A change plot identification label generation method based on style decoupling of a flood storage and detention area

Abstract

This invention discloses a style decoupling-based technology for generating change patch recognition labels in flood storage and detention areas, comprising: S1, construction and preprocessing of a change detection dataset; S2, construction of a conditional diffusion generation network based on a decoupled encoder; S3, decoupling extraction and orthogonalization representation of content-style features; S4, construction of a multi-objective training strategy and two-stage model training; S5, style feature injection and fusion based on cross-attention; S6, cross-domain style transfer and diversified label generation; and S7, outputting a high-quality change detection augmented dataset based on label quality screening under both physical and semantic constraints. Compared with existing technologies, this invention, by introducing a content-style decoupling mechanism, achieves independent and precise control over the style and content of generated samples, generating pixel-level precisely labeled change detection labels with consistent feature layouts but diverse imaging styles.

Description

Technical Field

[0001] This invention relates to the field of remote sensing information extraction and analysis, specifically to a method for generating labels for identifying change patches in flood storage and detention areas based on style decoupling. Background Technology

[0002] Flood storage and detention areas, as key projects for ensuring flood control safety in river basins, require standardized management of non-flood control construction projects within them. This is a crucial link in reducing flood diversion risks and protecting people's lives and property. Ensuring that flood storage and detention areas can be safely activated according to their designed functions during planning and scheduling is of paramount importance in improving flood control capabilities. The key to achieving this goal lies in high-precision, routine remote sensing monitoring of changes in vegetation within flood storage and detention areas.

[0003] However, in practical applications, the detection of changes in flood storage and detention areas faces severe challenges: First, the number of legally compliant change patches is scarce, resulting in a severe shortage of labels needed for training deep learning models. Second, different flood storage and detention areas are located in different climate zones with complex imaging conditions, leading to huge differences in image style and making it difficult for models to generalize. For example, the Jingjiang flood diversion area in the Yangtze River Basin is affected by clouds, rain, and dense vegetation all year round, resulting in blurred images and color distortion; the Dongdian flood storage and detention area in the Haihe River Basin is easily affected by ice and snow cover in winter, making it difficult to identify ground textures. This huge "style difference" (i.e., non-ground feature change factors such as clouds, rain, snow, vegetation, and lighting) seriously affects the generalization ability of the model, causing a model trained in one area to fail in another. In addition, traditional data augmentation methods (such as cropping, scaling, and rotation) cannot create new semantic scenes and are difficult to simulate complex and ever-changing imaging conditions.

[0004] Therefore, there is an urgent need for a data augmentation method that can automatically generate large-scale, high-quality data with rich style variations for change patch recognition, in order to support the training of high-precision, highly generalized change detection models. Summary of the Invention

[0005] The technical solution of this invention is as follows:

[0006] The first aspect of the present invention provides a method for generating labels for identifying change patches in flood storage and detention areas based on style decoupling, comprising the following steps:

[0007] S1. Construction and preprocessing of change detection dataset: Obtain pre-change images (I) containing different imaging conditions, geographical locations, and time points (t1, t2). pre Image after change I post The corresponding binary change region mask M is used as the original dataset and preprocessed to form the original training sample triplet {I}. pre， I postThe preprocessing includes: performing radiometric and / or atmospheric correction and geometric registration on all images to ensure I pre and I post The images are then fully aligned in space; subsequently, they are uniformly cropped into samples of a fixed size and normalized.

[0008] S2. Constructing a Conditional Diffusion Generation Network Based on a Decoupled Encoder: A network architecture for image generation is built, comprising two core modules: a content-style decoupled encoder E and a conditional denoising network U-Net. Encoder E uses a shared weight network and independent branch network structure for feature extraction. The conditional denoising network U-Net serves as the generation framework, employing an encoder-decoder structure. The encoder is used for feature extraction and downsampling, while the decoder is used for feature upsampling and image reconstruction. Skip connections preserve detail information, and iterative denoising achieves image generation. I... pre and I post The M and M are concatenated in the channel dimension and used as multimodal conditional information input to the encoder of U-Net, so that the generation process is guided by the original semantic information and geometric layout.

