A remote sensing image phased radiation correction method

Abstract

This invention discloses a staged radiometric correction method for remote sensing images, comprising: acquiring the target remote sensing image to be corrected and its corresponding metadata; preprocessing the target remote sensing image based on the metadata to generate multiple image block units containing overlapping regions; sequentially inputting the multiple image block units into a pre-trained lightweight segmentation network to obtain the problem type, land cover type, and season type corresponding to each image block unit, wherein the problem type includes at least one of cloud shadow, fog, and aerosol; and performing staged repair on the multiple image block units according to the problem type based on a preset multi-scene anomaly sample library and a preset repair strategy to obtain the corrected target remote sensing image. The staged radiometric correction method for remote sensing images of this application solves the problems of coupled detection and processing, lack of adaptation to land cover and season, and lack of decomposability in multi-anomaly processing in existing technologies, significantly improving the accuracy and efficiency of remote sensing image radiometric correction.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a method for staged radiometric correction of remote sensing images. Background Technology

[0002] In the process of removing atmospheric and lighting anomalies from remote sensing images, traditional global radiometric correction methods have many shortcomings in large-scale industrial applications, particularly regarding typical interference problems such as cloud shadows, fog, and aerosols.

[0003] The lack of separation between detection and processing leads to misjudgment and misrepair: Existing methods usually rely on simple thresholds or single spectral features to make a rough judgment on the entire image. The detection and repair processes are tightly coupled, lacking precise abnormal area masks and multi-dimensional labels, which easily leads to missed detections and false detections. This results in the over-processing of non-abnormal areas and damages the spectral consistency of normal areas.

[0004] The lack of adaptive grouping for ground features and seasons leads to unstable restoration results: Traditional methods often use a unified physical model or a single network parameter, which cannot fully consider the spectral differences of different ground features (farmland, cities, water bodies, etc.) and seasonal observation conditions. This can easily lead to inconsistent local restoration results, making it difficult for the algorithm to be automated for large-scale batch processing and hindering its industrial application.

[0005] The handling of multiple anomalies lacks decomposability and is not targeted enough: multiple anomalies such as cloud shadows, fog, and aerosols are often mixed in a single atmospheric model and repaired at once. They are not processed in stages according to the degree of physical reversibility and quantifiability. Different anomalies are prone to superimposed interference, resulting in poor repair accuracy and physical consistency, which makes it difficult to meet the requirements of subsequent high-precision physical quantity inversion.

[0006] Therefore, a new method for radiometric correction of remote sensing images is needed to solve the above problems. Summary of the Invention

[0007] To address the technical problems existing in the background art, this invention proposes a staged radiometric correction method for remote sensing images.

[0008] The present invention proposes a staged radiometric correction method for remote sensing images, comprising:

[0009] The target remote sensing image to be corrected and its corresponding metadata are acquired, and the target remote sensing image is preprocessed based on the metadata to generate multiple image block units with overlapping areas.

[0010] Multiple image patch units are sequentially input into a pre-trained lightweight segmentation network to obtain the problem type, land cover type, and season type corresponding to the image patch unit. The problem type includes at least one of cloud shadow, fog, and aerosol.

[0011] Based on a pre-set multi-scene anomaly sample library and a pre-set repair strategy, multiple image block units are repaired in stages according to the problem type to obtain the corrected target remote sensing image.

[0012] Preferably, the construction process of the multi-scenario anomaly sample library includes:

[0013] Collect remote sensing image data and corresponding metadata from multiple time periods and regions;

[0014] The collected remote sensing image data and corresponding metadata are classified according to land cover type and seasonal type to generate basic classification subsets;

[0015] For each basic classification subset, sample blocks containing cloud shadows, fog, and aerosol anomalies, as well as corresponding normal sample blocks, are collected to form corresponding sample pairs;

[0016] All sample blocks are uniformly cropped to 512×512 pixels, labeled with problem type tags, and stored as a sample library. The sample pairs include: cloud shadow sample pairs, fog sample pairs, and aerosol sample pairs. The cloud shadow sample pairs contain cloud shadow image blocks and no-shadow reference image blocks in the same area. The fog sample pairs contain fog-covered image blocks and no-fog reference image blocks in the same area. The aerosol sample pairs contain high-aerosol image blocks and low-aerosol reference image blocks in the same area.

