Atmospheric correction method and device based on deep learning inversion AOD (Argon Oxygen Decarburization) and medium

Abstract

The invention discloses a deep learning inversion AOD-assisted atmospheric correction-based method and device and a medium, and relates to the field of atmospheric remote sensing and deep learning cross technologies, and the method comprises the steps: constructing a multi-type training sample set fusing actual measurement and physical simulation data, and carrying out mass screening and layering processing to obtain a multi-type training sample set; obtaining a pre-training sample, a fine tuning sample and an extreme scene sample matched with the satellite remote sensing data; establishing a deep learning network model taking data loss and inverse operator constraint loss weighted fusion as a total loss function, and adopting a cross-satellite transfer learning strategy to sequentially complete pre-training, satellite exclusive fine tuning and extreme scene enhancement training to obtain a trained aerosol optical thickness inversion model; radiometric calibration, cloud detection, geometric correction preprocessing and input model reasoning calculation are carried out on satellite remote sensing original data, post-processing and quality grading are carried out on an output result, and an aerosol optical thickness inversion result used for assisting atmospheric correction is obtained.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of atmospheric remote sensing and deep learning, and in particular to a method and device for deep learning-based AOD inversion assisted atmospheric correction and a medium. BACKGROUND

[0002] Aerosol Optical Depth (AOD) is a core physical parameter representing the concentration of atmospheric aerosols. Accurate inversion results are the key basis for atmospheric correction of remote sensing images, and have irreplaceable important value in the fields of atmospheric environmental research and application such as air quality monitoring, climate effect evaluation, and atmospheric pollution early warning. Satellite remote sensing technology has become the mainstream method for large-scale AOD inversion due to its wide coverage and high observation efficiency. How to achieve high-precision and high-generalization AOD inversion based on satellite remote sensing data and further improve the effect of atmospheric correction has become a research hotspot and key technology demand in the field of atmospheric remote sensing.

[0003] The existing AOD inversion methods for satellite remote sensing data mainly include physical driving algorithms, data-driven models, and hybrid methods combining the two. Physical driving algorithms such as the dark target method and the deep blue algorithm rely on aerosol radiation transfer physical models for inversion and require specific surface reflection assumptions. Data-driven models such as traditional convolutional neural networks and deep neural networks achieve AOD prediction by mining the statistical correlation between satellite and ground observation data, and rely on a large number of spatiotemporally matched sample data. Hybrid methods attempt to combine physical models and machine learning techniques, mostly in the form of simple stacking of physical model outputs as model input features.

[0004] However, physical driving algorithms rely on specific surface assumptions, and their inversion accuracy decreases significantly in heterogeneous surfaces and bright surfaces. Data-driven models are limited by the coverage of spatiotemporally matched samples, and their bias is significant in non-training scenarios such as extreme pollution and high latitudes. They are also prone to output abnormal values without physical meaning. Simple stacking of hybrid methods cannot take full advantage of the synergy of physical models and machine learning, and still cannot solve the multi-solution problem of underdetermined inversion and effectively avoid error accumulation during data conversion. The existing AOD inversion methods all have obvious technical defects, making it difficult to meet the demand for high-precision and high-generalization AOD inversion in complex surface and atmospheric scenarios, and further affecting the actual effect of atmospheric correction. SUMMARY

[0005] Therefore, the present application provides a method and device for deep learning-based AOD inversion assisted atmospheric correction and a medium, which can solve the problem of existing AOD inversion methods having obvious technical defects and being difficult to meet the demand for high-precision and high-generalization AOD inversion in complex surface and atmospheric scenarios.