[0009] S3. Decoupling and Orthogonalization of Content-Style Features: Through the shared weight convolutional network in the encoder E, from I... pre and I post Extract the shared content encoding Z that is unaffected by the imaging style. c The shared content encoding characterizes the geometric structure and spatial layout of land features; through the independent branch network in the encoder E, it is respectively derived from I pre and I post Extract their respective style codes Z s The style encoding represents style information; simultaneously, an orthogonal constraint loss function is introduced to calculate Z. c With Z s The cosine similarity between them is calculated, and the optimization objective is to minimize the absolute value or square of the cosine similarity, thus forcing the Z... c With Z s They are independent of each other in the feature space, achieving effective separation of structural information and imaging style information.

[0010] S4. Construct a multi-objective training strategy and a two-stage model training: Define a multi-objective loss function, which includes an orthogonal constraint objective to ensure content and style separation, a reconstruction objective to guarantee the quality of generated images, and a diffusion objective to guide the diffusion model for denoising; adopt a two-stage training strategy: the first stage trains the content-style decoupling encoder E to optimize the decoupling effect; the second stage jointly trains the encoder E and the conditional denoising network U-Net, so that it can learn the conditional generation capability under the constraints of the multi-objective loss function.

[0011] S5. Style Feature Injection and Fusion Based on Cross-Attention: The shared content encoding Zc extracted in step S3 is fused with a target style encoding Zs_target to generate a conditional vector F. In the reverse denoising process of the U-Net decoder, the cross-attention mechanism is used to inject and guide the feature injection by using the intermediate layer features of U-Net as queries and the conditional vector F as keys and values. This mechanism injects Zc and the target style encoding Zs_target into each layer of U-Net to ensure that the generated image maintains the geometric layout constrained by Zc at the pixel level, while presenting the style features defined by Zs_target.

[0012] S6. Cross-domain style transfer and diversified label generation: The conditional denoising network U-Net trained in step S4 is used for inference. In the reverse denoising inference stage, the content encoding Z of the original sample is used. c Based on the control information, the terrain structure remains unchanged; style codes Z are obtained through random sampling, mixing, or specifying from different domains. s_new It enables cross-domain style transfer, generating enhanced image pairs with consistent ground feature distribution but different imaging styles; and through large-scale inference, it obtains an expanded label set with diverse styles that exceeds the size of the original dataset.

[0013] S7. Label Quality Screening Based on Dual Physical and Semantic Constraints: An automated quality assessment process is constructed to screen labels in the expanded label set generated in step S6, outputting a high-quality change detection expanded dataset. The label screening process is as follows: First, the distribution distance between the generated images in the expanded label set and the original images obtained after preprocessing in step S1 in the depth feature space is calculated, and the physical realism is evaluated using the FID index; second, the generated images are input into the pre-trained change detection network, and the IoU score between the predicted mask and the input change mask is calculated to evaluate semantic consistency; finally, unqualified labels are automatically filtered based on the comprehensive quality score of FID and IoU.

[0014] According to the method of the first aspect of the present invention, in step S3, the decoupling extraction and orthogonalization representation of the content-style features are specifically implemented as follows:

[0015] (1) Feature extraction of encoder E

[0016] The shared-weight convolutional network adopts a "ConvNeXt-ViT hybrid backbone architecture", such as... Figure 1 As shown:

[0017] The bottom layer uses 6 layers of ConvNeXt-Tiny convolutional blocks with a kernel size of 3×3 and the number of channels gradually increases from 64 to 256. It reduces the computational load through depthwise separable convolution while capturing basic structural features such as local edges and textures of the image.