[0017] Preferably, the land cover types include farmland, cities, water bodies, bare soil, and snowfields; the season types include spring, summer, autumn, and winter; and the preset restoration strategy specifically includes: for multiple image block units, first restore image block units of cloud shadow type, then restore image block units of fog type, and finally restore image block units of aerosol type.

[0018] Preferably, the restoration process for the cloud shadow type image block unit specifically includes:

[0019] Match a subset of cloud shadow sample pairs from a multi-scene anomaly sample library that are consistent with the current image block unit's land cover type and season type;

[0020] A pre-trained conditional generative adversarial network is loaded as the inpainting model. The reflectance of the shadowed area is inverted using the spectral features of the unshaded areas of the pixels in the current image block and adjacent blocks and multi-temporal reference data of the same area during the same period in the historical image database as input.

[0021] For multiple restoration results of the same pixel p in the overlapping area of ​​adjacent blocks, the weight is calculated by the cloud shadow restoration fusion formula and pixel-level weighted fusion is performed to obtain a restored image block unit with consistent reflectance of the shadow area and spectrum of the surrounding unshadowed area and multi-temporal reference data, and smooth boundary.

[0022] Preferably, the restoration process for the hazy image block unit specifically includes:

[0023] Input an image block unit restored by cloud shadows, and match a subset of fog sample pairs from a multi-scene anomaly sample library that are consistent with the land cover type and season type of the current image block unit restored by cloud shadows.

[0024] Based on the estimation of local fog transmittance T from a subset of fog sample pairs, a pre-trained conditional generative adversarial network is invoked to perform intra-block fog removal with transmittance T and spectral features of fog-free areas in the current cloud shadow repaired image block unit as input.

[0025] For multiple restoration results of the same pixel p in the overlapping area of ​​adjacent blocks, the weights are calculated using the haze restoration fusion formula and then pixel-level weighted fusion is performed to obtain a restored image block unit that has thoroughly removed haze, is spectrally continuous with the surrounding area, and has smooth boundaries.

[0026] Preferably, the aerosol-type image block unit for restoration includes:

[0027] Input an image block unit that has been restored by haze, and match a subset of aerosol sample pairs from a multi-scene anomaly sample library that are consistent with the land cover type and seasonal type of the current image block unit restored by haze.

[0028] By combining atmospheric parameters in the metadata with external aerosol optical thickness data, the local AOD value is estimated. A pre-trained aerosol residual inpainting model is called to perform aerosol correction with the local AOD value and the spectral characteristics of the low aerosol region in the current block as input. For multiple inpainting results of the same pixel p in the overlapping area of ​​adjacent blocks, the weight is calculated by the aerosol inpainting fusion formula and pixel-level weighted fusion is performed to obtain a well-inpainted image block unit with thorough aerosol correction, spectral characteristics consistent with the low aerosol reference data in the same area, and smooth boundaries.

[0029] Preferably, the metadata includes shooting time, land feature type, solar altitude angle, radiance, and atlantic top reflectivity; the preprocessing specifically includes:

[0030] Extract the solar altitude angle from the metadata and calculate the solar zenith angle, where the solar zenith angle = 90° - solar altitude angle;

[0031] Convert the radiance values ​​of the target remote sensing image into the top atmospheric reflectance;

[0032] The target remote sensing image is normalized based on the solar zenith angle and the atmospheric top reflectance corresponding to all pixels of the target remote sensing image to obtain the normalized target remote sensing image.

[0033] A sliding window is used to divide the normalized target remote sensing image into blocks to generate multiple image block units with overlapping areas. The sliding window has a size of 512×512 pixels and a sliding step size of 256 pixels, so that adjacent image block units form a 50% overlap area.

[0034] Preferably, the normalization process specifically includes:

[0035] The atmospheric top reflectance of all pixels in the target remote sensing image is divided by the cosine of the solar zenith angle to normalize to the vertical incidence illumination condition. The final output is a normalized image that retains the original spatial resolution and georeference, i.e., the normalized target remote sensing image.

[0036] Preferably, the lightweight segmentation network is an improved U-Net network, which uses depthwise separable convolutions instead of standard convolutions to reduce the number of parameters; an attention mechanism module is set in the encoding stage to enhance the extraction of abnormal region features; the training samples of the lightweight segmentation network are derived from a preset multi-scene abnormal sample library.

