A change detection method of remote sensing image

Abstract

This application relates to the interdisciplinary field of remote sensing image processing and computer vision, and discloses a method for detecting changes in remote sensing images. The method includes: obtaining bottom-level pixel features of a first-temporal remote sensing image and a second-temporal remote sensing image; obtaining enhanced visual features based on the first-temporal and second-temporal bottom-level pixel features; obtaining semantically guided features based on a first-pixel-text similarity score map and a second-pixel-text similarity score map; obtaining multi-level fusion features based on the enhanced visual features and semantically guided features; performing pixel-level binary classification on the multi-level fusion features, and outputting a change detection binary map. This application improves the accuracy and precision of change detection in remote sensing images.

Description

Technical Field

[0001] This application relates to the interdisciplinary fields of remote sensing image processing and computer vision, and in particular to a method for detecting changes in remote sensing images. Background Technology

[0002] Remote sensing image change detection identifies land surface change patterns by comparing observational data of the same area at different time phases, and has been widely applied in fields such as natural resource monitoring, disaster prevention and mitigation, and urban dynamic analysis. With the development of deep learning technology, Siamese network architectures based on Convolutional Neural Networks (CNNs) and Transformers (deep learning architectures based on self-attention mechanisms) have become the mainstream methods.

[0003] However, despite the progress made in feature extraction, existing technologies still face the following serious challenges when dealing with complex scenes, which limit further improvements in detection accuracy.

[0004] Lacking collaborative modeling of differential features and background information, existing difference-based methods typically perform direct differencing on bi-temporal features to highlight areas of change. This logic focuses primarily on areas that have changed, neglecting areas that have not. Due to the lack of explicit modeling of background consistency features and interaction suppression mechanisms between differential and background features, the model struggles to utilize background information to filter environmental noise, leading to a high likelihood of false alarms in complex scenes.

[0005] Visual features are often mixed with semantic information, lacking refined semantic guidance. Existing methods typically extract visual features that are a mixture of low-level spatial texture information and high-level semantic information. Without external, purely semantic guidance, networks struggle to distinguish between visual differences and semantic changes. Current technologies also struggle to generate pixel-level semantic confidence distributions as prior knowledge, making it impossible to achieve refined change perception at the semantic level and leading to a problem of both false positives and false negatives.

[0006] The depth of multimodal fusion is insufficient. Although some recent studies have attempted to introduce vision-language pre-trained models to assist detection, their fusion methods mostly remain at a shallow level, such as feature splicing. Existing technologies lack a mechanism to align and deeply fuse textual semantic features with visual features at multiple scales of the feature pyramid, which limits the complementary effect of multimodal information at different resolution levels and makes it difficult to accurately capture regions of subtle changes. Summary of the Invention

[0007] In view of this, this application provides a change detection method for remote sensing images. It utilizes the CLIP-ViT pre-trained model to introduce prior textual knowledge, and uses the dual-stream attention module (DAM) to explicitly separate and cross-fuse differential and consistency features, thereby achieving effective separation of semantic and spatial information. Furthermore, it performs multi-scale fusion and decoding based on the enhanced features after collaborative modeling, thus achieving high-precision change detection.

[0008] This application discloses a method for detecting changes in remote sensing images, which includes:

[0009] Step 1: Based on the first phase remote sensing image Second-phase remote sensing images The first phase low-level pixel features are obtained. Second phase bottom layer pixel features Based on the first and second text prompt lexicons, the first pixel-text similarity score map is obtained. Second pixel-text similarity score map The first and second text prompt vocabularies are both sets of text descriptions constructed based on preset remote sensing land cover categories, used to represent the semantic category information of land covers; the first temporal remote sensing image Second-phase remote sensing images It belongs to dual-temporal imagery;

[0010] Step 2: Based on the bottom pixel features of the first time phase Second phase bottom layer pixel features Enhanced visual features ;

[0011] Step 3: Based on the first pixel-text similarity score map Second pixel-text similarity score map Semantic guidance features are obtained. And obtain the total loss function to optimize the semantically guided features. ;

[0012] Step 4: Based on the enhanced visual features and semantic guidance features This yields fusion features at multiple levels;

[0013] Step 5: Perform pixel-level binary classification on the fusion features of multiple levels and output a change detection binary map; the change detection binary map is used to represent the final result map of the spatial distribution of land cover changes between the first and second temporal remote sensing images; each pixel value in the change detection binary map corresponds to a binary state of the land surface spatial location, and the binary state includes a first state and a second state; wherein, the first state represents the area where land cover has changed, and the second state represents the background area where land cover has not changed.

