Sandstorm image and label synchronous generation method and system based on cross-frequency coordination and coordinate perception

Abstract

The application belongs to the field of computer vision and image processing, and discloses a sandstorm image and label synchronous generation method and system based on cross-frequency cooperation and coordinate perception, which comprises the following steps: obtaining FY-4A AGRI sandstorm remote sensing images and their corresponding semantic label masks, and performing pretreatment; based on the pretreated FY-4A AGRI sandstorm remote sensing images and their corresponding semantic label masks, constructing a cross-frequency cooperation and coordinate perception generative adversarial network model; based on the cross-frequency cooperation and coordinate perception generative adversarial network model, generating sandstorm expansion samples with corresponding semantic labels, which meet preset requirements in structure, texture and morphology.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and image processing, specifically relating to a method and system for synchronously generating sandstorm images and labels based on cross-frequency collaboration and coordinate perception. Background Technology

[0002] Reviewing existing research, image sample construction and augmentation methods can be mainly divided into three categories: manual sample construction methods based on real satellite observation images, traditional data augmentation methods based on geometric or color transformations, and image generation methods based on generative models. Sample construction methods based on real satellite observation images can maintain the authenticity of dust storm remote sensing images relatively well, but their sample quality is limited by factors such as cloud cover, lighting conditions, imaging time, observation angle, the suddenness of dust storm processes, and the cost of manual pixel-by-pixel annotation. The number of high-quality labeled and effective samples that can be directly used for deep learning model training is limited, and it is difficult to fully cover complex scenes with different dust storm intensities, different surface backgrounds, different spatial morphologies, and different boundary structures.

[0003] The Advanced Radiation Imager (AGRI) onboard the Fengyun-4A FY-4A satellite can acquire large-scale, multi-channel, and continuous remote sensing data, providing a data foundation for identifying the spatial distribution, range extraction, and transport process analysis of dust storms. Specifically, AGRI's NOMChannel01, NOMChannel02, and NOMChannel03 cover the visible to near-infrared shortwave bands, reflecting differences in brightness, color, texture, and spatial morphology of dust regions. By using NOMChannel03, NOMChannel02, and NOMChannel01 as R, G, and B channels respectively for pseudo-RGB synthesis, the separability between dust regions and the surface background and clouds can be enhanced, providing an effective data source for constructing dust storm remote sensing image samples and training intelligent recognition models. However, each FY-4A AGRI image typically only reflects the dust state at a specific time, in a specific region, and under specific observation conditions, making it difficult to directly characterize the changing processes of dust storms across multiple scenes, intensities, and morphologies.

[0004] Existing traditional image augmentation methods often employ simple transformations of existing samples, such as rotation, cropping, flipping, scaling, or color perturbation. While these methods are simple to implement, they essentially only alter the appearance of existing samples and struggle to generate new samples that reflect the true spatial distribution of dust storms, edge morphology variations, and local texture differences. They also fail to effectively expand the morphological diversity of dust storm regions under complex background conditions. For dust storm remote sensing images, dust regions typically exhibit blurred boundaries, strong spatial diffusion, complex textures, and are easily confused with clouds and bare ground backgrounds. Simple data augmentation methods are insufficient to meet the diverse training sample requirements of dust detection and semantic segmentation models.

[0005] While single-image generation methods based on Generative Adversarial Networks (GANs) or SinGANs can learn image distributions and generate a certain number of new samples under limited sample conditions, existing methods lack the ability to co-model low-frequency large-scale spatial contours, high-frequency edge details, and local texture structures in dust storm remote sensing images. This leads to problems such as blurred dust boundaries, local texture distortion, spatial discontinuities, and confusion between dust and background areas. Although diffusion models have strong detail generation capabilities, their generation process is computationally expensive, leading to high training costs, low convergence efficiency, and insufficient structural constraints under limited remote sensing sample conditions. Furthermore, existing image generation methods typically only generate the remote sensing image itself, making it difficult to simultaneously generate semantic label masks that strictly correspond to the image's spatial structure. Subsequent manual annotation is still required, failing to truly reduce the cost of constructing a dust storm sample library.

[0006] To address the above issues, this invention proposes a method for the simultaneous generation of FY-4A sandstorm remote sensing images and semantic labels based on cross-frequency collaborative and coordinate-aware GAN. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a method and system for synchronously generating dust storm images and labels based on cross-frequency collaboration and coordinate awareness. Under the condition of limited high-quality labeled samples, it can generate dust storm augmentation samples with reasonable structure, continuous texture, diverse morphology and corresponding semantic labels, providing new technical means for dust storm detection, semantic segmentation model training and remote sensing sample library construction.

[0008] To achieve the above objectives, the present invention provides the following solution: A method for synchronously generating sandstorm images and labels based on cross-frequency collaboration and coordinate awareness, the method comprising: Acquire FY-4A AGRI dust storm remote sensing images and their corresponding semantic label masks, and perform preprocessing; Based on the preprocessed FY-4A AGRI dust storm remote sensing images and their corresponding semantic label masks, a cross-frequency collaborative and coordinate-aware generative adversarial network model is constructed. Based on cross-frequency collaboration and coordinate-aware generative adversarial network models, sandstorm augmentation samples are generated that meet preset requirements in terms of structure, texture, and morphology and have corresponding semantic labels.

[0009] Preferred methods for preprocessing FY-4A AGRI dust storm remote sensing images and their corresponding semantic label masks include: The raw HDF data of FY-4A AGRI is parsed to obtain the observation data of the required channels. Combined with the FY-4A AGRI 4 km full disk latitude and longitude lookup table FullMask_Grid_4000.raw, the latitude and longitude matrices corresponding to the row and column positions of the original observation grid are obtained. The correspondence between the row and column coordinates of the original scanned pixels and the geographic latitude and longitude coordinates is established. The observation data of each channel are uniformly mapped to the target regular latitude and longitude grid, thereby completing the geometric positioning and spatial registration of FY-4A AGRI image. The raw digital quantization values ​​were radiometrically calibrated based on the calibration parameters of the FY-4A AGRI product. For invalid pixels in the raw observation data, comprehensive identification and processing are performed based on the FY-4A AGRI product data structure and calibration results. Specifically, the raw digital quantization values ​​65534 and 65535 in NOMChannel are identified as invalid product codes, and their corresponding pixels are set to null values ​​during radiometric calibration. At the same time, invalid items, non-finite values, and outliers appearing after calibration are removed from the calibration lookup table. Furthermore, an effective Earth observation mask is constructed by combining the effective observation column range of each scan row recorded in NOMObsColumn to remove pixels outside the Earth and invalid scan areas. After completing radiometric calibration, geometric correction and invalid pixel processing, the FY-4A AGRI image was cropped based on the dust storm occurrence area and the study area to obtain a dust storm remote sensing image that includes the main dust storm area and the surrounding background. The cropped image is resampled to convert samples from different times and regions into regular grid samples of the same size. At the same time, the sample size is uniformly set to 256×256 pixels. NOMChannel01, NOMChannel02, and NOMChannel03 were selected for pseudo-RGB synthesis. Specifically, NOMChannel03 was used as the R channel, NOMChannel02 as the G channel, and NOMChannel01 as the B channel to construct a three-channel sandstorm remote sensing image. For semantic labeling masks, pixel-level annotations are performed based on the distribution of dust areas in remote sensing images of dust storms. Specifically, the dust storm occurrence area is determined by visual interpretation or existing annotation results, the dust area is marked as the target category, and the non-dust area is marked as the background category, thus forming a single-channel semantic labeling mask.

[0010] Preferably, the cross-frequency collaborative and coordinate-aware generative adversarial network model includes: a high- and low-frequency collaborative feature diffusion generator, a coordinate-aware cross-attention discriminator, and an image and label joint generation mechanism; A high- and low-frequency collaborative feature diffusion generator is used to learn the large-scale spatial structure, edge morphology, and local texture features of dust regions from joint samples of imagery and label four channels; A coordinate-aware cross-attention discriminator is used to discriminate and constrain generated samples based on spatial positional relationships, region boundary morphology, and local texture representation. An image and label joint generation mechanism is used to maintain spatial consistency between the generated dust image and the corresponding semantic label mask.

[0011] Preferably, in the low-resolution generation stage, the high-low frequency collaborative feature diffusion generator adopts a multi-scale pyramid generation structure, and each generation unit includes three basic parts: a head feature extraction module, a main feature extraction module, and a tail output module. The head feature extraction module is used to receive input features at the current scale and complete the initial feature mapping through convolutional layers, normalization layers and Leaky-ReLU activation function. The main feature extraction module consists of multi-layer convolutional blocks or feature enhancement modules, used to extract and enhance the spatial structure, edge contours and texture features of the dust region; The tail output module consists of a convolutional layer and a Tanh activation function, which is used to map the generated features into a four-channel joint output, which includes a three-channel dust image and a one-channel dust label. In the medium-to-high resolution generation stage, the high-to-low frequency collaborative feature diffusion generator sequentially embeds the high-to-low frequency feature decoupling module HLFDM based on Haar wavelet decomposition, the dual-frequency cross-shaped window attention fusion module DF-CWAFM, the cross-frequency collaborative attention enhancement module CF-AEM, and the diffusion feature refinement module DFRM into the basic generation unit. Among them, the high- and low-frequency feature decoupling module HLFDM based on Haar wavelet decomposition is used to decompose the current scale input features into low-frequency structural components and high-frequency detail components. The dual-frequency cross-shaped window attention fusion module DF-CWAFM is used to establish a complementary relationship between low-frequency structural information and high-frequency texture information; The Cross-Frequency Collaborative Attention Enhancement (CF-AEM) module is used to collaboratively enhance cross-frequency fusion features from both channel and spatial dimensions. The Diffusion Feature Refinement (DFRM) module is used to perform perturbation recovery and detail refinement on the fused generated features, suppress redundant background responses, and enhance the continuity of dust edges and local texture stability.