[0037] Preferably, the cloud shadow repair and blending formula is as follows:

[0038] ;

[0039] in, The weights for the cloud shadow at pixel p of the i-th block are adjusted. The confidence level for shadow inpainting of the i-th image block unit; Let p be the Euclidean distance from pixel p to the center of the i-th image block; This is the cloud shadow fusion attenuation constant, used to control the rate at which the weights decay with distance.

[0040] Preferably, the haze repair and fusion formula is as follows:

[0041] ;

[0042] in, The fog repair weight for the i-th block at pixel p; To restore confidence in haze removal; Let p be the Euclidean distance from pixel p to the center of the i-th image block; is the haze fusion attenuation constant.

[0043] Preferably, the aerosol remediation and fusion formula is as follows:

[0044] ;

[0045] in, The aerosol repair weight for the i-th image block unit at pixel p; To correct the confidence level for aerosols, the reliability of the aerosol residual remediation model in the current block is evaluated; Input the confidence level for AOD; The aerosol fusion attenuation constant; Let p be the Euclidean distance from pixel p to the center of the i-th image block.

[0046] This invention proposes a phased radiometric correction method for remote sensing images. By constructing a multi-scene sample library and classifying it by land cover type and season, it provides a high-quality data foundation for restoration. Solar elevation angle normalization reduces the impact of non-abnormal illumination differences. Accurate localization of anomalous areas is achieved through sliding window block detection and multi-dimensional label generation. Phased restoration of anomalous issues based on physical reversibility avoids interference between different anomalous areas. Fusion processing of overlapping areas between blocks ensures smooth boundaries and physical consistency of the restoration results, while quality assessment and closed-loop optimization continuously improve the restoration effect. Compared with traditional methods, this invention separates detection and restoration, avoiding misprocessing of normal areas and saving computational resources. Adaptation to land cover and season improves spectral consistency in restoration. Phased processing of multiple anomalous areas reduces local residuals. Block fusion improves spatial continuity. Continuous optimization ensures long-term stability. Attached Figure Description

[0047] Figure 1 This is a schematic diagram illustrating the workflow of a staged radiometric correction method for remote sensing images proposed in this invention.

[0048] Figure 2 This is a schematic diagram of a problem image illustrating one implementation of a phased radiometric correction method for remote sensing images proposed in this invention. Detailed Implementation

[0049] Reference Figure 1 and Figure 2 The present invention proposes a staged radiometric correction method for remote sensing images, comprising the following steps:

[0050] S1. Obtain the target remote sensing image to be corrected and the corresponding metadata of the target remote sensing image, and preprocess the target remote sensing image based on the metadata to generate multiple image block units with overlapping areas.

[0051] In this embodiment, the metadata includes the shooting time, land cover type, solar altitude angle, radiance, and atmospheric top reflectivity; the preprocessing specifically includes:

[0052] Extract the solar altitude angle from the metadata and calculate the solar zenith angle, where solar zenith angle = 90° - solar altitude angle;

[0053] Convert the radiance values ​​of the target remote sensing image into the top atmospheric reflectance;

[0054] The target remote sensing image is normalized based on the solar zenith angle and the atmospheric top reflectance corresponding to all pixels of the target remote sensing image to obtain the normalized target remote sensing image.

[0055] A sliding window is used to divide the normalized target remote sensing image into blocks to generate multiple image block units with overlapping areas. The sliding window size is 512×512 pixels and the sliding step size is 256 pixels, so that adjacent image block units form a 50% overlap area.

[0056] Specifically, the normalization process is as follows:

[0057] The atmospheric top reflectance of all pixels in the target remote sensing image is divided by the cosine of the solar zenith angle to normalize to the vertical incidence illumination condition. The final output is a normalized image that retains the original spatial resolution and georeference, i.e., the normalized target remote sensing image.

[0058] Specifically, for L1-level remote sensing image products acquired by low-cost satellite constellations such as Jilin-1, before performing subsequent problem area detection, anomaly identification, and phased repair, it is necessary to first normalize the original radiation values ​​or atmospheric top reflectivity based on the solar elevation angle information in the metadata. This is to eliminate non-abnormal illumination differences caused by different observation times and changes in solar elevation angle, ensuring that subsequent anomaly detection and repair operations only target real atmospheric or shadow problems.

