Remote sensing change detection method and system based on change decoupling enhancement

Abstract

The present application relates to the technical field of remote sensing change detection, and discloses a remote sensing change detection method and system based on change decoupling enhancement, comprising: acquiring double-time-phase remote sensing images, constructing a change decoupling enhancement network model, the model comprising a double-branch encoder, a change representation decoupling module, a semantic consistency enhancement module and a segmentation head, the double-branch encoder extracts double-time-phase features of a double-time-phase image pair at different scales, the change representation decoupling module explicitly identifies and separates unchanged components by calculating the similarity between the double-time-phase features, the semantic consistency enhancement module enhances intra-class consistency and inter-class separability by dividing the change representation into multiple semantic groups and independently refining them, and the segmentation head performs remote sensing change detection according to semantic features. The present application can improve the semantic discrimination ability of change detection, thereby improving the accuracy of remote sensing change detection.

Description

Technical Field

[0001] This invention relates to the field of remote sensing change detection technology, and in particular to a remote sensing change detection method and system based on change decoupling enhancement. Background Technology

[0002] Change detection (CD) is one of the core tasks in the field of remote sensing (RS). Remote sensing change detection (RSCD) aims to observe changes in the Earth's surface by analyzing bi-temporal or multi-temporal images of the same geographical location acquired at different times. With the development of ultra-high resolution (VHR) remote sensing imagery, change detection technology has been widely applied in many fields, including disaster assessment, urban sprawl research, land management, and environmental monitoring.

[0003] Early change detection relied primarily on manually designed features and traditional classification algorithms, such as methods based on image differences, ratios, and principal component analysis (PCA). These methods often suffer from poor robustness when faced with complex scenes and multi-source data, as they are affected by lighting, noise, and sensor variations.

[0004] With the rapid development of deep learning, techniques such as Convolutional Neural Networks (CNNs), Attention Mechanisms, and Transformers have been gradually introduced into the field of change detection. Deep learning models can automatically extract hierarchical features from data, freeing them from the limitations of manual feature design, and possess powerful nonlinear modeling capabilities, making them more adaptable to modeling spatiotemporal feature differences in complex scenes. Transformers, using self-attention, can model the global dependencies between any pixels in an image, overcoming the limitations of the local receptive field in traditional CNNs. Considering the insufficient modeling of local details, many studies have adopted a hybrid architecture combining Transformers and CNNs, thereby preserving both local texture and global semantic information.

[0005] In change detection tasks, facing strong interference from unchanged backgrounds, existing technologies include the Spatiotemporal Enhancement and Interlevel Fusion Network (SEIFNet), which achieves good results by enhancing the differences between two temporal images and fully utilizing the complementary advantages of multi-scale features. Another example is the Change Detection Network using Visual-Prompt Enhanced CLIP (CVNet), which introduces a spatial attention mechanism to enable the model to focus on semantically important key pixel regions, thereby extracting more discriminative features. To fully explore the correlation between two temporal features, there is also the Cross-temporal Attention and Adaptive Ensemble Kernel Learning Network (CA2KNet), which combines channel and spatial attention mechanisms to identify key features and utilizes temporal information to enhance the interaction and fusion of cross-temporal features. To achieve more accurate change detection, a cross-domain coarse-to-fine (CDC2F) network multi-scale architecture has been developed. This method utilizes both spatial and frequency domain information and employs a coarse-to-fine detection strategy to achieve more detailed and accurate change identification. Furthermore, inspired by human perception of change, there are building change detection models with context alignment and structure perception (CASP). By exploring context alignment strategies, the accuracy of building change detection has been further improved.

[0006] However, existing methods still have a key limitation: many change detection networks suffer from local attention bias, meaning the model overemphasizes regions with significant texture changes while neglecting global semantic consistency. This narrow focus often leads to fragmented change predictions, making it difficult to maintain semantic consistency among similar targets and more susceptible to spurious changes, thus hindering their applicability to complex urban and natural scenes. Summary of the Invention

[0007] Therefore, the technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a remote sensing change detection method and system based on change decoupling enhancement, which can improve the semantic discrimination ability of change detection and thus improve the accuracy of remote sensing change detection.