[0012] Preferably, the coordinate-aware cross-attention discriminator enhances the perception of differences between real and generated features by introducing a cross-attention mechanism, and strengthens spatial structure constraints by combining coordinate position information; specifically: The discriminator receives both real and generated samples. The real and generated samples extract shallow features through the Head module with shared parameters to obtain two representations. Each discriminator consists of three parts: a head feature extraction module, a main feature extraction module, and a tail output module. It also incorporates a multi-head cross-attention mechanism to enhance the ability to model global dependencies and capture local textures in the dusty area. The head feature extraction module extracts preliminary features through 3×3 convolution, spectral normalization, Leaky-ReLU activation and global average pooling, while generating queries and key-value pairs for attention calculation, thereby establishing a global association between the features of the generated image and the real image. The main feature extraction module integrates attention-enhanced features with convolutional features, and further extracts local texture, boundary transition and regional structure information through multi-layer convolution to improve the discriminator's ability to perceive the subtle differences in sand and dust areas. The tail output module consists of convolutional layers and global average pooling, outputting unnormalized real-valued discrimination scores to provide adversarial constraints at the structural and texture levels for the generator.

[0013] Preferably, the image and label joint generation mechanism models three-channel dust storm remote sensing images and single-channel semantic label masks as unified joint samples, enabling the generator to simultaneously learn the joint distribution relationship between images and labels; specifically: Original sandstorm images With corresponding tags By concatenating the input channels, a four-channel joint input is constructed: ; in, This represents the input image of a sandstorm. This indicates the corresponding semantic tag for sand and dust. This indicates a splicing operation at the channel level. This represents the constructed image-label joint sample. and These represent the height and width of the feature map, respectively.

[0014] The present invention also provides a sandstorm image and tag synchronous generation system based on cross-frequency collaboration and coordinate perception. The system is used to implement the aforementioned method and includes: a preprocessing module, a construction module, and a generation module. The preprocessing module is used to acquire FY-4A AGRI dust storm remote sensing images and their corresponding semantic label masks, and to perform preprocessing. The module is used to build a cross-frequency collaborative and coordinate-aware generative adversarial network model based on preprocessed FY-4A AGRI dust storm remote sensing images and their corresponding semantic label masks. The generation module is used to generate sandstorm augmentation samples with structure, texture, and morphology that meet preset requirements and have corresponding semantic labels, based on cross-frequency collaborative and coordinate-aware generative adversarial network models.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a method for the simultaneous generation of FY-4A dust storm remote sensing images and semantic labels based on a Cross-Frequency Collaborative and Coordinate-Aware Generative Adversarial Network (CFCA-GAN). This method takes FY-4A AGRI dust storm remote sensing images and their corresponding semantic label masks as input, and uses the image channels and label channels as unified samples for joint modeling. This allows the model to learn the spatial structure, edge morphology, and local texture features of the dust storm region while simultaneously learning the pixel-level correspondence between the image and the label, thus achieving the synchronous generation of the dust storm remote sensing image and the semantic label mask. The proposed CFCA-GAN mainly consists of a High-Frequency Collaborative Feature Diffusion Generator (HLF-CFDG), a Coordinate-Aware Cross-Attention Discriminator (CACA-D), and a joint image and label generation mechanism. HLF-CFDG, through high- and low-frequency feature decoupling, cross-frequency attention fusion, cross-frequency collaborative enhancement, and diffuse feature refinement, jointly models the large-scale spatial distribution, edge structure, and local texture of dust storm regions. CACA-D, through cross-attention and coordinate-aware mechanisms, discriminates and constrains the generated results based on spatial location, regional boundaries, and texture structure. The image and label joint generation mechanism maintains spatial consistency between the generated dust storm image and the corresponding semantic label mask. It should be noted that this invention uses the dust storm remote sensing image and the semantic label mask as a unified sample for joint modeling, rather than simply generating the image itself. Through joint input and synchronous output of the image and label channels, the model can obtain a semantic label mask consistent with its spatial structure while generating the dust storm remote sensing image, thereby reducing subsequent manual pixel-by-pixel annotation. This method can generate structurally reasonable, textured, morphologically diverse, and accurately labeled extended samples under limited high-quality labeled sample conditions, providing data support for dust storm detection, semantic segmentation model training, and sample library construction. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of a method for synchronously generating sandstorm images and tags based on cross-frequency collaboration and coordinate sensing, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of a cross-frequency collaborative and coordinate-aware generative adversarial network according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the high- and low-frequency feature decoupling module based on Haar wavelet decomposition in an embodiment of the present invention. Figure 4 This is a schematic diagram of the dual-frequency cross-shaped window attention fusion module structure according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the cross-frequency collaborative attention enhancement module structure according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of the diffusion-type feature refinement module in an embodiment of the present invention; Figure 7 This is a schematic diagram of the coordinate-aware cross-attention discriminator structure according to an embodiment of the present invention; Figure 8 This is a schematic diagram comparing a real image with labels and a generated image with labels according to an embodiment of the present invention; Figure 9 This is a comparison chart of the results of the algorithm proposed in this embodiment of the invention with other algorithms. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] Terminology Explanation: FY-4A: Fengyun-4A refers to the Fengyun-4A satellite, a three-axis stabilized geostationary orbit meteorological satellite independently developed by my country and launched in December 2016. FY-4A can conduct large-scale, multi-channel, and continuous Earth remote sensing observations using its onboard payloads, including an advanced radiation imager. AGRI: short for Advanced Geosynchronous Radiation Imager, is a multispectral imaging payload carried by the FY-4A satellite. This payload has 14 spectral channels, covering a band range of approximately 0.45–13.8 μm, with spatial resolutions of 0.5 km, 1 km, 2 km, 2–4 km, and 4 km. It can acquire multi-channel remote sensing data in visible, near-infrared, mid-infrared, and long-wave infrared light.

[0021] NOMObsColumn: The nominal observation column range dataset, used to record the start and end columns of the effective observation area of ​​Earth in each scan row of the FY-4A AGRI raw image. This data can be used to construct an effective observation mask for Earth, thereby eliminating pixels outside the Earth and invalid scan areas.

[0022] GAN: Generative Adversarial Network, which consists of a generator and a discriminator. It generates samples that are similar to the distribution of real data through adversarial training.

[0023] SinGAN: short for Single Image Generative Adversarial Network, is a generative model trained on a single image. It can learn the multi-scale structure and texture features of the image and generate new samples with a style similar to the original image.

[0024] Diffusion Model: A depth generation model that generates images through "stepwise noise addition" and "stepwise noise reduction," capable of generating high-quality, detailed image samples.

[0025] MLP stands for Multi-Layer Perceptron, a model composed of multiple fully connected neural networks that can perform nonlinear mapping and feature transformation on input features to enhance the model's feature representation capabilities.

[0026] Leaky-ReLU: Leaky-corrected linear unit is a neural network activation function used to enhance the non-linear expressiveness of a model and improve training stability.

[0027] Example 1 Existing methods for augmenting remote sensing images of dust storms can be mainly divided into two categories: traditional data augmentation methods and generative model-based data augmentation methods. Traditional data augmentation methods typically use techniques such as rotation, flipping, cropping, scaling, brightness adjustment, and color perturbation to augment the original samples. These methods are simple in principle and easy to implement, but they mainly involve geometric or color changes based on existing samples, making it difficult to generate new spatial morphologies of dust storms and failing to adequately increase the structural diversity and background complexity of the samples.

[0028] Generative model-based data augmentation methods can learn the data distribution of real samples and generate new samples with certain differences. Among them, single-image generation methods such as SinGAN can generate images under small sample conditions and are suitable for data augmentation under limited dust storm sample conditions. Improved SinGAN and HA-SinGAN methods have improved the detail representation ability of generated images to a certain extent by introducing attention mechanisms or structure enhancement methods. However, dust storm remote sensing images are characterized by blurred boundaries, complex morphological changes, uneven spatial distribution, and easy confusion with cloud and surface backgrounds. When existing generation methods are directly used to generate dust storm remote sensing images, problems such as unclear dust edges, unstable local textures, and insufficient overall structural continuity still easily occur.

[0029] Furthermore, most existing generation methods primarily focus on generating the image itself, failing to simultaneously generate semantic labels consistent with the spatial structure of the dust storm image. Therefore, when using generated images for downstream dust storm detection and segmentation tasks, additional manual annotation is still required, increasing sample construction costs and limiting the efficiency of generating samples in dust storm detection and segmentation tasks.

[0030] To address the above problems, this invention provides a method for synchronously generating dust storm images and labels based on cross-frequency collaboration and coordinate awareness. Using FY-4A AGRI dust storm remote sensing images and their corresponding semantic label masks as input, the remote sensing images and label masks are jointly modeled along the channel dimension to construct a data representation method for synchronous image and label generation. In terms of network structure, this invention designs a high- and low-frequency collaborative feature diffusion generator and a coordinate-aware cross-attention discriminator. The high- and low-frequency collaborative feature diffusion generator achieves explicit decoupling of low-frequency structural information and high-frequency detail information through Haar wavelet decomposition, and combines a dual-frequency cross-shaped window attention fusion module, a cross-frequency collaborative attention enhancement module, and a diffusion-based feature refinement module to progressively model the overall spatial distribution, edge morphology, and local texture of the dust region. The coordinate-aware cross-attention discriminator, through cross-attention and coordinate awareness mechanisms, discriminates and constrains the generated results from aspects such as spatial structure, boundary morphology, and local texture. Through the above technical solution, the present invention can generate sandstorm augmentation samples with reasonable structure, continuous texture, diverse morphology and corresponding semantic labels under the condition of limited high-quality labeled samples, providing new technical means for sandstorm detection, semantic segmentation model training and remote sensing sample library construction.