[0014] Further, step 1 includes:

[0015] Step 11: Load the CLIP-ViT pre-trained model, fix the text encoder parameters to retain prior knowledge of the text and images, and only adjust the image encoder to adapt to the first-phase remote sensing image. Second-phase remote sensing images The feature distribution; the CLIP-ViT pre-trained model includes the text encoder and the image encoder;

[0016] Step 12: Transfer the first temporal remote sensing image Second-phase remote sensing images The image encoder is input to extract the first temporal low-level pixel features respectively. Second phase bottom layer pixel features Among them, the bottom-layer pixel features of the first time phase Second phase bottom layer pixel features All contain a Transformer global token, which can represent global information and ensure the underlying pixel features of the first phase. Second phase bottom layer pixel features The feature dimensions are unified;

[0017] Step 13: Input the first text prompt dictionary and the second text prompt dictionary into the text encoder to generate a text feature matrix containing semantic information about land cover categories. ; Calculate the text feature matrix respectively Compared with the bottom pixel features of the first phase Second phase bottom layer pixel features The similarity between them is used to obtain the first pixel-text similarity score map. Second pixel-text similarity score map To reflect the first temporal remote sensing image The semantic attribution confidence distribution of each pixel in the first time phase, and the remote sensing image in the second time phase. The semantic attribution confidence distribution of each pixel in the second time phase.

[0018] Further, in step 13, the first pixel-text similarity score map is calculated using the following formula. Second pixel-text similarity score map :

[0019]

[0020] in, For cosine similarity, For the first Temporal underlying pixel features, The phase is 1 or 2.

[0021] Further, step 2 includes:

[0022] Step 21: Based on the first phase low-level pixel features Second phase bottom layer pixel features The difference features are obtained through difference and summation operations. and consistency features ;

[0023] Step 22: Difference features The input difference branch first extracts the difference features through a convolutional layer. The local spatial context information is then used to perform global average pooling (AvgPool) and global max pooling (MaxPool) to obtain differential features. Global statistical information; consistency characteristics Input the consistency branch, and extract the consistency features using global average pooling and local feature pooling respectively. Statistical information;

[0024] Step 23: Generate the weights of the difference branch and the consistency branch through the channel compression-expansion mechanism, that is, the difference features. and consistency features The feature dimensions are first compressed and then expanded, combined with the Sigmoid activation function. Output the weights of the difference branch and the consistency branch; utilize the difference features obtained in step 22. Global statistical information and consistency characteristics The statistical information is used to generate difference branch weights and consistency branch weights through a channel compression-expansion mechanism; that is, the feature dimensions of the statistical information are first compressed and then expanded, combined with the Sigmoid activation function. Output the final weights of the difference branches and the consistency branches;

[0025] Step 24: Employ a cross-integration strategy to combine the different features With consistency features By combining the weights of each party separately, the enhanced differential features are obtained. and enhanced consistency features Enhanced differential features and enhanced consistency features The images are stitched together and then fused through convolutional layers to output enhanced visual features. .

[0026] Further, step 21 includes:

[0027] The differential characteristics are obtained using the following formula. and consistency features :

[0028]

[0029]

[0030] in, To determine the sign of the absolute value, the difference feature Used to capture pixel-level change information, consistency features Used to characterize first-temporal remote sensing images Second-phase remote sensing images Shared background information enables explicit separation of changing and invariant features;

[0031] Step 23 includes:

[0032] The weights of the divergent branches are obtained using the following formula. :

[0033]

[0034] in, It is a multilayer perceptron. For global average pooling, For global max pooling, This is a convolution operation;

[0035] The weight of the consensus branch is obtained using the following formula. :

[0036]

[0037] in, Pooling for local features;

[0038] Step 24 includes:

[0039] The enhanced differential features are obtained using the following formula. and enhanced consistency features :

[0040]

[0041]

[0042] in, For element-wise multiplication;

[0043] The enhanced visual features are obtained using the following formula. :

[0044]

[0045] in, This is a one-dimensional convolution operation.

[0046] Further, step 3 includes:

[0047] Step 31: Based on the first pixel-text similarity score map Second pixel-text similarity score map Semantic change features are obtained. Semantic consistency features ;

[0048] Step 32: Semantic change features Semantic consistency features Channel attention ECA is applied separately, channel statistics are obtained through global average pooling (GAP), and attention weights are generated by combining one-dimensional convolutional Conv1D with adaptive kernel size to obtain semantic change features after channel enhancement. Semantic consistency features enhanced by channel ;

[0049] Step 33: Based on the complementarity of changing and invariant information, through learnable fusion coefficients A fusion gate is constructed to perform weighted fusion of the two enhanced semantic features, outputting semantic guidance features. ;

[0050] Step 34: Obtain the total loss function to constrain the boundary accuracy of the changing region and optimize the semantically guided features. The generation of .