[0008] To address the aforementioned technical problems, this invention provides a remote sensing change detection method based on change decoupling enhancement, comprising: Acquire dual-temporal remote sensing images, and use remote sensing images of the same area at different times as dual-temporal image pairs; A change decoupling enhancement network model is constructed, comprising a dual-branch encoder, a change representation decoupling module, a semantic consistency enhancement module, and a segmentation head. The dual-branch encoder extracts dual-temporal features of the dual-temporal image pairs at different scales. The change representation decoupling module explicitly identifies and separates unchanged components by calculating the similarity between the dual-temporal features to obtain change representations. The semantic consistency enhancement module divides the change representations into multiple semantic groups and refines them independently to enhance intra-class consistency and inter-class separability to obtain semantic features. The segmentation head performs remote sensing change detection based on the semantic features. The change decoupling enhancement network model is trained by combining the prediction results of the segmentation head, and the dual-temporal remote sensing image to be detected is input into the trained change decoupling enhancement network model to obtain the remote sensing change detection result.

[0009] Furthermore, the change characterization decoupling module explicitly identifies and separates the unchanged components and obtains the change characterization by calculating the similarity between the two temporal features, specifically as follows: Let the dual-phase characteristic be denoted as and , , These are the features at the i-th scale corresponding to two remote sensing images of the same region at different times, where i = 1, 2, ..., N, and N is the number of scales; The change representation decoupling module includes N change representation decoupling blocks, each of which includes two branches, which respectively obtain the change representation of the dual-temporal features through decoupling; and Input the i-th change representation decoupling block, and the output of the change representation decoupling block is the change representation.

[0010] Furthermore, the change representation of the i-th change representation decoupling block output is: , In the formula, This represents the change representation of the output of the i-th change representation decoupling block. This represents a convolution operation with a kernel of size k1×k1. This indicates a splicing operation. , They are respectively , The change representation obtained by decoupling in the i-th change representation decoupling block.

[0011] Furthermore, the two branches of the i-th change characterization decoupling block respectively calculate the... and In one branch, the The calculation method is as follows: , In the formula, The i-th change represents the decoupling block. The corresponding similarity score matrix, This indicates element-wise multiplication.

[0012] Furthermore, the aforementioned The calculation method is as follows: In the i-th change representation decoupling block, reshaping Let the query vector be denoted as Reshaping Let be the key vector, denoted as Attention calculation via scaling dot product and The feature similarity map obtained from pixel-level similarity is denoted as... , The i-th change represents the decoupling block. Corresponding feature similarity map; exist The maximum value in each row is selected as the co-significance probability of each pixel in the dual-time feature, denoted as . , Represents the cosignificance probability of the k-th pixel, reshaping get .

[0013] Furthermore, the semantic consistency enhancement module enhances intra-class consistency and inter-class separability by dividing the change representation into multiple semantic groups and refining them independently to obtain semantic features, specifically: The semantic consistency enhancement module includes N semantic consistency enhancement blocks, and the change representation of the output of the i-th change representation decoupling block is input to the i-th semantic consistency enhancement block. In each semantic consistency enhancement block, the change representation is divided into multiple semantic groups and refined independently to obtain semantic features.

[0014] Furthermore, within each semantic consistency enhancement block, the change representation is divided into multiple semantic groups and independently refined to obtain semantic features, specifically: Let the change in the output of the i-th change be represented as... ,right The data is randomly shuffled and re-divided into N1 groups, denoted as... , express The n1th grouping feature, where n1 = 1, 2, ..., N1; By utilizing multi-scale convolutional operations with different receptive fields, more class-discriminative semantic cues are extracted from each group of features at multiple scales, resulting in... Semantic clues; According to The query vector obtained from the semantic clues is denoted as The key vector is denoted as Value vectors are denoted as ,pass and The dot product operation calculates the channel relationship, T represents the transpose operation, and is further related to... Perform dot product operations, and finally obtain the channel-enhanced features through convolution operations; By fusing the channel-enhanced features corresponding to all grouped features, the semantic features output by each semantic consistency enhancement block are obtained.

[0015] Furthermore, the aforementioned The semantic clues are: , In the formula, express semantic clues This indicates a splicing operation. This represents a convolution operation with a kernel size of kk×kk. Representation layer normalization.

[0016] Furthermore, the semantic features output by each semantic consistency enhancement block are: , In the formula, This represents the semantic features output by the i-th semantic consistency enhancement block. This represents the enhanced feature corresponding to the n1th group feature.

