn ; Then, the high-level feature high is upsampled and aligned with the current layer feature current scale, and the foreground feature F fn Subtract to eliminate irrelevant noise and ensure that the foreground feature focuses on the changing subject to obtain refine1; then refine1 is combined with the background feature F bn Add, bn+relu, i.e. batch normalization and Relu activation function, to eliminate the generated negative values; use background information to supplement the missing details of the change area in the foreground to obtain the final refined features; α and β are two self-learning parameters to control the amplitude of elimination and supplementation; refine1=bn(relu(up(high)-α×F fn )) refine2=bn(relu(refine1+β×F bn )) outmap=Sigmoid(Conv(refine2)) Among them, Conv represents a convolution block with a convolution kernel of 3. The mask outmap is obtained by convolution and Sigmoid mapping the refined feature refine2, and is passed to the next semantic guidance module together with the refined feature refine2. The final predicted change map is obtained after decoding by three semantic guidance modules.

7. An optical remote sensing image change detection system based on change perception and semantic guidance, characterized in that: It includes a preprocessing module for standardizing the input bi-temporal optical remote sensing image; an encoder module for extracting multi-scale features of the bi-temporal image; and a spatiotemporal differential fusion module for differentially fusing the multi-scale bi-temporal features output by the encoder; The semantic aggregation module is used to aggregate deep difference features to obtain semantic masks; The semantic guidance module is used to guide the fusion and decoding operations of adjacent scale difference features; After preprocessing to ensure the alignment and consistency of the bi-temporal data, the bi-temporal remote sensing images are input into the encoder module. The encoder adopts a parameter-sharing ResNet-18 structure to extract multi-scale features of the bi-temporal images layer by layer to form the basis for difference fusion. Then, the bi-temporal features of the same scale are differentially fused scale by scale in the spatiotemporal difference fusion module to capture the spatiotemporal change information and generate multi-scale difference features. Next, the semantic aggregation module aggregates the high-level difference features to extract global semantic information and generate semantic masks to provide semantic guidance for the subsequent semantic guidance module. Finally, the semantic guidance module guides the fusion and decoding of the multi-scale difference features layer by layer during the decoding process and outputs the final change detection result map.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the device where the computer-readable storage medium is located executes the optical remote sensing image change detection method based on change perception and semantic guidance as described in any one of claims 1 to 6.

9. An electronic device, characterized in that: include: A memory and a processor, wherein the memory stores a program that can be run on the processor, and when the processor executes the program, the optical remote sensing image change detection method based on change perception and semantic guidance as described in any one of claims 1 to 6 is implemented.

Images