[0059] Specifically, the calculation process for atmospheric top reflectivity is as follows: ;

[0060] in, Pixel radiance; This is the normalization coefficient for the Earth-Sun distance; This represents the solar constant for the Sun in this wavelength band; This is the solar zenith angle.

[0061] S2. Input multiple image block units sequentially into a pre-trained lightweight segmentation network to obtain the problem type, land cover type, and season type corresponding to the image block unit. The problem type includes at least one of cloud shadow, fog, and aerosol.

[0062] In this embodiment, the land cover types include farmland, city, water body, bare soil, and snowfield; the season types include spring, summer, autumn, and winter; the preset restoration strategy specifically includes: for multiple image block units, first restore the image block units of cloud shadow type, then restore the image block units of fog type, and finally restore the image block units of aerosol type.

[0063] Specifically, the lightweight segmentation network is an improved U-Net network that uses depthwise separable convolutions instead of standard convolutions to reduce the number of parameters; an attention mechanism module is set in the encoding stage to enhance the extraction of abnormal region features; and the training samples of the lightweight segmentation network are derived from a pre-set multi-scene abnormal sample library.

[0064] S3. Based on the preset multi-scene anomaly sample library and preset repair strategy, multiple image block units are repaired in stages according to the problem type to obtain the corrected target remote sensing image.

[0065] In this embodiment, the process of constructing a multi-scenario anomaly sample library includes:

[0066] Collect remote sensing image data and corresponding metadata from multiple time periods and regions;

[0067] The collected remote sensing image data and corresponding metadata are classified according to land cover type and seasonal type to generate basic classification subsets;

[0068] For each basic classification subset, sample blocks containing cloud shadows, fog, and aerosol anomalies, as well as corresponding normal sample blocks, are collected to form corresponding sample pairs;

[0069] All sample blocks were uniformly cropped to 512×512 pixels, labeled with problem type tags, and stored as a sample library. Sample pairs included: cloud shadow sample pairs, fog sample pairs, and aerosol sample pairs. Cloud shadow sample pairs contained cloud shadow image blocks and no-shadow reference image blocks in the same area; fog sample pairs contained fog-covered image blocks and no-fog reference image blocks in the same area; aerosol sample pairs contained high-aerosol image blocks and low-aerosol reference image blocks in the same area.

[0070] Specifically, all sample blocks are uniformly made into 512×512 pixel blocks. Combined with the sub-meter resolution of image satellites, it can cover a ground area of ​​about 500m×500m, which not only preserves sufficient spatial context, but also meets the input size requirements of deep learning models, facilitating large-scale automated sliding window processing.

[0071] In this embodiment, the method of this application is used to process three types of image block units—cloud shadows, fog, and aerosols—in stages and independently, and the final target remote sensing image is obtained after merging the processes.

[0072] In this embodiment, the restoration process for image block units of the cloud shadow type specifically includes:

[0073] Match a subset of cloud shadow sample pairs from a multi-scene anomaly sample library that are consistent with the current image block unit's land cover type and season type;

[0074] A pre-trained conditional generative adversarial network is loaded as the inpainting model. The reflectance of the shadowed area is inverted using the spectral features of the unshaded areas of the pixels in the current image block and adjacent blocks and multi-temporal reference data of the same area during the same period in the historical image database as input.

[0075] For multiple restoration results of the same pixel p in the overlapping area of ​​adjacent blocks, the weight is calculated by the cloud shadow restoration fusion formula and pixel-level weighted fusion is performed to obtain a restored image block unit with consistent reflectance of the shadow area and spectrum of the surrounding unshadowed area and multi-temporal reference data, and smooth boundary.

[0076] Specifically, the cloud shadow repair and blending formula is as follows:

[0077] ;

[0078] in, The weights for the cloud shadow at pixel p of the i-th block are adjusted. The confidence level for shadow inpainting of the i-th image block unit; Let p be the Euclidean distance from pixel p to the center of the i-th image block; This is the cloud shadow fusion attenuation constant, used to control the rate at which the weights decay with distance.

[0079] Specifically, the formula for calculating the final merged pixel values ​​after cloud shadow restoration is as follows:

[0080] ;

[0081] in, These are the pixel values ​​after shadow inversion; This represents the final pixel value after cloud shadow restoration. This weighted fusion ensures that the restoration results for the cloud shadow area maintain spectral continuity at the boundary positions, avoiding seams between blocks.