[0051] Further, step 31 includes:

[0052] The semantic change features are obtained using the following formula. Semantic consistency features :

[0053]

[0054]

[0055] in, This is the first pixel-text similarity score map. The second pixel-text similarity score map; semantic variation features Used to highlight the true change areas in semantic category transformation, semantic consistency features Used to preserve first-time remote sensing images Second-phase remote sensing images Features of the background-invariant region;

[0056] Step 32 includes:

[0057] The semantic change features after channel enhancement are obtained using the following formula. Semantic consistency features enhanced by channel :

[0058]

[0059]

[0060] in, For element-wise multiplication, For a one-dimensional convolution with kernel size k, For global average pooling, the kernel size is... Based on the number of channels Adaptive determination:

[0061]

[0062] In the formula, This is the channel ratio parameter. For bias parameters, To take the nearest odd number;

[0063] Step 33 includes:

[0064] The semantic guidance features are obtained using the following formula. :

[0065]

[0066] in, These are learnable fusion coefficients, forming a fusion gate used to enhance the semantic change features of the channels. Semantic consistency features enhanced by channel Perform adaptive weighted fusion;

[0067] Step 34 includes:

[0068] The total loss function is obtained using the following formula:

[0069]

[0070] in, For the total loss function, To mitigate losses, Main loss, This is the loss weighting coefficient.

[0071] Further, step 4 includes:

[0072] Step 41: Enhance the visual features With semantic guidance features Matching is performed at multiple scales, and spatial resolution is adjusted through interpolation or downsampling to ensure enhanced visual features at the same scale. With semantic guidance features The dimensions and size are consistent;

[0073] Step 42: Input the features matched in Step 41 into the Feature Pyramid Network (FPN), and stitch together feature maps from different levels through top-down upsampling and bottom-up lateral connections.

[0074] Step 43: Perform convolution compression and activation processing on each layer of the spliced ​​feature map to output fused features of multiple layers; the number of layers is the same as the number of scales.

[0075] Further, step 5 includes:

[0076] Step 51: Input the fused features from multiple levels into the decoder for pixel-level binary classification and output a change detection binary map; the decoder consists of multiple convolutional layers and a Sigmoid activation function.

[0077] Furthermore, after step 5, the method further includes:

[0078] The detection performance is evaluated using multi-dimensional evaluation indicators, including precision, recall, F1 score, overall accuracy (OA), and intersection-over-union (IoU).

[0079] Due to the adoption of the above technical solution, this application has the following advantages:

[0080] 1. This application calculates pixel-text similarity score maps based on the CLIP-ViT pre-trained model. It utilizes the natural invariance of high-level semantic features (which start from generating similarity score maps from the text prompt vocabulary and are specifically applied in the semantic guidance enhancement module for refined processing and fusion) to suppress pseudo-changes, significantly reduce the false detection rate, and has strong anti-interference ability.

[0081] 2. The Dual Stream Attention (DAM) module of this application is the first to achieve explicit separation and cross-complementation of differential features and background consistency features (explicit separation generates differential and consistency features through mathematical operations; cross-complementation enhances the two types of features through a cross-fusion strategy, with differential features modulated by the attention weight of the background branch and consistency features modulated by the attention weight of the differential branch). It effectively filters noise by utilizing background information, making the model more accurate in locating changing regions in complex scenes and improving feature discrimination.

[0082] 3. The semantic guidance enhancement module (SGM) of this application (which outputs semantic guidance features based on the first pixel-text similarity score map and the second pixel-text similarity score map) combined with the auxiliary supervision mechanism (obtaining the total loss function to constrain the boundary accuracy of the change region and optimize the generation of semantic guidance features) enhances the capture of semantic category transformation. Attached Figure Description

[0083] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments recorded in the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings.

[0084] Figure 1 This is a flowchart illustrating a method for detecting changes in remote sensing images according to an embodiment of this application.

[0085] Figure 2 This is a schematic diagram of the dual-stream attention module (DAM) according to an embodiment of this application;

[0086] Figure 3 This is a schematic diagram of the semantic guidance enhancement module (SGM) according to an embodiment of this application;

[0087] Figure 4 This is a schematic diagram comparing the detection results of the method of this application with other methods on the LEVIR-CD test set according to an embodiment of this application;

[0088] Figure 5 This is a schematic diagram comparing the detection results of the method of this application with other methods on the SYSU-CD test set of this application embodiment;

[0089] Figure 6 This is a feature visualization diagram of the DAM module under different channel dimensionality reduction ratios in an embodiment of this application. Detailed Implementation

[0090] The present application will be further described in conjunction with the accompanying drawings and embodiments. The described embodiments are only some, not all, of the embodiments of the present application. All other embodiments obtained by those skilled in the art should fall within the protection scope of the embodiments of the present application.