[0031] The process of using this framework to augment sandstorm image samples mainly consists of four stages: 1. Data preprocessing; 2. Constructing a cross-frequency collaborative and coordinate-aware generative adversarial network model; 3. Model training and parameter optimization; 4. Model testing and augmented sample output.

[0032] The specific process of this method is as follows: Figure 1 As shown: The specific steps are as follows: 1. Data Preprocessing This invention first preprocesses FY-4A AGRI dust storm remote sensing images and their corresponding semantic label masks. The purpose of data preprocessing is mainly to extract input samples suitable for model training from the raw FY-4A AGRI multispectral remote sensing data and construct joint image and label samples. It is particularly important to note that the preprocessing stage does not only extract dust storm remote sensing images, but also stitches the dust storm images and their corresponding semantic label masks along the channel dimension to form four-channel joint samples, which serve as input data for the subsequent CFCA-GAN model. The remote sensing data used is FY-4A AGRI L1 level full-disk multispectral remote sensing image data. The raw FY-4A AGRI data is stored in HDF format, with a satellite nadir longitude of 104.7°E, a nominal projection, and observation time expressed in Coordinated Universal Time (UTC). AGRI has 14 spectral channels covering a band size of approximately 0.45–13.8 μm, enabling the acquisition of multi-channel remote sensing data including visible, near-infrared, mid-infrared, and long-wave infrared. Among these, NOMChannel01, NOMChannel02, and NOMChannel03, located in the visible to near-infrared short-wave band, effectively reflect the brightness, color, texture, and spatial morphological differences of dust storm regions. Therefore, this invention preferentially uses these three channels to construct dust storm remote sensing image samples.

[0033] First, the raw HDF format data of FY-4A AGRI is parsed to extract the required observation channels and perform pixel-level geolocation. Specifically, for each FY-4A AGRI L1-level HDF file, the raw observation data of NOMChannel01, NOMChannel02, and NOMChannel03 are read, along with the corresponding CALChannel01, CALChannel02, and CALChannel03 calibration lookup tables. For the k-th channel, its raw digital quantization matrix can be represented as: (1) In the formula, Indicates the first The original pixel matrix of each AGRI channel; Indicates the first The first channel in the line, number The original quantized value of the number in the column; and These represent the number of rows and columns of the original image, respectively. To convert the original digital quantized values ​​into physically meaningful observations, calibration is performed using the CALChannel calibration lookup table for the corresponding channel. This process can be represented as follows: (2) In the formula, Indicates the first The first channel in the line, number The calibrated observations are listed below; Indicates the first Each channel corresponds to a calibration lookup table. For invalid pixel values ​​65534 and 65535 in the original data, as well as outliers or non-finite values ​​that appear after calibration, they are set to null values ​​and do not participate in subsequent spatial registration and sample construction.

[0034] Subsequently, the FY-4A AGRI 4 km full-disk latitude and longitude lookup table FullMask_Grid_4000.raw was read. This lookup table has the same row and column dimensions as the original FY-4A AGRI 4 km observation grid, both being 2748×2748. The lookup table stores the latitude and longitude information corresponding to each original scanned pixel, which can be represented as follows: (3) In the formula, and These represent the original image number 1 and 2 respectively. line, number The table lists the latitude and longitude corresponding to the center of each pixel. Using this latitude and longitude lookup table, the correspondence between the row and column numbers of the FY-4A AGRI raw scan and their geographic coordinates can be established. (4) In the formula, Indicates the first line, number The multi-channel calibration observations of each pixel and their corresponding geographical locations are obtained. This transforms the scanned row and column coordinates of the original HDF data into geolocated pixels with clearly defined latitude and longitude, achieving pixel-level geometric positioning.

[0035] To further eliminate areas outside the Earth and invalid observations, the NOMObsColumn information in the HDF file is read to construct an effective Earth observation mask. For each row, the effective Earth observation range within that row is determined based on the effective start and end columns provided by NOMObsColumn. If the start or end columns are invalid, exceed the limits, or the end column is smaller than the start column, the corresponding row's pixels are marked as invalid. This yields the effective Earth pixel mask. Simultaneously, by combining null values, abnormal latitude and longitude values ​​from the latitude and longitude lookup table, and null values ​​after channel calibration, a comprehensive effective pixel condition is constructed.

[0036] After completing radiometric calibration, geometric localization, and invalid pixel processing, the FY-4A AGRI imagery was cropped and resampled using regular grids based on the dust storm occurrence area and the latitude and longitude range of the study area. Since the original FY-4A AGRI imagery is full-disk observation data, directly inputting it into the model would include a large amount of background areas unrelated to the dust storm process. Therefore, it is necessary to extract local image samples containing the main dust storm area and its surrounding background based on the study area's latitude and longitude range. Specifically, based on the longitude range of the study area... Latitude range is To ensure sufficient neighboring source pixels participate in interpolation at the boundary of the study area and to reduce the instability of edge resampling, a latitude and longitude buffer range is set outside the study area. The set of source pixels used for region cropping and resampling can then be represented as: (5) In the formula, and These represent the original image number 1 and 2 respectively. line, number The longitude and latitude corresponding to each pixel; Indicates the first The overall effective pixel mask for each channel; This indicates the buffer zone extending beyond the study area. (During the experiment...) Set to 1.0°.

[0037] By filtering according to the above criteria, effective source pixels of the study area and its adjacent areas can be extracted from the full-disk FY-4A AGRI observation data. This process is not simply cropping according to the original row and column numbers, but rather spatially filtering based on the latitude and longitude position of each pixel, thus ensuring that the cropped area is consistent with the actual geographical extent. Subsequently, a target regular latitude and longitude grid is constructed to uniformly map the filtered source pixels to regular image samples of a fixed size. Let the target grid width be... The height is Then the first one in the target grid Column, No. The longitude and latitude of the row cell center are defined as follows: (6) (7) In the formula, Indicates the target mesh number Longitude of the pixel center; Indicates the target mesh number The latitude of the row cell center. The latitude direction is from... Decrease to This ensures that the output image is oriented north-facing. In the experiment, , This means that each sample is uniformly resampled to 256×256 pixels.

[0038] For the Each channel will collect the source pixels. Using the latitude and longitude locations in the data as spatial interpolation points, the corresponding calibration observations are... The pixel values ​​on the target regular grid are calculated using a linear interpolation method as the interpolation values. This resampling process can be represented as: (8) in, This represents a linear interpolation operator. and This represents the latitude and longitude coordinates on the regular grid after resampling. This process maps the effective observation cells in the original FY-4A AGRI scan grid to a unified regular latitude and longitude grid, thereby achieving spatial registration and size uniformity.

[0039] For target grid cells that remain empty after linear interpolation, nearest neighbor interpolation is further used to complete them, in order to avoid invalid holes within the target sample. Specifically, the first... The linear interpolation result of each channel is denoted as When the result is null or invalid, in the first... The effective source pixel set of each channel In the search and target grid points The original valid cell with the closest geographical location. This nearest neighbor source cell can be represented as: (9) In the formula, This indicates the location of the nearest original valid source cell to the target regular grid point; This indicates the cell index that minimizes the distance metric within the parentheses; Indicates the first The set of valid source pixels for each channel; and These represent the longitude and latitude of the original source pixel, respectively. and These represent the longitude and latitude corresponding to the target rule grid points, respectively.

[0040] Finally, the first The resampling result of each channel on the target regular grid is defined as: (10) In the formula, Indicates the first The first channel in the target rule grid Column, No. The final resampling result for the row; This represents the pixel value obtained through linear interpolation; Indicates the first The original effective source pixel that is closest to the target grid point in each channel; This represents the radiometrically calibrated observation value at the nearest neighbor valid source pixel. Through the above processing, linear interpolation results are preferentially used to maintain the continuity of spatial variations in the image; when linear interpolation cannot obtain valid values, the observation values ​​of the nearest neighbor valid source pixels are used to complete the image, thereby reducing the number of missing pixels in the target regular grid and improving the integrity of the resampled image samples.

[0041] After completing region cropping and regular grid resampling for each channel, FY-4A AGRI dust storm remote sensing image samples with uniform spatial range and size can be obtained. For the first... The resampled image matrix of each channel can be represented as: (11) In the formula, Indicates the first The image matrix after region cropping and regular grid resampling of each channel; Indicates the first The first channel in the target rule grid Column, No. The cell value at the row; and These represent the height and width of the target regular mesh, respectively. In the experiment, , Each channel was then uniformly converted into a 256×256 pixel regular grid image to meet the input size consistency requirements of the generative adversarial network model. Since FY-4AAGRI does not have standard red, green, and blue channels that perfectly correspond to human visual response, NOMChannel01, NOMChannel02, and NOMChannel03 were selected for pseudo-RGB synthesis to facilitate dust storm area identification and model training sample construction. Specifically, NOMChannel03 was used as the R channel, NOMChannel02 as the G channel, and NOMChannel01 as the B channel to construct a three-channel dust storm remote sensing image. This channel combination enhances the separability between the dust storm area and the surface background and clouds, while preserving the overall outline, edge morphology, and local texture information of the dust storm area.

[0042] For the semantic labeling mask, pixel-level annotations are performed based on the distribution of dust regions in the dust storm remote sensing image. Specifically, the dust storm occurrence area is determined through visual interpretation or existing annotation results, and the dust regions are labeled as the target category, while non-dust regions are labeled as the background category, forming a single-channel semantic labeling mask. This semantic labeling mask is used to characterize the spatial distribution range of dust regions in the image and serves as supervision information for subsequent image and label synchronous generation models.