[0018] Middle layer: Embedded with 4 cross-scale attention fusion blocks, which associate 1 / 2, 1 / 4, and 1 / 8 scale feature maps across layers and model the dependencies of global ground feature layout (such as building outlines and road directions) through a self-attention mechanism;

[0019] Top layer: The Global Feature Aggregation (GFA) module is used to perform weighted summation on feature maps at each scale (the weights are adaptively allocated by the information entropy of the feature maps), and finally generate a shared content encoding vector Zc with dimension Dc=2560, ensuring that it covers both local details and global structural information.

[0020] The independent branch network is a symmetric structure of "lightweight Transformer + gated convolution" as follows: Figure 2 As shown:

[0021] Each branch contains 3 layers of MobileViT blocks (lightweight Transformer variants that adapt to fine-grained extraction of style features), which capture the local relevance of style information through local window attention;

[0022] Subsequently, two layers of gated convolution are connected in series, and the sigmoid activation function is used to generate channel gate vectors to filter out feature channels that are strongly correlated with illumination, season and imaging sensor type, and suppress irrelevant noise.

[0023] Finally, the feature dimension is compressed to Ds=384 by a 1×1 convolution layer, and the style encoding vectors Zs1 and Zs2 of Ipre and Ipost are output respectively, so as to achieve accurate capture of style information and dimensional optimization.

[0024] (2) Feature standardization preprocessing

[0025] To address the issue of inconsistent feature scales caused by radiometric differences in remote sensing images, a "local-global two-stage standardization" process is adopted:

[0026] Local standardization: for Z c Z s1 Z s2Local Response Normalization (LRN) is performed separately, using the local neighborhood (window size 5) of each feature vector as the unit to suppress redundant feature channels, as shown in the following formula:

[0027]

[0028] In the formula, Z local Z is the locally standardized feature vector, Z is the original input feature vector, and k is the window size. (i-j):(i+j) This represents the feature vector elements within a local neighborhood centered at position i with a window size of k, where ϵ is a smoothing term, typically taking the value 10. −7 Avoid having a denominator of 0;

[0029] Global standardization: The locally standardized feature vectors are subjected to L2 normalization, constraining the vector magnitude to 1, as shown in the following formula:

[0030]

[0031] In the formula, The feature vector is the globally standardized form, where ||Zlocal||2 represents Z. local The L2 norm.

[0032] Two-stage standardization ensures both the local discriminative power of features and the numerical stability of cosine similarity calculation, avoiding the failure of orthogonal constraints due to radiation differences.

[0033] (3) Optimization of orthogonal constraint loss

[0034] To overcome the limitations of a single cosine similarity constraint, a weighted hybrid orthogonal loss function combining "cosine similarity + mutual information" is adopted, simultaneously constraining both linear and nonlinear correlations. The mathematical expression is as follows:

[0035]

[0036] In the formula, L ortho This represents the mixed orthogonal loss function. The weighting coefficients of the cosine similarity loss Lcos are represented by... L represents the weighting coefficient of the cosine similarity loss Lmi, Lcos represents the cosine similarity loss, and Lmi represents the mutual information loss.

[0037] Cosine similarity loss Lcos: constrains the linear orthogonality of feature vectors, as shown in the following formula:

[0038]

[0039] In the formula, N represents the batch sample size. This represents the content feature vector of the i-th sample. This represents the first style feature vector of the i-th sample. Let represent the second style feature vector of the i-th sample, cos(.,.) represent the cosine similarity function, and cos2(.,.) represent the squared cosine similarity.

[0040]

[0041] In the formula, MI(,.,) represents the mutual information calculation function, and the meanings of the other parameters are the same as those in the Lcos function;

[0042] The weights λ1=0.6 and λ2=0.4 (the optimal ratio is determined by 5-fold cross-validation), and N is the batch size. During training, the weight ratio of this loss to other target losses is set to λortho=0.7, and gradient clipping (clip_value=0.5) is used to avoid gradient oscillation and ensure optimization stability.