[0017] The present invention also provides a remote sensing change detection system based on change decoupling enhancement, comprising: The image acquisition module is used to acquire dual-temporal remote sensing images, which are remote sensing images of the same area at different times as dual-temporal image pairs; A remote sensing change detection model construction module is used to construct a change decoupling enhancement network model. The change decoupling enhancement network model includes a dual-branch encoder, a change representation decoupling module, a semantic consistency enhancement module, and a segmentation head. The dual-branch encoder extracts dual-temporal features of the dual-temporal image pairs at different scales. The change representation decoupling module explicitly identifies and separates unchanged components by calculating the similarity between the dual-temporal features to obtain change representations. The semantic consistency enhancement module divides the change representations into multiple semantic groups and refines them independently to enhance intra-class consistency and inter-class separability to obtain semantic features. The segmentation head performs remote sensing change detection based on the semantic features. The model training module is used to train the variation decoupling enhancement network model by combining the prediction results of the segmentation head; The remote sensing change detection module is used to input the dual-temporal remote sensing image to be detected into the trained change decoupling enhancement network model to obtain the remote sensing change detection result.

[0018] Compared with the prior art, the above-described technical solution of the present invention has the following advantages: This invention explicitly identifies and separates unchanged components through a change characterization decoupling module. Based on this, it combines a semantic consistency enhancement module to enhance intra-class consistency and inter-class separability. Through the organic combination of feature decoupling and semantic enhancement mechanisms, semantic consistency among similar targets is maintained, improving the semantic discrimination capability of change detection and thus improving the accuracy of remote sensing change detection. Attached Figure Description

[0019] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart of a method in a preferred embodiment of the present invention.

[0020] Figure 2 This is a structural diagram of the variation decoupling enhancement network model in a preferred embodiment of the present invention.

[0021] Figure 3 This is a structural diagram of the variation characterization decoupling block in a preferred embodiment of the present invention.

[0022] Figure 4 This is a structural diagram of the semantic consistency enhancement block in a preferred embodiment of the present invention. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0024] Reference Figure 1As shown, this invention discloses a remote sensing change detection method based on change decoupling enhancement, comprising the following steps: S1: Acquire dual-temporal remote sensing images, and use remote sensing images of the same area at different times as dual-temporal image pairs; denote the dual-temporal image pair as ( , ), and These are remote sensing images of the same area at different times.

[0025] S2: Construct a change-decoupling enhancement network model, such as Figure 2 As shown, the change decoupling enhancement network model includes a dual-branch encoder, a change representation decoupling module, a semantic consistency enhancement module, and a segmentation head. The dual-branch encoder extracts dual-temporal features of the dual-temporal image pairs at different scales. The change representation decoupling module explicitly identifies and separates unchanged components by calculating the spatial similarity between the dual-temporal features, thereby obtaining a purer and more stable change representation. The semantic consistency enhancement module enhances intra-class consistency and inter-class separability by dividing the change representation into multiple semantic groups and refining them independently, thus obtaining semantic features. The segmentation head performs remote sensing change detection based on the semantic features.

[0026] S2-1: The dual-branch encoder extracts the dual-temporal features of the dual-temporal image pair at different scales, specifically as follows: Let the two remote sensing images in the dual-temporal image pair be denoted as , , , H represents the height of the remote sensing image, W represents the width of the remote sensing image, and 3 represents the channel dimension. In this embodiment, the dual-branch encoder is an image encoder, which is used to extract the data separately. and Global spatiotemporal features at different scales, denoted as and Let i = 1, 2, ..., N, where N is the number of scales. In this embodiment, global spatiotemporal features are extracted at four scales, i.e., N = 4. , , Let i be the channel dimension at the i-th scale. , These are the features at the i-th scale corresponding to two remote sensing images of the same region at different times. In this embodiment, the image encoder uses a Mix-Transformer encoder.

[0027] S2-2: When the change representation decoupling module explicitly identifies and separates the unchanged components by calculating the spatial similarity between the two temporal features to obtain a purer and more stable change representation, it captures the invariant features between the two images and then decouples the difference parts to model the feature representation of the change.

[0028] The change representation decoupling module includes N change representation decoupling blocks (CRDs), each of which includes two branches, which obtain the change representation of the dual-temporal features through decoupling. and Input the i-th change representation decoupling block, i=1,2,…,N, and the output of the change representation decoupling block is the change representation.