[0091] See Figure 1 This application provides an embodiment of a method for detecting changes in remote sensing images, which includes:

[0092] Step 1: Based on the first phase remote sensing image Second-phase remote sensing images The first phase low-level pixel features are obtained. Second phase bottom layer pixel features Based on the first and second text prompt lexicons, the first pixel-text similarity score map is obtained. Second pixel-text similarity score map The first and second text prompt vocabularies are both sets of text descriptions constructed based on preset remote sensing land cover categories, used to represent the semantic category information of land covers; the first temporal remote sensing image Second-phase remote sensing images It belongs to dual-temporal imagery;

[0093] Step 2: Based on the bottom pixel features of the first time phase Second phase bottom layer pixel features Enhanced visual features ;

[0094] Step 3: Based on the first pixel-text similarity score map Second pixel-text similarity score map Semantic guidance features are obtained. And obtain the total loss function to optimize the semantically guided features. ;

[0095] Step 4: Based on the enhanced visual features and semantic guidance features This yields fusion features at multiple levels;

[0096] Step 5: Perform pixel-level binary classification on the fused features of multiple levels and output a change detection binary map; the change detection binary map is used to represent the final result map of the spatial distribution of ground feature changes between the first and second time-phase remote sensing images.

[0097] In this binary change detection image, each pixel value corresponds to a binary state of the land surface spatial location, which includes a first state and a second state. The first state represents the area where the land features have changed, and the second state represents the background area where the land features have not changed. As the direct output product of the remote sensing image change detection method of this application, the binary change detection image realizes pixel-level positioning and recognition of land surface change information.

[0098] The remote sensing ground features categories include the following types: 1) Natural environment: including beaches, forests, lakes, grasslands, mountains, oceans, wetlands, farmland, deserts, rivers, and snowfields; 2) Transportation: including airports, bridges, highways, ports, railways, overpasses, intersections, and roads; 3) Recreation and sports: including basketball courts, stadiums, athletic fields, tennis courts, and golf courses; 4) Residential and building: including dense housing, detached houses, general buildings, churches, and cabins; 5) Commercial and industrial: including commercial areas, industrial areas, oil tanks, containers, and mines; 6) Other man-made structures: including automobiles, ships, airplanes, and impermeable surfaces.

[0099] Optionally, step 1 includes:

[0100] Step 11: Load the CLIP-ViT pre-trained model, fix the text encoder parameters to retain prior knowledge of the text and images, and only adjust the image encoder to adapt to the first-phase remote sensing image. Second-phase remote sensing images The feature distribution; the CLIP-ViT pre-trained model includes the text encoder and the image encoder;

[0101] Step 12: Transfer the first temporal remote sensing image Second-phase remote sensing images The image encoder is input to extract the first temporal low-level pixel features respectively. Second phase bottom layer pixel features Among them, the bottom-layer pixel features of the first time phase Second phase bottom layer pixel features All contain a Transformer global token, which can represent global information and ensure the underlying pixel features of the first phase. Second phase bottom layer pixel features The feature dimensions are unified;

[0102] Step 13: Input the first text prompt dictionary and the second text prompt dictionary into the text encoder to generate a text feature matrix containing semantic information about land cover categories. ; Calculate the text feature matrix respectively Compared with the bottom pixel features of the first phase Second phase bottom layer pixel features The similarity between them is used to obtain the first pixel-text similarity score map. Second pixel-text similarity score map To reflect the first temporal remote sensing image The semantic attribution confidence distribution of each pixel in the first time phase, and the remote sensing image in the second time phase. The semantic attribution confidence distribution of each pixel in the second time phase.

[0103] Optionally, in step 13, the first pixel-text similarity score map is calculated using the following formula. Second pixel-text similarity score map :

[0104]

[0105] in, For cosine similarity, For the first Temporal underlying pixel features, The phase is 1 or 2.