[0043] (4) Decoupling effectiveness verification mechanism

[0044] The "Multi-Dimensional Decoupling Evaluation Index (MDEI)" is designed to quantify the decoupling effect, comprehensively considering linear correlation, nonlinear correlation, and feature reconstruction capability.

[0045]

[0046] Where, cosˉ is and , The mean cosine similarity is MIˉ, the mean mutual information value is MIˉ, and the feature reconstruction error is REˉ (the mean MSE of the original image after stitching Zc and Zs together).

[0047] Weights α=0.3, β=0.3, γ=0.4, MDEI∈[0,1], and the decoupling effect is considered to be satisfactory when MDEI≥0.90.

[0048] If the target is not met, a tiered adjustment strategy will be activated:

[0049] If cosˉ>0.07: fix the parameters of the bottom 3 ConvNeXt blocks of the shared network, and only fine-tune the weights of the middle attention blocks and the top GFA module;

[0050] If MIˉ>0.12: Increase the weight λ2 of mutual information loss (adjust by +0.1 each time, with a maximum of 0.6);

[0051] If REˉ>0.05: Insert residual connections in the shared network to enhance feature propagation and reduce reconstruction error;

[0052] During the adjustment process, the MDEI is recalculated every 30 epochs until the target conditions are met, ensuring the reliability and stability of the decoupling effect.

[0053] A second aspect of this invention discloses a remote sensing change detection data augmentation system based on a style decoupling generative model, the system comprising:

[0054] The data preprocessing module is configured to perform step S1;

[0055] The model building module is configured to execute step S2;

[0056] The decoupled training module is configured to execute steps S3 and S4;

[0057] The controllable generation module is configured to execute steps S5 and S6;

[0058] The quality assessment and output module is configured to execute step S7.

[0059] A third aspect of the present invention discloses an electronic device. The electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the data augmentation method according to any one of the first aspects of the present disclosure.

[0060] A fourth aspect of the present invention discloses a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the data augmentation method according to any one of the first aspects of this disclosure.

[0061] The beneficial effects of this invention are as follows: Compared with the prior art, this invention, by introducing a content-style decoupling mechanism, achieves independent and precise control over the style and content of generated samples, generating pixel-level precisely labeled change detection tags with consistent ground feature layouts but diverse imaging styles. This overcomes the problem that traditional data augmentation methods cannot create new semantic scenes, improves the diversity and quality of generated tags, and provides ample data support for training change detection models. Attached Figure Description

[0062] Figure 1 ConvNeXt-ViT hybrid backbone architecture;

[0063] Figure 2 Lightweight Transformer + gated convolutional network;

[0064] Figure 3 Example of label enhancement for remote sensing image change detection based on style decoupling. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0066] A style-decoupling-based technique for identifying and generating labels for change patches in flood storage and detention areas is presented. This implementation uses remote sensing extraction of dikes in the Mengwa flood storage and detention area of ​​Anhui Province as a case study. It provides a method for enhancing remote sensing image change detection data based on style decoupling, the method comprising:

[0067] Step S1: Construction and Preprocessing of Change Detection Dataset: Acquire GF-1 / 6 2-meter high-resolution remote sensing images of the Mengwa flood storage area in Anhui Province at two time phases, t1 (before the change) and t2 (after the change), along with corresponding manually labeled or high-precision algorithm-generated change area mask data M. Perform radiometric and geometric correction on the images to eliminate radiometric errors and geometric distortions. Crop the images into uniform 512×512 pixel blocks and normalize the pixel values ​​to the range [-1, 1]. Finally, form 1500 original training sample triplets {I} containing the original image pairs and the change mask. pre， I post , M}.

[0068] Step S2: Construct a conditional diffusion generation network based on a decoupled encoder: A network architecture for image generation is built, comprising two core modules: a content-style decoupled encoder (E) and a conditional denoising network U-Net. The encoder E employs a shared weight network and independent branch network structure for feature extraction. The U-Net network serves as the generation framework; in this embodiment, an improved U-shaped structure is used, embedding Transformer blocks (multi-head self-attention mechanism) in key layers of the encoder and decoder, and residual connections are employed to achieve efficient iterative denoising and image generation. pre and I post The M and M are concatenated in the channel dimension and used as multimodal conditional information input to the encoder of U-Net to ensure that the generation process is guided by the original semantic information and geometric layout.