[0029] The i-th change characterizes the structure of the decoupling block as follows: Figure 3 As shown, the change representation of the i-th change representation decoupling block output is: , In the formula, This represents the change representation of the output of the i-th change representation decoupling block. This represents a convolution operation with a kernel of size k1×k1. This indicates a splicing operation. , They are respectively , The change representation (invariant feature) obtained by decoupling in the i-th change representation decoupling block; The four change representation decoupling blocks output four change representations respectively.

[0030] The two branches of the i-th change characterization decoupling block calculate the... and In one branch, the The calculation method is as follows: , In the formula, The i-th change represents the decoupling block. The corresponding similarity score matrix, This indicates element-wise multiplication. In another branch, Calculation method and Same. Guided by the similarity score matrix, and Decoupled, , .

[0031] The calculation method is as follows: In the i-th change representation decoupling block, reshaping Let the query vector be denoted as Reshaping Let be the key vector, denoted as , , Attention calculation via scaling dot product and The feature similarity map obtained from pixel-level similarity is denoted as... , The i-th change represents the decoupling block. The corresponding feature similarity map, . The calculation method is as follows: , In the formula, T represents the transpose operation. Indicates the vector dimension; Next, in Select the maximum value in each row (i.e.) Each pixel in the middle and The maximum similarity value corresponding to each pixel in the two-time feature is used as the co-significance probability of each pixel, denoted as . , This represents the cosignificance probability of the k-th pixel. The calculation formula is: , , ; Finally, reshaping get , , Used to indicate the invariant region between features in two-phase images. The reshaping method involves changing the shape of the tensor through the Reshape operation.

[0032] S2-3: The goal of the semantic consistency enhancement module is to enhance inter-class diversity and intra-class similarity. In this module, features are grouped and processed separately, enabling each group to learn semantic representations with categorical features, thereby highlighting discriminative inter-class cues and maintaining intra-class consistency. The semantic consistency enhancement module enhances intra-class consistency and inter-class separability by dividing change representations into multiple semantic groups and refining them independently, resulting in semantic features. Specifically, the module includes N semantic consistency enhancement blocks (SCEs), where the change representation output from the i-th change representation decoupling block is input to the i-th semantic consistency enhancement block; for example... Figure 4As shown, in each semantic consistency enhancement block, the change representation is divided into multiple semantic groups and refined independently to obtain semantic features.

[0033] Within each semantic consistency enhancement block, the change representation is divided into multiple semantic groups and refined independently to obtain semantic features, specifically: S2-3-1: Let the change representation of the i-th change representation decoupling block output be (as shown in the original text). ,right The data is randomly shuffled and re-divided into N1 groups, denoted as... , express The n1-th grouping feature, where n1 = 1, 2, ..., N1. (Before the features are input to the semantic consistency enhancement module, the channel dimension has been unified to C, which is C=64 in this embodiment); In this embodiment, N1=2, and the partitioning method is to randomly shuffle and then divide into two.

[0034] S2-3-2: The intra-group processing module mines category-specific semantic features for each group feature. Considering the diversity of target shape and scale in remote sensing imagery, the intra-group processing module first utilizes multi-scale convolutional operations with different receptive fields to mine more category-discriminative semantic cues for each group feature from multiple scales, resulting in... The semantic clues are: , In the formula, express semantic clues This indicates a splicing operation. This represents a convolution operation with a kernel size of kk×kk. Representation layer normalization; in this embodiment, four different convolutional operations with different kernels are used for mining. The semantic clues, namely: , .

[0035] S2-3-3: The in-group processing module further models the global relationships between channels. By understanding these relationships, each channel can perceive its similarities or differences with other channels, thus better mining category-specific semantic features. For According to The resulting query vector is denoted as The key vector is denoted as Value vectors are denoted as ,pass and The dot product operation calculates the channel relationship and further correlates it with... Perform dot product, and finally convolution operation to obtain the channel-enhanced features, denoted as . ; , The specific calculation formula is as follows: , In the formula, Represents the vector dimension. The normalization function represents the channel dimension.

[0036] S2-3-4: By fusing the channel-enhanced features corresponding to all grouped features, the semantic features output by each semantic consistency enhancement block are obtained as follows: , In the formula, This represents the semantic features output by the i-th semantic consistency enhancement block. In this embodiment, , , This represents the enhanced feature corresponding to the n1th group feature.