[0106] Optionally, step 2 includes:

[0107] Step 21: Based on the first phase low-level pixel features Second phase bottom layer pixel features The difference features are obtained through difference and summation operations. and consistency features ;

[0108] Step 22: Difference features The input difference branch first extracts the difference features through a convolutional layer. The local spatial context information is then used to perform global average pooling (AvgPool) and global max pooling (MaxPool) to obtain differential features. Global statistical information; consistency characteristics Input the consistency branch, and extract the consistency features using global average pooling and local feature pooling respectively. Statistical information;

[0109] Step 23: Generate the weights of the difference branch and the consistency branch through the channel compression-expansion mechanism, that is, the difference features. and consistency features The feature dimensions are first compressed and then expanded, combined with the Sigmoid activation function. Output the weights of the difference branch and the consistency branch; utilize the difference features obtained in step 22. Global statistical information and consistency characteristics The statistical information is used to generate difference branch weights and consistency branch weights through a channel compression-expansion mechanism; that is, the feature dimensions of the statistical information are first compressed and then expanded, combined with the Sigmoid activation function. Output the final weights of the difference branches and the consistency branches;

[0110] Step 24: Employ a cross-integration strategy to combine the different features With consistency features By combining the weights of each party separately, the enhanced differential features are obtained. and enhanced consistency features Enhanced differential features and enhanced consistency features The images are stitched together and then fused through convolutional layers to output enhanced visual features. .

[0111] Optionally, step 21 includes:

[0112] The differential characteristics are obtained using the following formula. and consistency features :

[0113]

[0114]

[0115] in, To determine the sign of the absolute value, the difference feature Used to capture pixel-level change information, consistency features Used to characterize first-temporal remote sensing images Second-phase remote sensing images Shared background information enables explicit separation of changing and invariant features;

[0116] Step 23 includes:

[0117] The weights of the divergent branches are obtained using the following formula. :

[0118]

[0119] in, It is a multilayer perceptron. For global average pooling, For global max pooling, This is a convolution operation;

[0120] The weight of the consensus branch is obtained using the following formula. :

[0121]

[0122] in, Pooling for local features;

[0123] Step 24 includes:

[0124] The enhanced differential features are obtained using the following formula. and enhanced consistency features :

[0125]

[0126]

[0127] in, For element-wise multiplication;

[0128] The enhanced visual features are obtained using the following formula. :

[0129]

[0130] in, This is a one-dimensional convolution operation.

[0131] Optionally, step 3 includes:

[0132] Step 31: Based on the first pixel-text similarity score map Second pixel-text similarity score map Semantic change features are obtained. Semantic consistency features ;

[0133] Step 32: Semantic change features Semantic consistency features Channel attention ECA is applied separately, channel statistics are obtained through global average pooling (GAP), and attention weights are generated by combining one-dimensional convolutional Conv1D with adaptive kernel size to obtain semantic change features after channel enhancement. Semantic consistency features enhanced by channel ;

[0134] Step 33: Based on the complementarity of changing and invariant information, through learnable fusion coefficients A fusion gate is constructed to perform weighted fusion of the two enhanced semantic features, outputting semantic guidance features. ;

[0135] Step 34: Obtain the total loss function to constrain the boundary accuracy of the changing region and optimize the semantically guided features. The generation of .

[0136] Optionally, step 31 includes:

[0137] The semantic change features are obtained using the following formula. Semantic consistency features :

[0138]

[0139]

[0140] in, This is the first pixel-text similarity score map. The second pixel-text similarity score map; semantic variation features Used to highlight the true change areas in semantic category transformation, semantic consistency features Used to preserve first-time remote sensing images Second-phase remote sensing images Features of the background-invariant region;

[0141] Step 32 includes:

[0142] The semantic change features after channel enhancement are obtained using the following formula. Semantic consistency features enhanced by channel :

[0143]

[0144]

[0145] in, For element-wise multiplication, For a one-dimensional convolution with kernel size k, For global average pooling, the kernel size is... Based on the number of channels Adaptive determination:

[0146]

[0147] In the formula, This is the channel ratio parameter. For bias parameters, To take the nearest odd number;

[0148] Step 33 includes:

[0149] The semantic guidance features are obtained using the following formula. :

[0150]

[0151] in, These are learnable fusion coefficients, forming a fusion gate used to enhance the semantic change features of the channels. Semantic consistency features enhanced by channel Perform adaptive weighted fusion;

[0152] Step 34 includes:

[0153] The total loss function is obtained using the following formula:

[0154]

[0155] in, For the total loss function, To mitigate losses, Main loss, This is the loss weighting coefficient.

[0156] Optionally, step 4 includes:

[0157] Step 41: Enhance the visual features With semantic guidance features Matching is performed at multiple scales, and spatial resolution is adjusted through interpolation or downsampling to ensure enhanced visual features at the same scale. With semantic guidance features The dimensions and size are consistent;

[0158] Step 42: Input the features matched in Step 41 into the Feature Pyramid Network (FPN), and stitch together feature maps from different levels through top-down upsampling and bottom-up lateral connections.

[0159] Step 43: Perform convolution compression and activation processing on each layer of the spliced ​​feature map to output fused features of multiple layers; the number of layers is the same as the number of scales.

[0160] Optionally, step 5 includes:

[0161] Step 51: Input the fused features from multiple levels into the decoder for pixel-level binary classification and output a change detection binary map; the decoder consists of multiple convolutional layers and a Sigmoid activation function.