[0006] According to a first aspect of the present application, a method for AOD auxiliary atmospheric correction based on deep learning inversion is provided, comprising: A multi-type training sample set is constructed by fusing measured data and physically simulated data, quality screening and hierarchical processing are performed on the multi-type training sample set, and pre-training samples, fine-tuning samples and extreme scene samples suitable for satellite remote sensing data are obtained; A deep learning network model is created, the deep learning network model updates parameters through a total loss function, the total loss function is a weighted fusion function of data loss and inverse operator constraint loss, the data loss represents the deviation between the model prediction value and the measured value, and the inverse operator constraint loss represents the degree to which the model prediction value conforms to the aerosol radiation transfer physical law; The deep learning network model is sequentially pre-trained, satellite-specific fine-tuned and subjected to extreme scene enhanced training through a cross-satellite transfer learning strategy, and a trained aerosol optical thickness inversion model is obtained; Radiometric calibration, cloud detection and geometric correction preprocessing operations are performed on satellite remote sensing raw data, the preprocessed satellite remote sensing raw data is input into the trained aerosol optical thickness inversion model for inference calculation, post-processing and quality grading are performed on the model output results, and aerosol optical thickness inversion results for auxiliary atmospheric correction are obtained.

[0007] According to a second aspect of the present application, an apparatus for AOD auxiliary atmospheric correction based on deep learning inversion is provided, comprising: A construction module is configured to construct a multi-type training sample set by fusing measured data and physically simulated data, perform quality screening and hierarchical processing on the multi-type training sample set, and obtain pre-training samples, fine-tuning samples and extreme scene samples suitable for satellite remote sensing data; A creation module is configured to create a deep learning network model, the deep learning network model updates parameters through a total loss function, the total loss function is a weighted fusion function of data loss and inverse operator constraint loss, the data loss represents the deviation between the model prediction value and the measured value, and the inverse operator constraint loss represents the degree to which the model prediction value conforms to the aerosol radiation transfer physical law; A training module is configured to sequentially pre-train, satellite-specific fine-tune and perform extreme scene enhanced training on the deep learning network model through a cross-satellite transfer learning strategy, and obtain a trained aerosol optical thickness inversion model; The processing module is used to perform preprocessing operations such as radiometric calibration, cloud detection, and geometric correction on the raw satellite remote sensing data. The preprocessed raw satellite remote sensing data is then input into the trained aerosol optical thickness inversion model for inference calculation. The model output results are then post-processed and quality graded to obtain aerosol optical thickness inversion results used to assist atmospheric correction.

[0008] According to a third aspect of this application, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described method for AOD-assisted atmospheric correction based on deep learning inversion.

[0009] According to a fourth aspect of this application, an electronic device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described method for AOD-assisted atmospheric correction based on deep learning inversion.

[0010] By employing the above technical solutions, this application provides a method, apparatus, and medium for AOD-assisted atmospheric correction based on deep learning inversion. By constructing a multi-type training sample set that integrates measured and physical simulation data and performing quality screening and stratification, it can compensate for the deficiency of insufficient spatiotemporal matching sample coverage and effectively solve the problem of large inversion bias in extreme and non-training scenarios for data-driven models. By creating a deep learning network model that integrates data loss and inverse operator constraint loss into a total loss function, the physical laws of aerosol radiative transfer are deeply integrated into the model parameter update process. This avoids the dependence of physical-driven algorithms on specific surface assumptions and solves the problem of their application on heterogeneous surfaces. The problem of decreased accuracy in bright surfaces can be addressed by fundamentally constraining model outputs to conform to physical laws, avoiding the generation of physically meaningless outliers. It also overcomes the drawbacks of simple superposition in traditional hybrid methods, effectively solving the multiple solutions of underdetermined inverse problems and avoiding the accumulation of data transformation errors. Through phased training using cross-satellite transfer learning strategies, the model's generalization ability and adaptability to extreme scenarios can be further improved. Combined with satellite remote sensing data preprocessing and inversion result post-processing and quality grading, high-precision, high-generalization inversion of aerosol optical thickness (AOD) in complex surface and atmospheric scenarios can be achieved, significantly enhancing the practical effect of AOD inversion results in assisting atmospheric correction.