[0037] With input Compare, It has clearer category characteristics and is effectively enhanced in terms of inter-class differences and intra-class consistency.

[0038] This decoupling method, achieved through the change representation decoupling module, guides the model to focus on the regions that have actually changed, while suppressing background noise and providing more reliable input features for the subsequent semantic consistency enhancement module.

[0039] S2-4: Semantic features output from the four semantic consistency enhancement blocks ( , , , The input is fed into the segmentation head to obtain the final predicted change detection map.

[0040] S3: Train the variation decoupling enhancement network model based on the prediction results of the segmentation head.

[0041] In this embodiment, the total loss function is composed of Dice loss and binary cross-entropy loss, and the formula for calculating the total loss function is as follows: , In the formula, Represents the total loss function. This indicates Dice's loss. Represents the binary cross-entropy loss. , This represents the weighting coefficient. , These represent the actual labels and their corresponding predicted changes. In the middle, the pixel value is either 0 or 1; while in... In this context, the pixel value ranges from 0 to 1, representing the probability that a pixel is predicted to be a region of variation.

[0042] The calculation method is as follows: .

[0043] The calculation method is as follows: .

[0044] S4: Input the dual-temporal remote sensing image to be detected into the trained change decoupling enhancement network model to obtain the remote sensing change detection result.

[0045] The present invention also discloses a remote sensing change detection system based on change decoupling enhancement, including an image acquisition module, a remote sensing change detection model construction module, a model training module, and a remote sensing change detection module.

[0046] The image acquisition module is used to acquire dual-temporal remote sensing images, which are remote sensing images of the same area at different times as dual-temporal image pairs.

[0047] The remote sensing change detection model construction module is used to construct a change decoupling enhancement network model. The change decoupling enhancement network model includes a dual-branch encoder, a change representation decoupling module, a semantic consistency enhancement module, and a segmentation head. The dual-branch encoder extracts dual-temporal features of the dual-temporal image pairs at different scales. The change representation decoupling module explicitly identifies and separates unchanged components by calculating the similarity between the dual-temporal features to obtain change representations. The semantic consistency enhancement module divides the change representations into multiple semantic groups and refines them independently to enhance intra-class consistency and inter-class separability to obtain semantic features. The segmentation head performs remote sensing change detection based on the semantic features.

[0048] The model training module is used to train the variation decoupling enhancement network model by combining the prediction results of the segmentation head.

[0049] The remote sensing change detection module is used to input the dual-temporal remote sensing image to be detected into the trained change decoupling enhancement network model to obtain the remote sensing change detection result.

[0050] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a remote sensing change detection method based on change decoupling enhancement.

[0051] The present invention also discloses an apparatus including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a remote sensing change detection method based on change decoupling enhancement.

[0052] In remote sensing change detection tasks, spurious changes caused by unchanged noise are the most typical source of interference. Perturbations from the unchanged background cause spurious changes to become highly entangled with real ground feature changes in the feature space, thus severely affecting the model's ability to identify real changes. An effective remote sensing change detection model should possess two capabilities simultaneously: (1) decoupling real semantic changes from unchanged background features at the feature level; and (2) maintaining semantic consistency among similar targets at the semantic level. This invention can capture change features during feature decoupling and improve intra-class feature consistency. Compared with existing technologies, the advantages of this invention are: 1. A change representation decoupling module was designed to learn feature representations related to change. Unchanged features are captured through similarity matching and decoupled from the bi-temporal features to infer stable change features. This effectively distinguishes change information from unchanged components, thereby enhancing the model's ability to identify change regions.

[0053] 2. A semantic consistency enhancement module is introduced to divide features into multiple semantic groups and process them separately. In this way, features with category-specific semantic information can be generated, thereby enhancing the inter-class separability and intra-class similarity of semantic features.