[0082] In this embodiment, the restoration process for hazy image block units specifically includes:

[0083] Input an image block unit restored by cloud shadows, and match a subset of fog sample pairs from a multi-scene anomaly sample library that are consistent with the land cover type and season type of the current image block unit restored by cloud shadows.

[0084] Based on the estimation of local fog transmittance T from a subset of fog sample pairs, a pre-trained conditional generative adversarial network is invoked to perform intra-block fog removal with transmittance T and spectral features of fog-free areas in the current cloud shadow repaired image block unit as input.

[0085] For multiple restoration results of the same pixel p in the overlapping area of ​​adjacent blocks, the weights are calculated using the haze restoration fusion formula and then pixel-level weighted fusion is performed to obtain a restored image block unit that has thoroughly removed haze, is spectrally continuous with the surrounding area, and has smooth boundaries.

[0086] Specifically, the haze repair and fusion formula is as follows:

[0087] ;

[0088] in, The fog repair weight for the i-th block at pixel p; To restore confidence in haze removal; Let p be the Euclidean distance from pixel p to the center of the i-th image block; is the haze fusion attenuation constant.

[0089] Specifically, the formula for calculating the final fused pixel value in the haze restoration stage is as follows:

[0090] ;

[0091] in, These are the pixel values ​​after dehazing; The final pixel value is obtained after further dehazing based on the restored cloud shadows. This fusion process ensures that the image after haze removal is spatially continuous and consistent, while avoiding visible transition bands caused by differences in dehazing intensity between different blocks.

[0092] In this embodiment, the aerosol-type image block repair unit includes:

[0093] Input an image block unit that has been restored by haze, and match a subset of aerosol sample pairs from a multi-scene anomaly sample library that are consistent with the land cover type and seasonal type of the current image block unit restored by haze.

[0094] By combining atmospheric parameters in the metadata with external aerosol optical thickness data, the local AOD value is estimated. A pre-trained aerosol residual inpainting model is called to perform aerosol correction with the local AOD value and the spectral characteristics of the low aerosol region in the current block as input. For multiple inpainting results of the same pixel p in the overlapping area of ​​adjacent blocks, the weight is calculated by the aerosol inpainting fusion formula and pixel-level weighted fusion is performed to obtain a well-inpainted image block unit with thorough aerosol correction, spectral characteristics consistent with the low aerosol reference data in the same area, and smooth boundaries.

[0095] Specifically, the aerosol remediation and fusion formula is as follows:

[0096] ;

[0097] in, The aerosol repair weight for the i-th image block unit at pixel p; To correct the confidence level for aerosols, the reliability of the aerosol residual remediation model in the current block is evaluated; Input the confidence level for AOD; The aerosol fusion attenuation constant; Let p be the Euclidean distance from pixel p to the center of the i-th image block.

[0098] Specifically, the final fused pixel values ​​in the aerosol repair stage The calculation formula is:

[0099] ;

[0100] in, These are pixel values ​​after aerosol correction; The final pixel value is obtained after aerosol correction following the restoration of cloud shadows and haze. This fusion method can fully utilize the correction reliability of each block and the credibility of physical parameters, making the final output aerosol correction result more stable in terms of spectral consistency and spatial continuity.

[0101] It should be noted that during the three sequential processing stages, the relevant methods only apply independently to the segmented cloud shadow image patch, haze image patch, and aerosol image patch. The final image result after the entire process is completed is as follows:

[0102] ;

[0103] Wherein, R(p) is the final pixel value of the pixel after unification, smoothing, and fusion in this anomaly repair stage.

[0104] In this embodiment, atmospheric parameters include atmospheric models, aerosol models, water vapor content, and ozone content. The estimation of local AOD values ​​employs a hybrid estimation method combining physical models and data-driven approaches, including:

[0105] Input initialization: Combine the atmospheric and aerosol models in the external low-resolution AOD product and metadata to generate the initial AOD field;

[0106] Local inversion: Identify dark targets in the image and estimate their surface reflectance in the visible light band using empirical relationships based on the reflectance of the dark targets in the shortwave infrared band.