[0162] Optionally, after step 5, the method further includes:

[0163] The detection performance is evaluated using multi-dimensional evaluation indicators, including precision, recall, F1 score, overall accuracy (OA), and intersection-over-union (IoU).

[0164] This application innovatively introduces the CLIP cross-modal model to construct a land cover category prompting lexicon suitable for remote sensing scenarios, and generates a pixel-text similarity score map, providing clean and robust high-level semantic guidance for change detection, thus solving the core pain point of traditional methods where semantic and spatial information are mixed.

[0165] The dual-stream attention module (DAM) designed in this application achieves explicit separation and cross-fusion of differential features and background consistency features for the first time. By strengthening feature complementarity through bidirectional information flow interaction, it effectively suppresses background noise and false detections, and improves the discrimination power of changing regions.

[0166] The semantic guidance enhancement module (SGM) proposed in this application combines channel attention (ECA), fusion gate mechanism, and auxiliary head loss constraint to accurately capture semantic change signals, significantly suppress pseudo-change interference such as illumination and shadow, and improve boundary detection accuracy.

[0167] like Figure 2 As shown, the dual-stream attention module (DAM) first receives differential features. With consistency features As input, the differential features are processed through 3 After feature extraction via convolution, global average pooling and global max pooling are combined to obtain statistical information; consistency features are further processed through global average pooling and local feature pooling to extract statistical information. These two sets of statistical information are then input into a multilayer perceptron (e.g., convolution). c / r represents the number of channels after compression, convolution. Here, c represents the number of channels, indicating the channel dimension of the input feature map; r represents the compression ratio, a hyperparameter used to control the degree of channel compression. Channel compression and restoration are performed, and the difference branch weights and consistency branch weights are generated using the Sigmoid activation function. Then, a cross-fusion strategy is employed, multiplying the difference features element-wise with the consistency branch weights and vice versa, to obtain the enhanced difference features. and enhanced consistency features Finally, and Concatenate along the channel dimension, and use 1 1. Convolutional fusion outputs enhanced visual features .

[0168] like Figure 3 As shown, the semantic guidance enhancement module uses the first pixel-text similarity score map. Second pixel-text similarity score map As input, semantic change features and semantic consistency features are first calculated. Then, both types of features are compressed into one-dimensional vectors using global average pooling, and channel interaction is performed using one-dimensional convolutions with adaptive kernel size. After sigmoid activation, channel attention weights are generated and multiplied element-wise with the original features to achieve feature enhancement. The enhanced features are then input into a fusion gate, utilizing learnable fusion coefficients. The semantic change features and semantic consistency features are weighted and summed to output the semantic guidance features. Furthermore, this module introduces an auxiliary head during the training phase to calculate the auxiliary loss between semantically guided features and the true labels, in order to supervise and optimize the generation of semantic features.

[0169] like Figure 4 As shown, the qualitative comparison results of the proposed method with other mainstream methods (such as STANet, ChangeCLIP, etc.) on the LEVIR-CD test set are presented. Each row in the figure represents a typical test scenario, from left to right: the first temporal image Img1, the second temporal image Img2, the ground truth change label GT, and the detection results of each method. The visualization results intuitively show that the prediction map of our proposed method is closest to GT, and the red false positive region FP and the green false negative region FN are significantly less than the comparison methods. Combined with quantitative experimental data, it is shown that the F1 score and IoU index of our proposed method on this benchmark dataset are superior to existing mainstream methods such as ChangeCLIP and STANet. In particular, it can be clearly seen in the visualization that our proposed method has a significant advantage in preserving the details of fine structures such as road and building edges, effectively avoiding the boundary breakage and blurring phenomena common in other methods.

[0170] like Figure 5As shown in the figure, a comparison of the detection results of the proposed method with other methods on the SYSU-CD test set is presented. The SYSU-CD test set contains more complex urban and natural integrated scenes, which places higher demands on the robustness of the algorithm. By comparing the detection results of STANet (Spatiotemporal Attention Network), BIT (Bi-temporal Image Transformer), BAN (Bilateral Attention Network), ChangeCLIP (CLIP-based change detection method), and Ours of this application, it can be seen that other methods are prone to generating more noise in complex backgrounds, while the proposed method can effectively suppress spurious changes. Experimental statistics further confirm that the proposed method also comprehensively outperforms existing mainstream methods in terms of F1 score and IoU index on the SYSU-CD dataset. This advantage is specifically manifested in the figure as follows: this method can not only accurately locate the change region, but also depict the edge details more completely and sharply when dealing with irregular targets, demonstrating its excellent detail preservation ability and feature discrimination ability in complex scenes.