[0043] 2. Construct a cross-frequency collaborative and coordinate-aware generative adversarial network model. The cross-frequency collaborative and coordinate-aware generative adversarial network model consists of a high- and low-frequency collaborative feature diffusion generator, a coordinate-aware cross-attention discriminator, and an image-label joint generation mechanism. Specifically, the high- and low-frequency collaborative feature diffusion generator learns the large-scale spatial structure, edge morphology, and local texture features of the dust region from joint samples from the four channels of the image and label. The coordinate-aware cross-attention discriminator constrains the generated samples based on spatial relationships, regional boundary morphology, and local texture representation. The image-label joint generation mechanism maintains spatial consistency between the generated dust image and the corresponding semantic label mask.

[0044] A schematic diagram of the cross-frequency collaboration and coordinate-aware generative adversarial network model is shown below: The proposed multimodal cross-branch fusion network, such as Figure 2 As shown, specifically, this model mainly consists of three parts: a high- and low-frequency collaborative feature diffusion generator, a coordinate-aware cross-attention discriminator, and an image and label joint generation part. Each part will be described in detail below.

[0045] The first part is the High-Low Frequency Collaborative Feature Diffusion Generator (HLF-CFDG), which employs a multi-scale pyramid-shaped generation structure to generate dust images and their corresponding labels progressively from low resolution to high resolution. Each generation unit at each scale comprises three basic parts: a head feature extraction module, a main feature extraction module, and a tail output module. The head feature extraction module receives the input features at the current scale and performs preliminary feature mapping through convolutional layers, normalization layers, and the Leaky-ReLU activation function. The main feature extraction module consists of multiple convolutional blocks or feature enhancement modules, used to extract and enhance the spatial structure, edge contours, and texture features of the dust region. The tail output module consists of convolutional layers and the Tanh activation function, used to map the generated features into a four-channel joint output, which includes a three-channel dust image and a one-channel dust label. In the low-resolution generation stage, since the input features mainly characterize the overall spatial layout and coarse-scale structure of the dust region, and do not yet contain stable edge textures and local details, a frequency decomposition module is not introduced at this stage. Instead, the global contour and basic spatial distribution of the dust region are initialized through a head feature extraction module, a main feature extraction module, and a tail output module. This setting avoids premature high- and low-frequency splitting at the initial scale, which would weaken the global semantic structure expression. As the generation scale gradually increases, the current scale features gradually contain richer edge structures, local textures, and regional transition information. At this point, relying solely on convolution operations is insufficient to simultaneously maintain the consistency of large-scale spatial structure and the authenticity of local details. Therefore, in the medium- and high-resolution generation stage, the HLF-CFDG sequentially embeds a high- and low-frequency feature decoupling module HLFDM based on Haar wavelet decomposition, a dual-frequency cross-shaped window attention fusion module DF-CWAFM, a cross-frequency collaborative attention enhancement module CF-AEM, and a diffusion-based feature refinement module DFRM into the basic generation unit. Specifically, HLFDM decomposes the current-scale input features into low-frequency structural components and high-frequency detail components; DF-CWAFM establishes a complementary relationship between low-frequency structural information and high-frequency texture information; CF-AEM collaboratively enhances cross-frequency fusion features from both channel and spatial dimensions; and DFRM performs perturbation recovery and detail refinement on the fused generated features, thereby suppressing redundant background responses and enhancing the continuity of dust edges and the stability of local textures. During multi-scale generation, the generation result from the previous scale is upsampled and used together with random noise at the current scale as input to the current generation unit, allowing the model to gradually supplement high-scale detail information while inheriting the low-scale global structure. Through this hierarchical generation method, HLF-CFDG can improve the edge contour clarity, texture expression integrity, and spatial consistency between the generated image and the label while ensuring the overall spatial morphology of the dust region.

[0046] ① High- and low-frequency feature decoupling module based on Haar wavelet decomposition To extract different frequency information from the generated features of FY-4A dust storm remote sensing images, this invention incorporates a high-low frequency feature decoupling module based on Haar wavelet decomposition within the high-low frequency collaborative feature diffusion generator. This module performs a two-dimensional discrete wavelet transform on the input feature map, decomposing the generated features at the current scale into low-frequency structural components and high-frequency detail components, thereby modeling the large-scale spatial distribution, overall contour, edge structure, and local texture information of the dust region.

[0047] Specifically, such as Figure 3 As shown, given the input feature map at the current generation stage ,in , Indicates the number of channels. and These represent the height and width of the feature map, respectively. First, for the input feature map... Perform a two-dimensional Haar discrete wavelet transform on each channel to decompose it into a low-frequency approximate subband. and three high-frequency detail sub-bands , , Mid- and low-frequency approximate sub-bands It is mainly used to characterize the overall spatial distribution, main morphology and large-scale continuous structure of dust areas; the high-frequency detail sub-bands in the horizontal, vertical and diagonal directions are mainly used to characterize the edge contour, local texture undulation and detail changes of dust.

[0048] The two-dimensional Haar wavelet decomposition process can be represented as: Low-frequency approximate subband: (12) Horizontal high-frequency detail subband: (13) Vertical high-frequency detail subband: (14) Diagonal high-frequency detail subband: (15) in, This indicates that the input feature map is at the location The characteristic response at that location, and These represent the low-pass and high-pass filters of the Haar wavelet, respectively. This is the position index of the output subband, with a downsampling factor of 2.

[0049] The filter coefficients of the Haar wavelet are defined as follows: Low-pass filter: (16) High-pass filter: (17) After the wavelet decomposition described above, the low-frequency structural features are obtained. and three high-frequency detail features , , To avoid insufficient expression of structural or detailed information due to the use of fixed weights for different frequency subbands, frequency-aware modulation branches are further set for low-frequency and high-frequency components respectively.

[0050] For low-frequency approximate subband A frequency-aware adaptive gain modulation module is used to generate a low-frequency response modulation coefficient diagram. This modulation coefficient diagram... This is used to enhance the effective response in low-frequency components related to the main morphology, spatial continuity, and large-scale structure of dust. The calculation process is as follows: (18) (19) in, This represents the convolution mapping function in the low-frequency branch; Learnable parameters representing the low-frequency branch; This represents the Sigmoid activation function; This represents the low-frequency response modulation coefficient diagram; This indicates element-wise multiplication.

[0051] For high-frequency components, a frequency-aware adaptive detail modulation module is introduced to enhance the representation of edges, textures, and local details. Specifically, the three high-frequency sub-bands are first jointly modeled, a process that can be represented as a feature... : (20) in, This represents a feature concatenation operation along the channel dimension; Represents the high-frequency joint mapping function; This represents the learnable parameters in the high-frequency joint mapping.

[0052] Subsequently, using convolution mapping and Activation function generates high-frequency response modulation coefficient diagram and high-frequency joint features Perform element-wise modulation: (twenty one) (twenty two) in, This represents the convolution mapping function in the high-frequency detail modulation branch; This represents the learnable parameters in the high-frequency detail modulation branch; This represents a high-frequency response modulation coefficient diagram. This indicates the high-frequency detail features after modulation. This modulation process is mainly used to enhance the response of dust edge contours, texture variation areas, and local detail areas, thereby improving the edge continuity and texture representation of the generated image.

[0053] After modulating the low-frequency structural features and high-frequency detail features, HLFDM does not directly fuse the two into a single feature. Instead, it modulates the low-frequency structural features... and high-frequency detail features As a dual-branch feature output: (twenty three) in, , These represent the low-frequency structural branch features and high-frequency detail branch features of the HLFDM output, respectively. These two feature paths serve as inputs to the subsequent dual-frequency cross-window attention fusion module DF-CWAFM, further establishing the complementary relationship between low-frequency structural information and high-frequency detail information.

[0054] Through the above processing, HLFDM can explicitly separate low-frequency structural information and high-frequency detail information in the generated features of dust images in the frequency domain, and enhance the effective response of each using learnable modulation branches. Compared with the method of directly extracting features using ordinary convolution in the spatial domain, this module can more clearly preserve the large-scale structure of the dust region, while enhancing edge contours and local texture details. Since the modulated low-frequency and high-frequency branches are still passed to subsequent modules in a dual-branch form, it is beneficial for DF-CWAFM to further carry out cross-frequency interactive modeling, thereby improving the structural consistency, edge continuity, and detail expression ability of the generated images.

[0055] ② Dual-frequency cross-shaped window attention fusion module To further establish the complementary relationship between low-frequency structural information and high-frequency detail information, this invention sets up a dual-frequency cross-window attention fusion module DF-CWAFM after the high-low frequency feature decoupling module HLFDM. This module receives the low-frequency structural features and high-frequency detail features output by HLFDM and enhances them using different attention modeling methods. Specifically, the high-frequency branch is used to model the edge contours, texture undulations, and local detail changes of the dust region; the low-frequency branch is used to model the overall spatial distribution, main morphology, and large-scale continuous structure of the dust region. Subsequently, the two types of frequency features are adaptively integrated through a gated fusion method, thereby improving the structural consistency, edge continuity, and local texture representation ability of the generated image.

[0056] like Figure 4 As shown, given the high-frequency characteristics from HLFDM and low frequency characteristics First, 3×3 convolutions are applied to enhance local spatial representation capabilities. Then, the convolutional features are flattened and mapped to sequential patch representations, and learnable positional encodings are introduced to preserve spatial location information and enhance the model's ability to perceive long-range dependencies. For the encoded high-frequency and low-frequency sequences, layer normalization is first performed to standardize the feature distribution, and then the sequences are input into the high-frequency and low-frequency branches respectively for differentiated modeling. Specifically, the high-frequency branch inputs Cross-Window Attention (CSWA) to model the local spatial dependencies and long-range relationships of dust edges and banded texture structures; the low-frequency branch inputs Multi-Head Self-Attention (MHSA) to model the global background context. Through this branch design, key structural information of high- and low-frequency features can be extracted separately, and feature complementarity is enhanced based on encoded positional information.