[0069] Step S3, Decoupling Extraction and Orthogonalization Representation of Content-Style Features: Through the shared weight convolutional network in the encoder E, from I... pre and I post Extracting shared content encoding (Z) that is unaffected by imaging style. c This encoding characterizes the geometric structure and spatial layout of ground features; through the independent branch network in the encoder E, it respectively extracts data from I... pre and Ipost Extract their respective style codes (Z s1 Z s2 This encoding represents style information such as illumination, season, and imaging conditions; simultaneously, an orthogonal constraint loss function is introduced, where the balance coefficient λ is set to 0.5, to calculate Z. c With Z s The cosine similarity between them is calculated, and the optimization objective is to minimize the absolute value or square of this cosine similarity, thus forcing the Z... c With Z s They are independent of each other in the feature space, achieving effective separation of structural information and imaging style information.

[0070] Step S4: Construct a multi-objective training strategy and a two-stage model training: Define a multi-objective loss function, which includes an orthogonal constraint objective to ensure content and style separation, a reconstruction objective to guarantee the quality of generated images, and a diffusion objective to guide the diffusion model for denoising; adopt a two-stage training strategy: The first stage trains the content-style decoupling encoder E to optimize the decoupling effect, training for 100 epochs with a learning rate of 1e-4; the second stage jointly trains the encoder E and the conditional denoising model U-Net to learn the conditional generation capability under the constraints of the multi-objective loss function, training for 500 epochs with a learning rate of 5e-5, using the AdamW optimizer.

[0071] Step S5: The shared content encoding Zc extracted in step S3 is fused with a target style encoding Zs_target to generate a conditional vector F. Zs_target is either the original style encoding Zs1 or Zs2 corresponding to the input image during model training, and can be any newly specified style encoding during generation. In the reverse denoising process of the U-Net decoder, a cross-attention mechanism is used, with U-Net intermediate layer features as queries and the fused conditional vector F as keys and values ​​for feature injection and guidance. This mechanism injects Zc and the target style encoding Zs_target into various layers of the U-Net to ensure that the generated image maintains the geometric layout constrained by Zc at the pixel level, while simultaneously presenting the style features defined by Zs_target.

[0072] Step S6, Cross-domain Style Transfer and Diverse Label Generation: The conditional denoising network U-Net trained in step S4 is used for inference. In the reverse denoising inference stage of the model, the content encoding Z of the original sample is used. c Based on the control information, the terrain structure remains unchanged; style codes Z are obtained through random sampling, mixing, or specifying from different domains. s_newIt achieves cross-domain style transfer, generating enhanced image pairs with consistent ground feature distribution but different imaging styles; through large-scale inference, such as by repeatedly executing the reverse denoising process and traversing or randomly sampling large-scale style codes, it obtains an expanded label set with diverse styles that far exceeds the number of the original dataset, successfully expanding the original 1,500 labels to 25,000 labels.

[0073] Step S7: Label quality screening based on both physical and semantic constraints: An automated quality assessment process is constructed to screen the labels in the expanded label set generated in Step S6. First, the distribution distance between the generated images in the expanded label set and the original images obtained after preprocessing in Step S1 in the depth feature space is calculated, and the physical realism is evaluated using the FID index. Second, the generated images are input to a pre-trained change detection network, and the IoU score between the prediction mask and the input change mask is calculated to evaluate semantic consistency. Finally, unqualified labels are automatically filtered based on the comprehensive quality score of FID and IoU, with an FID threshold of less than 35 and an IoU threshold of greater than 0.65, ultimately obtaining 25,000 high-quality labels.