[0107] Iterative solution: Substitute the estimated surface reflectance into the radiative transfer equation and find the optimal local AOD value through an iterative algorithm;

[0108] Fusion Output: The high-resolution local AOD results obtained from physical inversion are fused with the external AOD initial field to generate a high spatial resolution local AOD map.

[0109] Among them, the Atmospheric Model describes the classification of standard atmospheric temperature and humidity profiles (such as mid-latitude summer, mid-latitude winter, and tropical), and is used to determine the characteristics of atmospheric molecular (Rayleigh) scattering.

[0110] Aerosol Model: Defines aerosol type and particle size distribution (e.g., continental, urban, marine) to determine aerosol scattering and absorption characteristics. Water Vapour Content: Total precipitable water (TPW), affecting near-infrared and short-wave infrared bands. Ozone Content: The total atmospheric column content, affecting ultraviolet and part of the visible light band.

[0111] Specifically, after cloud shadows and haze removal, local residual aerosol block units are subjected to local optical thickness estimation and restoration by combining external AOD products or multi-temporal reference images, thereby improving the accuracy and traceability of atmospheric residual influence removal. This invention utilizes the high repetition rate of small satellite constellations and the ability to rapidly acquire multiple images from the same period, facilitating the construction and continuous updating of multi-scene sample libraries, supporting a phased restoration network and a closed-loop quality verification system, and meeting the requirements of large-scale automated industrial processing.

[0112] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for staged radiometric correction of remote sensing images, characterized in that, include: The target remote sensing image to be corrected and its corresponding metadata are acquired, and the target remote sensing image is preprocessed based on the metadata to generate multiple image block units with overlapping areas. Multiple image patch units are sequentially input into a pre-trained lightweight segmentation network to obtain the problem type, land cover type, and season type corresponding to the image patch unit. The problem type includes at least one of cloud shadow, fog, and aerosol. Based on a pre-set multi-scene anomaly sample library and a pre-set repair strategy, multiple image block units are repaired in stages according to the problem type to obtain the corrected target remote sensing image. The land cover types include farmland, cities, water bodies, bare soil, and snowfields; the season types include spring, summer, autumn, and winter; the preset restoration strategy specifically includes: for multiple image block units, first restore image block units of cloud shadow type, then restore image block units of fog type, and finally restore image block units of aerosol type. The restoration process for the cloud shadow type image block unit specifically includes: Match a subset of cloud shadow sample pairs from a multi-scene anomaly sample library that are consistent with the current image block unit's land cover type and season type; A pre-trained conditional generative adversarial network is loaded as the inpainting model. The reflectance of the shadowed area is inverted using the spectral features of the unshaded areas of the pixels in the current image block and adjacent blocks and multi-temporal reference data of the same area during the same period in the historical image database as input. For multiple restoration results of the same pixel p in the overlapping area of ​​adjacent blocks, the weight is calculated by the cloud shadow restoration fusion formula and pixel-level weighted fusion is performed to obtain a restored image block unit with consistent reflectance of the shadow area and spectrum of the surrounding unshadowed area and multi-temporal reference data, and smooth boundary.

2. The method for staged radiometric correction of remote sensing images according to claim 1, characterized in that, The construction process of the multi-scenario anomaly sample library includes: Collect remote sensing image data and corresponding metadata from multiple time periods and regions; The collected remote sensing image data and corresponding metadata are classified according to land cover type and seasonal type to generate basic classification subsets; For each basic classification subset, sample blocks containing cloud shadows, fog, and aerosol anomalies, as well as corresponding normal sample blocks, are collected to form corresponding sample pairs; All sample blocks are uniformly cropped to 512×512 pixels, labeled with problem type tags, and stored as a sample library. The sample pairs include: cloud shadow sample pairs, fog sample pairs, and aerosol sample pairs. The cloud shadow sample pairs contain cloud shadow image blocks and no-shadow reference image blocks in the same area. The fog sample pairs contain fog-covered image blocks and no-fog reference image blocks in the same area. The aerosol sample pairs contain high-aerosol image blocks and low-aerosol reference image blocks in the same area.