[0057] For high-frequency branches, CSWA is used to jointly model local details and directional dependencies. Specifically, the input features are linearly mapped to generate queries. ,key Sum Based on the spatial structure, the features are reorganized into row-wise windows and column-wise windows, and self-attention calculations are performed in parallel in both directions. Through this bidirectional row-column cross-modeling approach, the high-frequency branch can simultaneously capture the vertical and horizontal structural extensibility of the dust edge and the banded texture structure, while also considering both local detail modeling and long-range dependency modeling with relatively low computational complexity. The attention weights in the row and column directions can be expressed as follows: Linear attention weights: (twenty four) The corresponding row direction output is: (25) Column-direction attention weights: (26) The corresponding column direction output is: (27) in: (28) in, These represent the features after recombination along the row and column directions, respectively. Let be the dimension of the key vector. This is a scaling factor used to mitigate the numerical instability caused by inner product operations on high-dimensional features. Local location coding is used to enhance the network's ability to perceive local spatial or temporal location information of the input sequence. This represents the normalization function.

[0058] Then, output the row branch. Column-direction branch output Element-wise addition is performed to fuse complementary information from two directions, resulting in the output features of the high-frequency branch after attention modeling via a cross-shaped window.

[0059] (29) For the low-frequency branch, multi-head self-attention (MHSA) is employed to model the global context and large-scale structural information. Specifically, the input low-frequency features are normalized by layers and then generated through a linear mapping. ,key Sum Furthermore, global self-attention is computed in parallel across multiple attention heads, thereby enhancing the ability of low-frequency branches to represent the overall structural distribution and macroscopic regional correlations. This can be expressed as: (30) Among them, the An attention head is defined as: (31) in, Indicates the feature embedding dimension. Indicates the number of heads of attention. Indicates matrix transpose. This is for outputting the projection matrix.

[0060] After obtaining the multi-head self-attention output, an intermediate representation is obtained by further combining linear mapping and residual connection.

[0061] (32) Subsequently, nonlinear enhancement is performed by layer normalization and multilayer perceptron, and residual connections are used to more fully preserve global background and large-scale structural information, thereby improving feature representation capabilities.

[0062] (33) After completing the differentiated modeling of the high-frequency branch and the low-frequency branch, the two features are spliced ​​together in the channel dimension to form a joint feature representation.

[0063] (34) Subsequently, the contribution of high- and low-frequency features is adaptively adjusted through a gating fusion module to achieve effective integration of global background information and local detail information. This process can be represented as follows: (35) in, This represents the learnable parameters in the gated branch. for Activation function This represents the convolution operation. This represents element-wise multiplication. Finally, the gated and fused sequence features are reconstructed back to their original spatial structure by adjusting the dimensionality, yielding the module output features: (36) Through the above process, DF-CWAFM can achieve effective interaction between low-frequency structural information and high-frequency detail information within the current generation scale. Specifically, the high-frequency branch enhances the directional coherence of dust edge contours and texture details through cross-shaped window self-attention; the low-frequency branch enhances the consistency of the main dust morphology and large-scale spatial structure through multi-head self-attention; and the gated fusion module adaptively adjusts the contributions of high- and low-frequency features based on different spatial locations and channel responses. Thus, low-frequency structural information provides global structural constraints for high-frequency texture representation, while high-frequency detail information supplements the low-frequency structural representation with edge and texture information, thereby improving the spatial structural rationality, edge continuity, and detail representation capability of the generated dust image.

[0064] ③ Cross-frequency collaborative attention enhancement module As the generator upsamples layer by layer, the feature maps in the third and higher generation stages gradually contain richer information on dust edge contours, local texture variations, and regional transitions. Although the second generation stage has achieved high- and low-frequency feature decoupling and preliminary cross-frequency interaction through the Haar wavelet decomposition-based high- and low-frequency feature decoupling module HLFDM and the dual-frequency cross-window attention fusion module DF-CWAFM, in the higher resolution generation stage, the fine-grained textures, edge structures, and spatial local responses in the features are more complex, and relying solely on the previous dual-frequency interaction is still insufficient to fully highlight key channels and key spatial regions. Therefore, this invention further introduces the cross-frequency collaborative attention enhancement module CF-AEM in the third and higher medium- and high-resolution generation stages, the structure of which is as follows: Figure 5 As shown. Specifically, CF-AEM first mines key channel responses in different frequency branches by constructing channel description vectors and grouping channels; then it enhances the complementary relationship between high-frequency texture information and low-frequency structural information through cross-frequency channel modulation; finally, it highlights the main dust region, edge region, and local texture salient region through spatial response modulation, thereby improving the ability of generated features to express the overall spatial structure, edge morphology, and local details of dust.

[0065] First, regarding the high-frequency features of the input. and low frequency characteristics Channel description vectors are constructed separately. For the high-frequency branch, to fully characterize the complementary relationship between local texture response and global channel statistics, max pooling and average pooling are used in parallel to extract channel information, resulting in two 1×1×C channel features. and The max pooling branch captures the most salient local responses in each channel, highlighting key dust textures and detailed structures. The average pooling branch characterizes the overall response trend of the channels, helping to stabilize channel weights and preserve global statistical information. Correspondingly, the low-frequency feature branch primarily provides global background information and large-scale structural information; therefore, global average pooling is used to extract its channel-level statistical features to enhance the stable representation of the overall structural distribution and suppress local noise interference. Through this differentiated channel description construction method, the high-frequency branch focuses more on local details and salient textures, while the low-frequency branch emphasizes global structure and background continuity. Accordingly, the high-frequency and low-frequency channel description vectors can be represented as: (37) (38) in, and These represent the channel description vectors for the high-frequency branch and the low-frequency branch, respectively.

[0066] To achieve adaptive enhancement at the channel dimension, the channel description vector is first processed... Normalization was performed, and a channel similarity metric was constructed based on cosine distance. Subsequently, the number of groups was adaptively determined by combining channel variance statistics. The channels are clustered using K-means based on cosine distance to obtain group labels for each channel and complete channel splitting. This allows us to obtain group labels for different semantically related channels. This process can be represented as: (39) in, This represents the K-means clustering operation based on cosine distance. and These represent the channel grouping results for the high-frequency branch and the low-frequency branch, respectively.

[0067] After obtaining the channel groups, the module independently performs grouped channel attention modeling within each group and combines multi-scale dilated convolution to expand the receptive field, thereby enhancing the joint representation ability of intra-group features on local texture and global structure. Subsequently, the outputs of each group are concatenated along the channel dimension to obtain the channel-enhanced features for the high-frequency and low-frequency branches, respectively. (40) (41) in, This represents the grouped channel attention function. This represents a concatenation operation along the channel dimension. Through this grouping modeling approach, the module can, while maintaining manageable computational overhead, more finely mine the response differences of high- and low-frequency features across different semantic channel groups.

[0068] At the same time, the obtained high-frequency features After convolution and activation function processing, information from different branches is fused through element-wise addition, and then high-frequency features are multiplied element-wise to obtain grouped channel attention enhancement features. This enhances the interactivity between local textures and the global background. Subsequently, low-frequency features are obtained. The features processed by convolution and activation functions are combined with information from different branches through element-wise addition, and then multiplied element-wise with high-frequency features to obtain grouped channel attention-enhanced features for low-frequency features. To achieve high- and low-frequency synergistic enhancement, and ultimately... and The fusion feature map is obtained by adding the high and low frequency co-enhanced features. .

[0069] (42) in, This represents an element-wise additive fusion operation. This process enables the module to simultaneously preserve high-frequency texture responses and low-frequency structural priors along the channel dimension, thereby enhancing the correlation and complementarity between cross-frequency features.

[0070] Spatial interactive attention is used to evaluate the relative importance of each spatial location in the fused feature map and generate attention weights to enhance key regions. Specifically, firstly, max pooling and average pooling are applied in parallel to the input feature map to capture local extrema and global mean responses, respectively, resulting in two spatially mapped features of size H×W×1. Subsequently, these two features are concatenated along the channel dimension, and a normalized spatial attention weight map is generated through convolution mapping and sigmoid activation. Finally, the weight map is multiplied element-wise with the original feature map to achieve adaptive enhancement of salient regions, thereby improving the model's ability to perceive key spatial information. This process can be formally represented as follows: (43) in; For the Sigmoid activation function, A convolutional layer with a kernel size of 7×7.

[0071] Furthermore, in addition to the fusion characteristics In addition to applying spatial attention, the original high-frequency features Low-frequency characteristics They also undergo the same spatial attention processing to obtain the corresponding spatial augmentation maps. and Finally, the three spatial enhancement features are summed element-wise to obtain the final feature map after spatial collaborative attention enhancement. .

[0072] in, This represents element-wise multiplication. Through the aforementioned channel-based and spatial-based synergistic enhancement processes, CF-AEM can highlight key responses in the high- and low-frequency branches in the channel dimension, and enhance the main dust region, edge region, and locally textured regions in the spatial dimension, thereby achieving synergistic enhancement of low-frequency structural information and high-frequency detail information.

[0073] ④ Diffusion-type feature refinement module To further improve the problems of insufficient local texture representation, weak edge structure continuity, and unstable response in the boundary area between sand and background during the mid-to-late generation stage, this invention sets up a diffusion-based feature refinement module (DFRM) in the mid-to-high resolution generation stage of the generator. Its structure is as follows: Figure 6 As shown, this module does not directly perform full diffusion sampling in the image space, but introduces a perturbation and restorative thinning mechanism controlled by time step conditions in the generator feature space. Through feature perturbation, residual restoration, adaptive modulation and multi-stage convolutional thinning, the current generated features are progressively corrected, thereby improving the spatial continuity, edge integrity and local texture stability of the generated image.