[0074] The above-disclosed technical solution of this invention yields the following beneficial results: Based on a style decoupling generative model, this invention effectively overcomes the problem that traditional data augmentation methods cannot create new semantic scenarios, significantly improves the diversity and quality of generated labels, and provides ample data support for training change detection models (see appendix). Figure 3 .

[0075] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for generating labels for identifying change patches in flood storage and detention areas based on style decoupling, characterized in that: Includes the following steps: S1, change detection dataset construction and preprocessing: obtain the pre-change image I pre , the post-change image I post , and the corresponding binary change region mask M as the original dataset and perform preprocessing to constitute the original training sample triple {I pre， I post , M}; S2. Constructing a Conditional Diffusion Generation Network Based on a Decoupled Encoder: A network architecture for image generation is built, comprising two core modules: a content-style decoupled encoder E and a conditional denoising network U-Net. Encoder E uses a shared weight network and independent branch network structure for feature extraction. The conditional denoising network U-Net serves as the generation framework, employing an encoder-decoder structure. The encoder is used for feature extraction and downsampling, while the decoder is used for feature upsampling and image reconstruction. Skip connections preserve detail information, and iterative denoising achieves image generation. I... pre and I post M is concatenated with M in the channel dimension and used as multimodal conditional information input to the encoder of U-Net, so that the generation process is guided by the original semantic information and geometric layout. S3. Decoupling and Orthogonalization of Content-Style Features: Through the shared weight convolutional network in the encoder E, from I... pre and I post Extract the shared content encoding Z that is unaffected by the imaging style. c The shared content encoding characterizes the geometric structure and spatial layout of land features; through the independent branch network in the encoder E, it is respectively derived from I pre and I post Extract their respective style codes Z s The style encoding represents style information; simultaneously, an orthogonal constraint loss function is introduced to calculate Z. c With Z s The cosine similarity between them is calculated, and the optimization objective is to minimize the absolute value or square of the cosine similarity, thus forcing the Z... c With Z s They are independent of each other in the feature space, achieving effective separation of structural information and imaging style information; S4. Constructing a multi-objective training strategy and a two-stage model training: Define a multi-objective loss function, which includes an orthogonal constraint objective to ensure content and style separation, a reconstruction objective to guarantee the quality of generated images, and a diffusion objective to guide the diffusion model for denoising; adopt a two-stage training strategy: the first stage trains the content-style decoupling encoder E to optimize the decoupling effect; the second stage jointly trains the encoder E and the conditional denoising network U-Net, so that it can learn the conditional generation capability under the constraints of the multi-objective loss function. S5. Style feature injection and fusion based on cross-attention: The shared content encoding Zc extracted in step S3 is fused with a target style encoding Zs_target to generate a conditional vector F. In the reverse denoising process of the U-Net decoder, the cross-attention mechanism is used to inject and guide the feature injection by using the intermediate layer features of U-Net as the query and the conditional vector F as the key and value. This mechanism injects Zc and the target style encoding Zs_target into each layer of U-Net to ensure that the generated image maintains the geometric layout constrained by Zc at the pixel level, while presenting the style features defined by Zs_target. S6. Cross-domain style transfer and diversified label generation: The conditional denoising network U-Net trained in step S4 is used for inference. In the reverse denoising inference stage, the content encoding Z of the original sample is used. c Based on the control information, the terrain structure remains unchanged; style codes Z are obtained through random sampling, mixing, or specifying from different domains. s_new It achieves cross-domain style transfer, generating enhanced image pairs with consistent ground feature distribution but different imaging styles; through large-scale inference, it obtains an expanded label set with diverse styles that exceeds the size of the original dataset; S7. Label quality screening based on both physical and semantic constraints: Construct an automated quality assessment process to screen the labels in the expanded label set generated in step S6 and output a high-quality change detection expanded dataset.

2. The method for generating labels for identifying change patches in flood storage and detention areas based on style decoupling according to claim 1, characterized in that: The preprocessing described in S1 includes: radiometric and / or atmospheric correction and geometric registration of all images to ensure that I pre and I post The images are then fully aligned in space; subsequently, they are uniformly cropped into samples of a fixed size and normalized.