[0054] To further demonstrate the beneficial effects of this invention, this embodiment uses the Change Decoupling Enhancement Network (CDENet) model from this invention and existing technologies such as FC-EF (see the paper "Daudt RC, Saux BL, Boulch A. Fully convolutional siamese networks for change detection [J]. IEEE, 2018.DOI:10.1109 / ICIP.2018.8451652"), FC-Siam-Diff (see the paper "Daudt RC, Saux BL, Boulch A. Fully convolutional siamese networks for change detection [J]. IEEE, 2018.DOI:10.1109 / ICIP.2018.8451652"), and FC-Siam-Conc (see the paper "Daudt RC, Saux BL, Boulch A. Fully convolutional siamese networks for change detection [J]. IEEE, 2018.DOI:10.1109 / ICIP.2018.8451652") on the WHU-CD dataset, LEVIR-CD dataset, and SYSU-CD dataset, respectively. Boulch A. Fully convolutional siamese networks for change detection [J].IEEE, 2018.DOI:10.1109 / ICIP.2018.8451652"), BIT (bitemporal image transformer, see the paper "Chen H, Qi Z, Shi Z.Remote SensingImage Change Detection With Transformers[J].IEEE Transactions on Geoscienceand Remote Sensing, 2022, 60.DOI:10.1109 / TGRS.2021.3095166"), DMINet (see the paper "Feng Y, Jiang J, Xu H, Zheng J. Change detection on remote sensing images using dual-branch multilevel intertemporal network. IEEE Transactions onGeoscience and Remote Sensing.2023 Feb 1;61:1-5"), EATDer (see the paper "Ma J, DuanJ, Tang X, Zhang Spatiotemporal enhancement and interlevel fusion network forremote sensing images change detection. IEEE Transactions on Geoscience andRemote Sensing. 2024 Feb 2;62:1-4"), SChanger (see the paper "Zhou Z, Hu K, Fang Y, Rui X. SChanger: Change Detection from a Semantic Change and SpatialConsistency Perspective. IEEE Journal of Selected Topics in Applied EarthObservations and Remote Sensing. 2025 Mar A simulation experiment for remote sensing change detection was conducted using 28" as an example. The detection results were represented by precision (Precision), recall (Recall), F1 score, intersection-to-union (IoU), and overall accuracy (OA). Table 1 shows the results of different models on the WHU-CD dataset, Table 2 shows the results of different methods on the LEVIR-CD dataset, and Table 3 shows the results of different methods on the SYSU-CD dataset.

[0055] Table 1. Results of remote sensing change detection using different methods on the WHU-CD dataset.

[0056] Table 2. Results of remote sensing change detection using different methods on the LEVIR-CD dataset.

[0057] Table 3. Results of remote sensing change detection using different methods on the SYSU-CD dataset.

[0058] As can be seen from Tables 1, 2, and 3, this invention achieves excellent performance in F1 score and IoU index, effectively suppressing interference from unchanged features and improving the accuracy of remote sensing change detection, demonstrating the beneficial effects of this invention. This invention organically combines feature decoupling and semantic enhancement mechanisms, providing a novel and effective solution for change detection in complex remote sensing scenarios.

[0059] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0060] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0061] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0062] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0063] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A remote sensing change detection method based on change decoupling enhancement, characterized in that, include: Acquire dual-temporal remote sensing images, and use remote sensing images of the same area at different times as dual-temporal image pairs; A change decoupling enhancement network model is constructed, comprising a dual-branch encoder, a change representation decoupling module, a semantic consistency enhancement module, and a segmentation head. The dual-branch encoder extracts dual-temporal features of the dual-temporal image pairs at different scales. The change representation decoupling module explicitly identifies and separates unchanged components by calculating the similarity between the dual-temporal features to obtain change representations. The semantic consistency enhancement module divides the change representations into multiple semantic groups and refines them independently to enhance intra-class consistency and inter-class separability to obtain semantic features. The segmentation head performs remote sensing change detection based on the semantic features. The change decoupling enhancement network model is trained by combining the prediction results of the segmentation head, and the dual-temporal remote sensing image to be detected is input into the trained change decoupling enhancement network model to obtain the remote sensing change detection result.

2. The remote sensing change detection method based on change decoupling enhancement according to claim 1, characterized in that: The change characterization decoupling module explicitly identifies and separates the unchanged components by calculating the similarity between the two-phase features to obtain the change characterization, specifically: Let the dual-phase characteristic be denoted as and , , These are the features at the i-th scale corresponding to two remote sensing images of the same region at different times, where i = 1, 2, ..., N, and N is the number of scales; The change representation decoupling module includes N change representation decoupling blocks, each of which includes two branches, which respectively obtain the change representation of the dual-temporal features through decoupling; and Input the i-th change representation decoupling block, and the output of the change representation decoupling block is the change representation.