[0074] Specifically, firstly, for each time step The input features are positionally encoded and mapped using a multilayer perceptron to obtain a condition vector with the same dimension as the input feature space. This conditional vector controls the perturbation intensity and feature recovery method at the current stage, enabling the model to adaptively adjust the feature optimization process according to different time steps. To simulate perturbations in the feature space and enhance the model's robustness, the input features... By introducing time-step controlled Gaussian noise, a noisy feature is constructed: (44) in, Indicates the first Noisy features at each time step This represents the noise control factor associated with the time step. It is random noise that follows a standard Gaussian distribution.

[0075] Then, the noise characteristics With time step embedding vector Broadcast data is spliced ​​and merged to form the current stage of fusion input. And input it into a multi-stage convolutional feature optimization network for residual noise prediction: (45) in, This indicates a broadcast splicing operation. Represents a noise prediction network. This represents the residual noise predicted at the current stage. Further, the denoised intermediate feature representation can be obtained.

[0076] (46) Building upon noise perturbation and recovery, an adaptive feature modulation branch is further introduced to enhance the adaptability of feature representation. Specifically, after fusing conditional features with original features, learnable scaling parameters are generated through convolutional mapping. and bias parameters This allows for modulation of the input features: (47) in, This represents the Sigmoid activation function. This represents element-wise multiplication. and These represent the scale and bias parameters, respectively. Through this modulation process, the model can dynamically adjust the feature response intensity according to the current time step and input conditions, thereby enhancing the adaptability and controllability of feature representation.

[0077] Then, the denoising features With modulation characteristics The data are input into a multi-stage convolutional feature enhancement network, which sequentially performs global structure restoration, local texture refinement, and feature fusion optimization. Specifically, the first stage... Convolution, batch normalization, and Leaky ReLU activation functions are used to extract global spatial structure information, focusing on restoring the macroscopic distribution and background contours of the dust storm region; the second stage employs... Convolution, batch normalization, and Leaky ReLU activation functions are used to further enhance local texture and boundary details; the third stage optimizes the features from the first two stages. Convolutional mapping achieves joint optimization of global structural information and local detailed information, resulting in fused and enhanced features. .

[0078] Based on this, sigmoid activation is used to generate gating weights, and differential gating modulation is used to suppress the response of the input features and select features, thereby reducing redundant information and highlighting key structural regions. The process can be represented as follows: (48) in, The output features are represented. This represents the Sigmoid activation function.

[0079] Unlike traditional additive residual enhancement, the above process introduces a subtraction operation to explicitly suppress redundant activations and noise components in the input features. Specifically, this operation, guided by the fusion enhancement features, weakens background responses and interference information unrelated to the target structure, while highlighting the effective representation of key structures and detailed regions.

[0080] The second part is the coordinate-aware cross-attention discriminator CACA-D. This discriminator enhances the perception of differences between real and generated features by introducing a cross-attention mechanism, and strengthens spatial structure constraints by combining coordinate position information, thereby improving its joint discrimination ability against differences in local texture, edge morphology, and overall spatial structure. Figure 7 As shown, the discriminator simultaneously receives real samples. With generated samples Both extract shallow features through a Head module with shared parameters, obtaining two representations. and Each discriminator layer consists of three parts: a head feature extraction module, a main feature extraction module, and a tail output module. It incorporates a multi-head cross-attention mechanism to enhance global dependency modeling and local texture capture capabilities for dust regions. The head feature extraction module extracts preliminary features through 3×3 convolution, spectral normalization, Leaky-ReLU activation, and global average pooling, while generating queries and key-value pairs for attention calculation, thus establishing a global correlation between the generated image and the real image features. The main feature extraction module fuses attention-enhanced features with convolutional features and further extracts local texture, boundary transitions, and regional structure information through multiple convolutions to improve the discriminator's ability to perceive subtle differences in dust regions. The tail output module, composed of convolutional layers and global average pooling, outputs unnormalized real-valued discriminant scores to provide adversarial constraints on the generator at the structural and texture levels. This design enhances the discriminator's ability to distinguish the spatial distribution, boundary morphology, and local texture expression of dust, thereby guiding the generator to produce dust images with more reasonable structures and more complete detail representation.

[0081] In multi-head attention mechanisms, real image features capture internal long-range dependencies through self-attention, while generated image features are compared with real image features through cross-attention to identify structural differences and local artifacts. Specifically, when features When entering the multi-head self-attention, it is divided along the channel dimension into Subset: Each subset The query is generated by processing the header feature extraction module. Key Sum (Value) It also calculates self-attention to capture long-range dependencies within true features. The specific calculation is as follows: (49) in: Indicates the scaling factor. The embedding dimension of the feature. Then, By splicing together individual self-attention heads, a multi-head self-attention system is obtained. As shown below: (50) Simultaneously, generate image features Generate the corresponding key using the same Head module. Sum This is used for cross-attention calculation. Specifically, a query vector is constructed using real features. And construct keys using generated features Sum This establishes a global correspondence between real and generated features, using real features as a reference, and weights and aggregates inconsistencies in generated features to form a fusion representation for comparison and discrimination. The specific calculation is as follows: (51) Will Cross-attention head By splicing the data, we can obtain multi-head cross-attention. As shown below.

[0082] (52) Subsequently, the multi-head self-attention and multi-head cross-attention are added together to obtain the fused attention. As shown below: (53) Will merge attention After linear layer mapping, and By summing the residuals, the fusion features are obtained. As shown below: (54) Next, the fusion features will be... The input layer is normalized, and the spatial location of the fused features is reweighted using a dual-coordinate attention module, as shown below.

[0083] (55) The coordinate-aware enhanced features are fed into the main feature extraction module and the tail output module. The main feature extraction module refines the features step by step through multiple convolutional layers, improving the ability to discriminate image details. The tail output module compresses the features into an unnormalized real-valued score through convolutional mapping and global average pooling. (56) in, This represents the discriminator's score for distinguishing between the input image and the output real-valued image.

[0084] The CFCA-GAN network model employs a layer-by-layer multi-scale training strategy, which, while ensuring the spatial structure rationality and local texture detail integrity of the generated dust image regions, further enhances the morphological diversity of the generated samples. Each discriminator... The task is to determine the generated dust images Image after downsampling of real image The differences in scale provide global and local joint supervision for the generator, ensuring that the generated results remain consistent in structure and texture. Training begins with the lowest-scale GAN network, where the first-layer generator uses Gaussian noise as input to generate initial features; higher-scale generators then input the upsampled image from the previous layer along with random noise, refining the generated results layer by layer. After training at each scale, its optimized parameters are passed to the next scale as initial values, achieving cross-scale feature inheritance and progressive enhancement.

[0085] In this process, by injecting different random noises into each layer, the model can generate diverse image samples and achieve synchronous annotation by combining a label generation mechanism, providing high-quality data support for subsequent image augmentation and semantic segmentation tasks. This training strategy effectively leverages the synergistic optimization advantages of the High-Low Frequency Collaborative Feature Diffusion Generator (HLF-CFDG) and Coordinate-Aware Cross-Attention Discrimination (CACA-D), significantly improving the global structural consistency, local texture detail representation, and sample diversity of the generated images.

[0086] The third part is the joint generation mechanism of dust storm images and labels. The purpose of this part is to simultaneously generate semantic label masks corresponding to the spatial structure of dust storm remote sensing images, thereby reducing subsequent manual pixel-by-pixel annotation work and improving the usability of generated samples in downstream dust detection and semantic segmentation tasks. In remote sensing image dust detection tasks, high-quality semantic labels are a crucial foundation for ensuring the model training effect. However, existing dust labels usually rely on manual drawing, which is time-consuming. Furthermore, different annotators often have subjective differences in their judgments of dust boundaries, thin dust areas, and mixed background areas, easily leading to label inconsistencies and failing to meet the high-quality sample requirements of large-scale dust detection and segmentation tasks. Therefore, this invention designs a joint generation mechanism of images and labels, using three-channel dust storm remote sensing images and single-channel semantic label masks as unified joint samples for modeling, enabling the generator to simultaneously learn the joint distribution relationship between images and labels.

[0087] Specifically, the original dust images With corresponding tags By concatenating the input channels, a four-channel joint input is constructed: (57) in, This represents the input image of a sandstorm. This indicates the corresponding semantic tag for sand and dust. This indicates a splicing operation at the channel level. This represents the constructed image and label joint sample.

[0088] 3. Model Training After the network model was constructed, the Cross-Frequency Collaborative and Coordinate-Aware Generative Adversarial Network (CFCA-GAN) was trained. The model input data consisted of a four-channel joint sample composed of FY-4A AGRI dust storm remote sensing imagery and corresponding semantic label masks, where the first three channels were dust storm remote sensing imagery channels and the fourth channel was the dust semantic label mask channel. The training objective was to enable the generator to learn the joint distribution relationship between dust storm images and label masks under limited real dust sample conditions, and to generate dust storm remote sensing augmentation samples with reasonable structure, continuous edges, realistic texture, and synchronous labels.

[0089] Specifically, during training, the generator no longer learns the RGB image distribution separately, but instead models the four-channel joint samples, establishing a correspondence between the dust image and the labeled region in the feature space. For the nth scale, the real joint samples and the generated joint samples can be represented as follows: (58) in, and They represent the first Real-world dust images and their labels at various scales. and These represent the dust storm images output by the generator and their corresponding labels, respectively. This approach allows the model to output semantic labels consistent with the spatial structure of the dust storm images while generating them, thereby improving the usability of the generated samples in downstream segmentation tasks.