[0171] like Figure 6 The image shows a feature visualization heatmap of the two-stream attention module under different channel dimensionality reduction ratios. The image compares... =4、 =8 and Feature response under three configurations: =16. Visualization results show that when... When the value is 16, the feature heatmap provides the clearest boundary localization of the changing region and exhibits the weakest background noise response; while as... As the value decreases, noise increases in the feature map and the boundaries become blurred. This figure visually verifies the parameter... When the model has a value of 16, it can balance feature compression and information preservation, thereby achieving the best feature representation capability. This provides model parameter support for the excellent F1 and IoU performance of this application on the above dataset.

[0172] Experiments show that this application outperforms existing mainstream methods (such as ChangeCLIP and STANet) in both F1 score and IoU (Intersection over Union) on benchmark datasets such as LEVIR-CD and SYSU-CD, especially in preserving details such as road and building edges. The F1 score is the harmonic mean of precision and recall; IoU is a standard method for measuring the degree of overlap between two pixel regions, typically used to determine the matching degree between predicted and ground truth bounding boxes.

[0173] Precision refers to the proportion of pixels predicted as changed regions in a change detection binary map that are also actually labeled as changed regions. Recall refers to the proportion of samples that are actually positive that are correctly predicted as positive. Overall accuracy (OA) is the proportion of all correctly classified pixels out of the total number of pixels.

[0174] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and not to limit them. Although this application has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of this application. Any modifications or equivalent substitutions that do not depart from the spirit and scope of this application should be covered within the protection scope of the claims of this application.

Claims

1. A method for detecting changes in remote sensing images, characterized in that, include: Step 1: Based on the first phase remote sensing image Second-phase remote sensing images The first phase low-level pixel features are obtained. Second phase bottom layer pixel features Based on the first and second text prompt lexicons, the first pixel-text similarity score map is obtained. Second pixel-text similarity score map The first and second text prompt vocabularies are both sets of text descriptions constructed based on preset remote sensing land cover categories, used to represent the semantic category information of land covers; the first temporal remote sensing image Second-phase remote sensing images It belongs to dual-temporal imagery; Step 2: Based on the bottom pixel features of the first time phase Second phase bottom layer pixel features Enhanced visual features ; Step 3: Based on the first pixel-text similarity score map Second pixel-text similarity score map Semantic guidance features are obtained. And obtain the total loss function to optimize the semantically guided features. ; Step 4: Based on the enhanced visual features and semantic guidance features This yields fusion features at multiple levels; Step 5: Perform pixel-level binary classification on the fused features of multiple levels and output a change detection binary map; the change detection binary map is used to represent the final result map of the spatial distribution of land cover changes between the first and second temporal remote sensing images; each pixel value in the change detection binary map corresponds to a binary state of the land surface spatial location, and the binary state includes a first state and a second state; wherein, the first state represents the area where land cover has changed, and the second state represents the background area where land cover has not changed; Step 2 includes: Step 21: Based on the first phase low-level pixel features Second phase bottom layer pixel features The difference features are obtained through difference and summation operations. and consistency features ; Step 22: Difference features The input difference branch first extracts the difference features through a convolutional layer. The local spatial context information is then used to perform global average pooling (AvgPool) and global max pooling (MaxPool) to obtain differential features. Global statistical information; consistency characteristics Input the consistency branch, and extract the consistency features using global average pooling and local feature pooling respectively. Statistical information; Step 23: Utilize the differential features obtained in Step 22 Global statistical information and consistency characteristics The statistical information is used to generate difference branch weights and consistency branch weights through a channel compression-expansion mechanism; that is, the feature dimensions of the statistical information are first compressed and then expanded, combined with the Sigmoid activation function. Output the final weights of the difference branches and the consistency branches; Step 24: Employ a cross-integration strategy to combine the different features With consistency features By combining the weights of each party separately, the enhanced differential features are obtained. and enhanced consistency features Enhanced differential features and enhanced consistency features The images are stitched together and then fused through convolutional layers to output enhanced visual features. ; Step 3 includes: Step 31: Based on the first pixel-text similarity score map Second pixel-text similarity score map Semantic change features are obtained. Semantic consistency features ; Step 32: Semantic change features Semantic consistency features Channel attention ECA is applied separately, channel statistics are obtained through global average pooling (GAP), and attention weights are generated by combining one-dimensional convolutional Conv1D with adaptive kernel size to obtain semantic change features after channel enhancement. Semantic consistency features enhanced by channel ; Step 33: Based on the complementarity of changing and invariant information, through learnable fusion coefficients A fusion gate is constructed to perform weighted fusion of the two enhanced semantic features, outputting semantic guidance features. ; Step 34: Obtain the total loss function to constrain the boundary accuracy of the changing region and optimize the semantically guided features. The generation of .