3. The remote sensing change detection method based on change decoupling enhancement according to claim 2, characterized in that: The change in the output of the i-th change characterization decoupling block is characterized as follows: ， In the formula, This represents the change representation of the output of the i-th change representation decoupling block. This represents a convolution operation with a kernel of size k1×k1. This indicates a splicing operation. , They are respectively , The change representation obtained by decoupling in the i-th change representation decoupling block.

4. The remote sensing change detection method based on change decoupling enhancement according to claim 3, characterized in that: The two branches of the i-th change characterization decoupling block calculate the... and In one branch, the The calculation method is as follows: ， In the formula, The i-th change represents the decoupling block. The corresponding similarity score matrix, This indicates element-wise multiplication.

5. The remote sensing change detection method based on change decoupling enhancement according to claim 4, characterized in that: The The calculation method is as follows: In the i-th change representation decoupling block, reshaping Let the query vector be denoted as Reshaping Let be the key vector, denoted as Attention calculation via scaling dot product and The feature similarity map obtained from pixel-level similarity is denoted as... , The i-th change represents the decoupling block. Corresponding feature similarity map; exist The maximum value in each row is selected as the co-significance probability of each pixel in the dual-time feature, denoted as . , Represents the cosignificance probability of the k-th pixel, reshaping get .

6. The remote sensing change detection method based on change decoupling enhancement according to claim 2, characterized in that: The semantic consistency enhancement module enhances intra-class consistency and inter-class separability by dividing change representations into multiple semantic groups and refining them independently, thereby obtaining semantic features. Specifically: The semantic consistency enhancement module includes N semantic consistency enhancement blocks, and the change representation of the output of the i-th change representation decoupling block is input to the i-th semantic consistency enhancement block. In each semantic consistency enhancement block, the change representation is divided into multiple semantic groups and refined independently to obtain semantic features.

7. The remote sensing change detection method based on change decoupling enhancement according to claim 6, characterized in that: In each semantic consistency enhancement block, the change representation is divided into multiple semantic groups and independently refined to obtain semantic features, specifically: Let the change in the output of the i-th change be represented as... ,right The data is randomly shuffled and re-divided into N1 groups, denoted as... , express The n1th grouping feature, where n1 = 1, 2, ..., N1; By utilizing multi-scale convolutional operations with different receptive fields, more class-discriminative semantic cues for each group of features are extracted from multiple scales. Semantic clues; According to The query vector obtained from the semantic clues is denoted as The key vector is denoted as Value vectors are denoted as ,pass and The dot product operation calculates the channel relationship, T represents the transpose operation, and is further related to... Perform dot product operations, and finally obtain the channel-enhanced features through convolution operations; By fusing the channel-enhanced features corresponding to all grouped features, the semantic features output by each semantic consistency enhancement block are obtained.

8. The remote sensing change detection method based on change decoupling enhancement according to claim 7, characterized in that: The The semantic clues are: ， In the formula, express semantic clues This indicates a splicing operation. This represents a convolution operation with a kernel size of kk×kk. Representation layer normalization.

9. The remote sensing change detection method based on change decoupling enhancement according to claim 7, characterized in that: The semantic features output by each semantic consistency enhancement block are: ， In the formula, This represents the semantic features output by the i-th semantic consistency enhancement block. This represents the enhanced feature corresponding to the n1th group feature.

10. A remote sensing change detection system based on change decoupling enhancement, characterized in that... include: The image acquisition module is used to acquire dual-temporal remote sensing images, which are remote sensing images of the same area at different times as dual-temporal image pairs; A remote sensing change detection model construction module is used to construct a change decoupling enhancement network model. The change decoupling enhancement network model includes a dual-branch encoder, a change representation decoupling module, a semantic consistency enhancement module, and a segmentation head. The dual-branch encoder extracts dual-temporal features of the dual-temporal image pairs at different scales. The change representation decoupling module explicitly identifies and separates unchanged components by calculating the similarity between the dual-temporal features to obtain change representations. The semantic consistency enhancement module divides the change representations into multiple semantic groups and refines them independently to enhance intra-class consistency and inter-class separability to obtain semantic features. The segmentation head performs remote sensing change detection based on the semantic features. The model training module is used to train the variation decoupling enhancement network model by combining the prediction results of the segmentation head; The remote sensing change detection module is used to input the dual-temporal remote sensing image to be detected into the trained change decoupling enhancement network model to obtain the remote sensing change detection result.

Images