[0090] To achieve a progressive generation from coarse to fine, this invention employs a multi-scale pyramid training strategy. First, it trains the four-channel joint samples... By performing progressive downsampling, a multi-scale joint sample sequence of real data arranged from low resolution to high resolution is constructed: (59) in, Indicates the first Real-world images and labeled joint samples at various scales This represents the total number of scales. The model is trained starting from the lowest resolution scale and progressing scale by scale towards higher resolution. Each scale corresponds to a generator for that current scale. and a current scale discriminator Once training at the current scale is complete, the generator parameters, fixed noise, real sample pyramid, and noise amplitude parameters for that scale are saved and used as the basis for subsequent scale training.

[0091] For the first scale, the generator With random noise As input, generate initial four-channel image-label joint samples: (60) in, This represents the image and label joint sample generated at the first scale.

[0092] For the One scale, First, utilize the already trained... Each scale generator generates the output of the previous scale step by step and upsamples it to the current scale to obtain the prior input of the current scale. Then, the current scale random noise With prior input The noise amplitude coefficients are superimposed to form the current scale generator input: (61) in, Indicates the first Generator input at each scale, This represents the prior features generated and upsampled from the previous scale. This represents the noise amplitude coefficient at the current scale. This represents random noise at the current scale.

[0093] At each scale, the generator and discriminator Training is performed using an alternating optimization approach. First, the generator parameters are fixed, and the discriminator parameters are updated; then, the discriminator parameters are fixed again, and the generator parameters are updated. (Discriminator...) Receive real joint samples and generating joint samples It outputs unnormalized real-valued scores to determine the differences between the two in terms of spatial structure, edge morphology, local texture, and label spatial correspondence.

[0094] To ensure the rationality of the generated four-channel image and label joint samples in terms of overall distribution, spatial structure correspondence, and visible light channel response, this invention constructs a joint loss function during model training. This loss function consists of three parts: adversarial loss, reconstruction loss, and spectral ratio difference constraint loss, which can be expressed as follows: (62) In the formula, Represents resistance to loss, Represents reconstruction losses, express The weight is set to 10 by default. express The weight is set to 1.5 by default.

[0095] Combat losses: To overcome the instability issues that arise during the training of traditional GANs, WGAN-GP loss is adopted as an adversarial loss. .

[0096] (63) (64) in It is the distribution of sandstorm image generation data. This is the actual data distribution of sandstorm images; and They represent Generated and real images at different scales yes and Random interpolation between, and ; for The gradient; This is the weight of GP, which was set to 0.1 in the experiment.

[0097] Reconstruction losses: To constrain the generator to maintain consistency of image structure and detail at different scales, a reconstruction loss is introduced. This enables the generator to reconstruct the original image from a specified noise map. The calculation formula is as follows: (65) in, The original dust image was reconstructed. . The input is a fixed noise map. The first term generated by noise Large-scale dust storm images.

[0098] To enhance the spectral consistency of the generated dust storm images in the visible light channels, this invention further introduces a spectral constraint loss. This loss primarily affects the first three image channels of the generated joint sample, and does not directly affect the label channels. Under dust storm conditions, the relative response relationships among the blue, green, and red visible light channels can reflect the important spectral characteristics of the dust storm images. By constraining the normalized relative differences between the generated images and the real images in the blue and green channels and the blue and red channels, the generator can be guided to learn the channel response relationships in real dust storm scenes, thereby improving the color consistency and spectral rationality of the generated images.

[0099] Specifically, These represent the first and second parts of the generated image, respectively. The response values ​​of each pixel in the blue, green, and red channels. Let represent the response values ​​of the corresponding pixels in the blue, green, and red channels of the real dust image, respectively. Then, the spectral ratio difference constraint loss is defined as: (66) in, This is a constant used to prevent numerical instability in low-response regions. It should be noted that this spectral constraint loss does not require the generated image to be completely identical to the real image at every pixel value, but rather constrains the relative response relationship between different visible light channels. Therefore, it is more suitable for generating remote sensing images of dust storms with blurred boundaries, significant morphological variations, and irregular spatial distribution.

[0100] 4. Model Testing After training the CFCA-GAN model, the trained multi-scale generator was used to simultaneously generate and test dust storm remote sensing images and their semantic labels. To verify whether the generated results meet the application requirements for dust sample expansion and downstream segmentation tasks, this invention evaluates the model from two aspects: generated image quality and downstream task applicability. The generated image quality evaluation measures the differences between the generated dust images and real dust images in terms of visual quality, sample diversity, and consistency of statistical distribution. The downstream task applicability evaluation further verifies its actual role in improving the performance of downstream models by using the generated images and their synchronized labels as expanded samples in dust region segmentation training.

[0101] 1) Generate quality evaluation indicators Image recognition (IS) is used to evaluate the visual discernibility and sample diversity of generated images. Its core idea is that if a single generated image has a relatively clear semantic structure, then the classification network's prediction distribution for that image... The samples should be relatively concentrated; if the generated samples as a whole have good diversity, then the marginal distribution... It should have a large degree of dispersion. The formula for calculating IS is: (67) in, This represents the generated sandstorm image. This indicates the distribution of generated dust samples. This indicates that the classification network affects the generated dust images. The predicted probability of the category. This represents the marginal category distribution of all generated dust samples. This represents the KL divergence. A higher IS value indicates better performance in terms of visual representation and sample diversity in the generated image.

[0102] Fréchet Distance (FID) is used to measure the difference in distribution between real and generated dust images in the deep feature space. It first uses a pre-trained feature extraction network to extract high-dimensional features from both real and generated dust samples, and then calculates the Fréchet distance between their feature distributions, defined as: (68) in, and Let represent the mean vector and covariance matrix of the feature distribution of the real dust image, respectively. and Let represent the mean vector and covariance matrix of the generated dust image feature distribution, respectively. This represents the trace of the matrix. The smaller the FID value, the closer the generated dust image is to the real dust image in terms of statistical feature distribution, and the higher the generation quality.

[0103] KID measures the difference in distribution between real and generated samples in the feature space based on Maximum Mean Discrepancy (MMD). Unlike FID, KID does not directly rely on the Gaussian distribution assumption, but instead calculates the difference between the feature distributions of the two sets of samples through a kernel function. The formula is as follows: (69) in, and These represent the feature vectors obtained by independent sampling from the feature distribution of the real dust image; and This represents the feature vector obtained by independently sampling from the feature distribution of the generated dust image; Represents the kernel function.

[0104] 2) Segmentation performance evaluation indicators Relying solely on generation quality metrics such as IS, FID, and KID is insufficient to fully demonstrate the application value of generated dust images in practical detection tasks. Therefore, this invention further utilizes generated dust images as expanded samples in the training of the dust region segmentation model, and employs accuracy (ACC), recall, precision, F1 score, and intersection-over-union (IoU) to assess the auxiliary role of segmented samples in downstream dust region identification tasks.

[0105] Overall Accuracy (ACC) represents the proportion of all pixels that are correctly classified, and its calculation formula is: (70) in, For the model to correctly predict dusty areas as dust storms. For the model to correctly predict non-dust-free areas as non-dust-free areas, For non-dust-affected areas that the model incorrectly predicted as dust storms, The dusty area is the one that the model incorrectly predicted as not being a dusty area.

[0106] Recall represents the proportion of pixels that were actually dusty areas, and it is defined as follows: (71) Precision represents the proportion of actual dust in the detected dust area, defined as follows: (72) The F1-score considers both precision and recall, and is defined as follows: (73) In the formula, Represents accuracy. Recall rate.

[0107] IoU represents the intersection-union ratio between the predicted dust category area and the actual dust category area. It is used to measure the model's ability to identify the degree of spatial overlap between dust categories, and is defined as follows: (74) The aforementioned segmentation performance metrics can be used to further evaluate the actual effectiveness of the generated samples as supplementary data in downstream dust storm segmentation training. When the segmentation model trained based on generated samples or jointly trained with real samples achieves good results in metrics such as ACC, Recall, Precision, F1-score, and IoU, or shows improvement compared to training with only a small number of real samples, it indicates that the dust storm images and labeled samples generated by this invention have good applicability.

[0108] To demonstrate the advantages of the method of this invention, using FY-4A AGRI dust storm remote sensing imagery as an example, the method of this invention is compared with the following three methods: the traditional SinGAN method, the improved SinGAN method, and the HA-SinGAN method. The experiment selected 50 real dust storm remote sensing images and their corresponding labels as the sample base. Each real image was used as a single training sample input into the generation model, generating multiple sets of dust storm augmented samples. To intuitively demonstrate the generation effect, five typical samples were selected from the 50 real samples for visualization. Each set displays the real image, the real label, and the corresponding generated image and generated label, such as... Figure 8 As shown.

[0109] Furthermore, the generation results of the method of the present invention are compared with those of the traditional SinGAN method, the improved SinGAN method, and the HA-SinGAN method, such as... Figure 9 As shown, traditional SinGAN methods often suffer from blurred edges, discontinuous regional structures, and missing local textures when generating dusty areas. While improved SinGAN methods offer some improvement in dust contour reproduction, their representation of background texture and dust boundary details remains unstable. HA-SinGAN methods enhance local feature representation by introducing an attention mechanism, but still fall short in terms of dust edge sharpness and the distinction between dust, clouds, and the ground background. In contrast, the method of this invention better preserves the overall spatial structure, edge continuity, and local texture features of the dusty area, resulting in generated images that more closely resemble real images in terms of spatial distribution and texture representation.