3. The method for generating labels for identifying change patches in flood storage and detention areas based on style decoupling according to claim 1, characterized in that: In step S3: The shared-weight convolutional network adopts a ConvNeXt-ViT hybrid backbone architecture: The bottom layer employs 6 layers of ConvNeXt-Tiny convolutional blocks with a kernel size of 3×3 and the number of channels gradually increasing from 64 to 256. Depthwise separable convolutions are used to reduce computational load while capturing local structural features of the image; these structural features include edges and textures. Middle layer: Embedded with 4 cross-scale attention fusion blocks, which associate 1 / 2, 1 / 4, and 1 / 8 scale feature maps across layers, and model the dependencies of global feature layout through a self-attention mechanism; the global feature layout includes building outlines and road directions; Top layer: A global feature aggregation module is used to perform weighted summation on feature maps at various scales, and finally generate a shared content encoding vector Zc with dimension Dc=2560, ensuring that it covers both local details and global structural information.

4. The method for generating labels for identifying change patches in flood storage and detention areas based on style decoupling according to claim 3, characterized in that: In step S3: The independent branch network is a lightweight Transformer + gated convolutional symmetric structure: Each branch contains 3 layers of MobileViT blocks, which capture the local relevance of style information through local window attention; Two gated convolutional layers are then concatenated, and a channel-gated vector is generated using the Sigmoid activation function to select feature channels that are strongly correlated with illumination, season, and imaging sensor type, while suppressing irrelevant noise. Finally, a single 1×1 convolution layer is used to compress the feature dimension to Ds=384, and I is output respectively. pre I post The style encoding vectors Zs1 and Zs2 are used to achieve accurate capture of style information and dimensional optimization.

5. The method for generating labels for identifying change patches in flood storage and detention areas based on style decoupling according to claim 1, characterized in that: In step S3: the mathematical expression for the loss function is: ； In the formula, L ortho Represents the mixed orthogonal loss function. The weighting coefficients of the cosine similarity loss Lcos are represented by... L represents the weighting coefficient of the cosine similarity loss Lmi, Lcos represents the cosine similarity loss, and Lmi represents the mutual information loss. The cosine similarity loss formula Lcos is as follows: ； In the formula, N represents the batch sample size. This represents the content feature vector of the i-th sample. This represents the first style feature vector of the i-th sample. Let represent the second style feature vector of the i-th sample, cos(.,.) represent the cosine similarity function, and cos2(.,.) represent the squared cosine similarity. The formula for mutual information loss Lmi is as follows: ； In the formula, MI(,.,) represents the mutual information calculation function.

6. The method for generating labels for identifying change patches in flood storage and detention areas based on style decoupling according to claim 1, characterized in that: The label selection process in S7 is as follows: First, the distribution distance between the generated image in the expanded label set and the original image obtained after preprocessing in step S1 in the depth feature space is calculated, and the physical realism is evaluated using the FID index; Second, the generated image is input to the pre-trained change detection network, and the IoU score between the prediction mask and the input change mask is calculated to evaluate semantic consistency; Finally, unqualified labels are automatically filtered based on the comprehensive quality score of FID and IoU.

7. A remote sensing change detection data augmentation system based on a style decoupling generative model, the system comprising: The data preprocessing module is configured to perform step S1 as described in claim 1; The model building module is configured to perform step S2 as described in claim 1; The decoupled training module is configured to perform steps S3 and S4 as described in claim 1; The controllable generation module is configured to perform steps S5 and S6 as described in claim 1; The quality assessment and output module is configured to perform step S7 as described in claim 1.

8. An electronic device, characterized in that: The electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the method of claim 1.

9. A computer-readable storage medium, characterized in that: A computer program is stored on a computer-readable storage medium, and when executed by a processor, the computer program implements the steps of the method of claim 1.

Images