3. The method for staged radiometric correction of remote sensing images according to claim 1, characterized in that, The restoration process for the hazy image block unit specifically includes: Input an image block unit restored by cloud shadows, and match a subset of fog sample pairs from a multi-scene anomaly sample library that are consistent with the land cover type and season type of the current image block unit restored by cloud shadows. Based on the estimation of local fog transmittance T from a subset of fog sample pairs, a pre-trained conditional generative adversarial network is invoked to perform intra-block fog removal with transmittance T and spectral features of fog-free areas in the current cloud shadow repaired image block unit as input. For multiple restoration results of the same pixel p in the overlapping area of ​​adjacent blocks, the weights are calculated using the haze restoration fusion formula and then pixel-level weighted fusion is performed to obtain a restored image block unit that has thoroughly removed haze, is spectrally continuous with the surrounding area, and has smooth boundaries.

4. The method for staged radiometric correction of remote sensing images according to claim 1, characterized in that, The image block unit for restoring aerosol types includes: Input an image block unit that has been restored by haze, and match a subset of aerosol sample pairs from a multi-scene anomaly sample library that are consistent with the land cover type and seasonal type of the current image block unit restored by haze. By combining atmospheric parameters in the metadata with external aerosol optical thickness data, the local AOD value is estimated. A pre-trained aerosol residual inpainting model is called to perform aerosol correction with the local AOD value and the spectral characteristics of the low aerosol region in the current block as input. For multiple inpainting results of the same pixel p in the overlapping area of ​​adjacent blocks, the weight is calculated by the aerosol inpainting fusion formula and pixel-level weighted fusion is performed to obtain a well-inpainted image block unit with thorough aerosol correction, spectral characteristics consistent with the low aerosol reference data in the same area, and smooth boundaries.

5. The method for staged radiometric correction of remote sensing images according to claim 1, characterized in that, The metadata includes the shooting time, land cover type, solar altitude angle, radiance, and atmospheric reflectivity; The preprocessing specifically includes: Extract the solar altitude angle from the metadata and calculate the solar zenith angle, where the solar zenith angle = 90° - solar altitude angle; Convert the radiance values ​​of the target remote sensing image into the top atmospheric reflectance; The target remote sensing image is normalized based on the solar zenith angle and the atmospheric top reflectance corresponding to all pixels of the target remote sensing image to obtain the normalized target remote sensing image. A sliding window is used to divide the normalized target remote sensing image into blocks to generate multiple image block units with overlapping areas. The sliding window has a size of 512×512 pixels and a sliding step size of 256 pixels, so that adjacent image block units form a 50% overlap area.

6. The method for staged radiometric correction of remote sensing images according to claim 5, characterized in that, The normalization process specifically involves: The atmospheric top reflectance of all pixels in the target remote sensing image is divided by the cosine of the solar zenith angle to normalize to the vertical incidence illumination condition. The final output is a normalized image that retains the original spatial resolution and georeference, i.e., the normalized target remote sensing image.

7. The method for staged radiometric correction of remote sensing images according to claim 1, characterized in that, The lightweight segmentation network is an improved U-Net network, which uses depthwise separable convolutions instead of standard convolutions to reduce the number of parameters; an attention mechanism module is set in the encoding stage to enhance the extraction of abnormal region features; the training samples of the lightweight segmentation network are derived from a pre-set multi-scene abnormal sample library.

8. The method for staged radiometric correction of remote sensing images according to claim 1, characterized in that, The cloud shadow repair and fusion formula is as follows: ； in, The weights for the cloud shadow at pixel p of the i-th block are adjusted. The confidence level for shadow inpainting of the i-th image block unit; Let p be the Euclidean distance from pixel p to the center of the i-th image block; This is the cloud shadow fusion attenuation constant, used to control the rate at which the weights decay with distance.

9. The method for staged radiometric correction of remote sensing images according to claim 3, characterized in that, The specific formula for haze repair and fusion is as follows: ； in, The fog repair weight for the i-th block at pixel p; To restore confidence in haze removal; Let p be the Euclidean distance from pixel p to the center of the i-th image block; is the haze fusion attenuation constant.

10. The method for staged radiometric correction of remote sensing images according to claim 4, characterized in that, The specific formula for aerosol remediation and fusion is as follows: ； in, The aerosol repair weight for the i-th image block unit at pixel p; To correct the confidence level for aerosols, the reliability of the aerosol residual remediation model in the current block is evaluated; Input the confidence level for AOD; The aerosol fusion attenuation constant; Let p be the Euclidean distance from pixel p to the center of the i-th image block.

Images