[0110] To quantitatively evaluate the generation quality of different methods, three metrics—FID, KID, and IS—were used for comparison. FID and KID measure the difference in feature distribution between the generated image and the real image; lower values ​​indicate a closer resemblance. IS evaluates the sharpness and diversity of the generated image; higher values ​​indicate better generation. The evaluation results for different methods are shown in Table 1.

[0111] Table 1 Quality Evaluation Indicators of Generated Results By applying this technical solution to specific examples, the present invention can draw the following two conclusions: (1) The generation effect of the method of the present invention was verified using FY-4A AGRI dust storm remote sensing image data, with FID of 220.9140, KID of 0.2133, and IS of 2.9557. The generation results show that the method of the present invention can better preserve the spatial distribution, edge contours, and local texture features of the dust area, proving the effectiveness of the method in the task of generating dust storm remote sensing images.

[0112] (2) Compared with the SinGAN method, the improved SinGAN method, and the HA-SinGAN method, the method of the present invention performs better in all three evaluation indicators: FID, KID, and IS, indicating that the method can effectively improve the generation quality of sandstorm remote sensing images. At the same time, the method of the present invention can simultaneously generate sandstorm images and their corresponding semantic labels, reducing the cost of subsequent manual annotation and improving the expansion efficiency of labeled sandstorm samples.

[0113] Example 2 The present invention also provides a sandstorm image and tag synchronous generation system based on cross-frequency collaboration and coordinate perception. The system is used to implement the method described in Embodiment 1. The system includes: a preprocessing module, a construction module, and a generation module. The preprocessing module is used to acquire FY-4A AGRI dust storm remote sensing images and their corresponding semantic label masks, and to perform preprocessing. The module is used to build a cross-frequency collaborative and coordinate-aware generative adversarial network model based on preprocessed FY-4A AGRI dust storm remote sensing images and their corresponding semantic label masks. The generation module is used to generate sandstorm augmentation samples with structure, texture, and morphology that meet preset requirements and have corresponding semantic labels, based on cross-frequency collaborative and coordinate-aware generative adversarial network models.

[0114] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for synchronously generating sandstorm images and tags based on cross-frequency collaboration and coordinate sensing, characterized in that, The method includes: Acquire FY-4A AGRI dust storm remote sensing images and their corresponding semantic label masks, and perform preprocessing; Based on the preprocessed FY-4A AGRI dust storm remote sensing images and their corresponding semantic label masks, a cross-frequency collaborative and coordinate-aware generative adversarial network model is constructed. Based on cross-frequency collaboration and coordinate-aware generative adversarial network models, sandstorm augmentation samples are generated that meet preset requirements in terms of structure, texture, and morphology and have corresponding semantic labels.

2. The method according to claim 1, characterized in that, Methods for preprocessing FY-4A AGRI dust storm remote sensing images and their corresponding semantic label masks include: The raw HDF data of FY-4A AGRI is parsed to obtain the observation data of the required channels. Combined with the FY-4A AGRI 4km full disk latitude and longitude lookup table FullMask_Grid_4000.raw, the latitude and longitude matrices corresponding to the row and column positions of the original observation grid are obtained. The correspondence between the row and column coordinates of the original scanned pixels and the geographic latitude and longitude coordinates is established. The observation data of each channel are uniformly mapped to the target regular latitude and longitude grid, thereby completing the geometric positioning and spatial registration of FY-4A AGRI image. The raw digital quantization values ​​were radiometrically calibrated based on the calibration parameters of the FY-4A AGRI product. For invalid pixels in the raw observation data, comprehensive identification and processing are performed based on the FY-4A AGRI product data structure and calibration results. Specifically, the raw digital quantization values ​​65534 and 65535 in NOMChannel are identified as invalid product codes, and their corresponding pixels are set to null values ​​during radiometric calibration. At the same time, invalid items, non-finite values, and outliers appearing after calibration are removed from the calibration lookup table. Furthermore, an effective Earth observation mask is constructed by combining the effective observation column range of each scan row recorded in NOMObsColumn to remove pixels outside the Earth and invalid scan areas. After completing radiometric calibration, geometric correction and invalid pixel processing, the FY-4A AGRI image was cropped based on the dust storm occurrence area and the study area to obtain a dust storm remote sensing image that includes the main dust storm area and the surrounding background. The cropped image is resampled to convert samples from different times and regions into regular grid samples of the same size. At the same time, the sample size is uniformly set to 256×256 pixels. NOMChannel01, NOMChannel02, and NOMChannel03 were selected for pseudo-RGB synthesis. Specifically, NOMChannel03 was used as the R channel, NOMChannel02 as the G channel, and NOMChannel01 as the B channel to construct a three-channel sandstorm remote sensing image. For semantic labeling masks, pixel-level annotations are performed based on the distribution of dust areas in remote sensing images of dust storms. Specifically, the dust storm occurrence area is determined by visual interpretation or existing annotation results, the dust area is marked as the target category, and the non-dust area is marked as the background category, thus forming a single-channel semantic labeling mask.

3. The method according to claim 1, characterized in that, The cross-frequency collaborative and coordinate-aware generative adversarial network model includes: a high- and low-frequency collaborative feature diffusion generator, a coordinate-aware cross-attention discriminator, and a joint image and label generation mechanism; A high- and low-frequency collaborative feature diffusion generator is used to learn the large-scale spatial structure, edge morphology, and local texture features of dust regions from joint samples of imagery and label four channels; A coordinate-aware cross-attention discriminator is used to discriminate and constrain generated samples based on spatial positional relationships, region boundary morphology, and local texture representation. An image and label joint generation mechanism is used to maintain spatial consistency between the generated dust image and the corresponding semantic label mask.

4. The method according to claim 3, characterized in that, In the low-resolution generation stage, the high- and low-frequency collaborative feature diffusion generator adopts a multi-scale pyramid generation structure. Each generation unit includes three basic parts: a head feature extraction module, a main feature extraction module, and a tail output module. The head feature extraction module is used to receive input features at the current scale and complete the initial feature mapping through convolutional layers, normalization layers and Leaky-ReLU activation function. The main feature extraction module consists of multi-layer convolutional blocks or feature enhancement modules, used to extract and enhance the spatial structure, edge contours and texture features of the dust region; The tail output module consists of a convolutional layer and a Tanh activation function, which is used to map the generated features into a four-channel joint output, which includes a three-channel dust image and a one-channel dust label. In the medium-to-high resolution generation stage, the high-to-low frequency collaborative feature diffusion generator sequentially embeds the high-to-low frequency feature decoupling module HLFDM based on Haar wavelet decomposition, the dual-frequency cross-shaped window attention fusion module DF-CWAFM, the cross-frequency collaborative attention enhancement module CF-AEM, and the diffusion feature refinement module DFRM into the basic generation unit. Among them, the high- and low-frequency feature decoupling module HLFDM based on Haar wavelet decomposition is used to decompose the current scale input features into low-frequency structural components and high-frequency detail components. The dual-frequency cross-shaped window attention fusion module DF-CWAFM is used to establish a complementary relationship between low-frequency structural information and high-frequency texture information; The Cross-Frequency Collaborative Attention Enhancement (CF-AEM) module is used to collaboratively enhance cross-frequency fusion features from both channel and spatial dimensions. The Diffusion Feature Refinement (DFRM) module is used to perform perturbation recovery and detail refinement on the fused generated features, suppress redundant background responses, and enhance the continuity of dust edges and local texture stability.

5. The method according to claim 3, characterized in that, The coordinate-aware cross-attention discriminator enhances the perception of differences between real and generated features by introducing a cross-attention mechanism, and strengthens spatial structure constraints by combining coordinate position information; specifically: The discriminator receives both real and generated samples. The real and generated samples extract shallow features through the Head module with shared parameters to obtain two representations. Each discriminator consists of three parts: a head feature extraction module, a main feature extraction module, and a tail output module. It also incorporates a multi-head cross-attention mechanism to enhance the ability to model global dependencies and capture local textures in the dusty area. The head feature extraction module extracts preliminary features through 3×3 convolution, spectral normalization, Leaky-ReLU activation and global average pooling, while generating queries and key-value pairs for attention calculation, thereby establishing a global association between the features of the generated image and the real image. The main feature extraction module integrates attention-enhanced features with convolutional features, and further extracts local texture, boundary transition and regional structure information through multi-layer convolution to improve the discriminator's ability to perceive the subtle differences in sand and dust areas. The tail output module consists of convolutional layers and global average pooling, outputting unnormalized real-valued discrimination scores to provide adversarial constraints at the structural and texture levels for the generator.

6. The method according to claim 3, characterized in that, The image and label joint generation mechanism models three-channel dust storm remote sensing images and single-channel semantic label masks as unified joint samples, enabling the generator to simultaneously learn the joint distribution relationship between images and labels; specifically: Original sandstorm images With corresponding tags By concatenating the input channels, a four-channel joint input is constructed: ； in, This indicates the input image of a sandstorm. This indicates the corresponding semantic tag for sand and dust. This indicates a splicing operation at the channel level. This represents the constructed image-label joint sample. and These represent the height and width of the feature map, respectively.

7. A system for synchronously generating sandstorm images and tags based on cross-frequency collaboration and coordinate sensing, the system being used to implement the method described in any one of claims 1-6, characterized in that, The system includes: a preprocessing module, a construction module, and a generation module; The preprocessing module is used to acquire FY-4A AGRI dust storm remote sensing images and their corresponding semantic label masks, and to perform preprocessing. The module is used to build a cross-frequency collaborative and coordinate-aware generative adversarial network model based on preprocessed FY-4A AGRI dust storm remote sensing images and their corresponding semantic label masks. The generation module is used to generate sandstorm augmentation samples with structure, texture, and morphology that meet preset requirements and have corresponding semantic labels, based on cross-frequency collaborative and coordinate-aware generative adversarial network models.

Images