2. The method according to claim 1, characterized in that, Step 1 includes: Step 11: Load the CLIP-ViT pre-trained model, fix the text encoder parameters to retain prior knowledge of the text and images, and only adjust the image encoder to adapt to the first-phase remote sensing image. Second-phase remote sensing images The feature distribution; the CLIP-ViT pre-trained model includes the text encoder and the image encoder; Step 12: Transfer the first temporal remote sensing image Second-phase remote sensing images The image encoder is input to extract the first temporal low-level pixel features respectively. Second phase bottom layer pixel features Among them, the bottom-layer pixel features of the first time phase Second phase bottom layer pixel features All contain a Transformer global token, which can represent global information and ensure the underlying pixel features of the first phase. Second phase bottom layer pixel features The feature dimensions are unified; Step 13: Input the first text prompt dictionary and the second text prompt dictionary into the text encoder to generate a text feature matrix containing semantic information about land cover categories. ; Calculate the text feature matrix respectively Compared with the bottom pixel features of the first phase Second phase bottom layer pixel features The similarity between them is used to obtain the first pixel-text similarity score map. Second pixel-text similarity score map To reflect the first temporal remote sensing image The semantic attribution confidence distribution of each pixel in the first time phase, and the remote sensing image in the second time phase. The semantic attribution confidence distribution of each pixel in the second time phase.

3. The method according to claim 2, characterized in that, In step 13, the first pixel-text similarity score is calculated using the following formula. Second pixel-text similarity score map : in, For cosine similarity, For the first Temporal underlying pixel features, The phase is 1 or 2.

4. The method according to claim 1, characterized in that, Step 21 includes: The differential characteristics are obtained using the following formula. and consistency features : in, To determine the sign of the absolute value, the difference feature Used to capture pixel-level change information, consistency features Used to characterize first-temporal remote sensing images Second-phase remote sensing images Shared background information enables explicit separation of changing and invariant features; Step 23 includes: The weights of the divergent branches are obtained using the following formula. : in, It is a multilayer perceptron. For global average pooling, For global max pooling, This is a convolution operation; The weight of the consensus branch is obtained using the following formula. : in, Pooling for local features; Step 24 includes: The enhanced differential features are obtained using the following formula. and enhanced consistency features : in, For element-wise multiplication; The enhanced visual features are obtained using the following formula. : in, This is a one-dimensional convolution operation.

5. The method according to claim 1, characterized in that, Step 31 includes: The semantic change features are obtained using the following formula. Semantic consistency features : in, This is the first pixel-text similarity score map. The second pixel-text similarity score map; semantic variation features Used to highlight the true change areas in semantic category transformation, semantic consistency features Used to preserve first-time remote sensing images Second-phase remote sensing images Features of the background-invariant region; Step 32 includes: The semantic change features after channel enhancement are obtained using the following formula. Semantic consistency features enhanced by channel : in, For element-wise multiplication, For a one-dimensional convolution with kernel size k, For global average pooling, the kernel size is... Based on the number of channels Adaptive determination: In the formula, This is the channel ratio parameter. For bias parameters, To take the nearest odd number; Step 33 includes: The semantic guidance features are obtained using the following formula. : in, These are learnable fusion coefficients, forming a fusion gate used to enhance the semantic change features of the channels. Semantic consistency features enhanced by channel Perform adaptive weighted fusion; Step 34 includes: The total loss function is obtained using the following formula: in, For the total loss function, To mitigate losses, Main loss, This is the loss weighting coefficient.

6. The method according to claim 1, characterized in that, Step 4 includes: Step 41: Enhance the visual features With semantic guidance features Matching is performed at multiple scales, and spatial resolution is adjusted through interpolation or downsampling to ensure enhanced visual features at the same scale. With semantic guidance features The dimensions and size are consistent; Step 42: Input the features matched in Step 41 into the Feature Pyramid Network (FPN), and stitch together feature maps from different levels through top-down upsampling and bottom-up lateral connections. Step 43: Perform convolution compression and activation processing on each layer of the spliced ​​feature map to output fused features of multiple layers; the number of layers is the same as the number of scales.

7. The method according to claim 1, characterized in that, Step 5 includes: Step 51: Input the fused features from multiple levels into the decoder for pixel-level binary classification and output a change detection binary map; the decoder consists of multiple convolutional layers and a Sigmoid activation function.

8. The method according to claim 1, characterized in that, After step 5, the following is also included: The detection performance is evaluated using multi-dimensional evaluation indicators, including precision, recall, F1 score, overall accuracy (OA), and intersection-over-union (IoU).

Images