[0011] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0012] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 The diagram shows a flowchart of a method for AOD-assisted atmospheric correction based on deep learning inversion, according to an embodiment of this application. Figure 2 A flowchart illustrating a method for AOD-assisted atmospheric correction based on deep learning inversion, according to another embodiment of this application, is shown. Figure 3 This illustration shows a flowchart of a multi-type training sample set construction provided in an embodiment of this application; Figure 4 This illustration shows a flowchart of a phased training process for a deep learning network model according to an embodiment of this application. Figure 5 This paper illustrates a flowchart of a performance verification process for aerosol optical thickness inversion results provided in an embodiment of this application. Figure 6 A schematic diagram of a device for AOD-assisted atmospheric correction based on deep learning inversion is shown in an embodiment of this application. Figure 7 A schematic diagram of a device for AOD-assisted atmospheric correction based on deep learning inversion is shown in another embodiment of this application. Detailed Implementation

[0013] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0014] Existing AOD retrieval methods for satellite remote sensing data can be mainly classified into three categories: physics-driven algorithms, data-driven models, and hybrid methods combining both. Physics-driven algorithms, such as the dark target method and the Deep Blue algorithm, rely on aerosol radiative transfer physics models for retrieval and require the establishment of specific surface reflection assumptions. Data-driven models, such as traditional convolutional neural networks and deep neural networks, achieve AOD prediction by mining the statistical correlation between satellite and ground observation data, relying on a large amount of spatiotemporally matched sample data. Hybrid methods attempt to integrate physical models and machine learning techniques, often using a simple superposition of physical model outputs as model input features.

[0015] However, physics-driven algorithms, relying on specific surface assumptions, suffer significant accuracy drops in inversion scenarios such as heterogeneous or bright surfaces. Data-driven models, limited by the coverage of spatiotemporal matching samples, exhibit significant biases in non-training scenarios such as extreme pollution and high latitudes, and are prone to outputting physically meaningless outliers. Simple superposition hybrid methods fail to leverage the synergistic advantages of physical models and machine learning, still struggle to address the multiple solutions inherent in the underdetermined inversion problem, and cannot effectively avoid error accumulation during data transformation. All of these existing AOD inversion methods suffer from significant technical deficiencies, making it difficult to meet the high-precision, high-generalization AOD inversion requirements of complex surface and atmospheric scenarios, thus affecting their actual effectiveness in assisting atmospheric correction.

[0016] To solve the above technical problems, such as Figure 1 As shown, this embodiment of the invention provides a method for AOD-assisted atmospheric correction based on deep learning inversion, the method comprising: Step 110: Construct a multi-type training sample set that integrates measured data and physical simulation data. Perform quality screening and stratification on the multi-type training sample set to obtain pre-training samples, fine-tuning samples and extreme scenario samples adapted to satellite remote sensing data.

[0017] Among them, the measured data refers to the raw observation data from satellite remote sensing, the aerosol optical thickness data obtained from actual monitoring at ground stations, and atmospheric-related auxiliary data used for auxiliary data processing. It is the data source that directly reflects the actual observation characteristics of the atmosphere and the land surface. The physical simulation data is aerosol optical thickness-related simulated observation data generated by operators based on the radiative transfer model, combined with multiple atmospheric models, aerosol models, surface reflectance, and geometric parameters. It can be used to supplement the insufficient coverage of the measured data. The multi-type training sample set is a comprehensive training dataset formed by integrating the measured data and the physical simulation data. It contains samples under different observation scenarios, different aerosol types, and different land surface and atmospheric conditions. The quality screening is the abnormal data filtering process performed on the multi-type training sample set to remove samples with values ​​exceeding reasonable limits. Invalid samples with no physical meaning in their scope or parameter combinations are excluded to ensure the validity and physical rationality of the sample data. Layered processing involves classifying and organizing the valid samples after quality screening according to different stages of model training, based on dimensions such as satellite sensor adaptability, study area characteristics, and scene type. Pre-training samples, after quality screening, are samples adapted to cross-satellite learning requirements, supporting the model in learning the general mapping relationship between aerosols and radiation signals without relying on specific satellite sensor characteristics. Fine-tuning samples are selected based on the characteristics of the target satellite sensor and the atmospheric and surface characteristics of the target study area, used for dedicated model adaptation training. Extreme scenario samples are constructed for scenarios with insufficient coverage of measured samples, such as extreme pollution, to improve the model's ability to invert extreme aerosol events.

[0018] In this embodiment of the disclosure, a multi-type training sample set covering multiple scenarios and types can be formed by first fusing satellite remote sensing, ground-based measured data, and atmospheric-assisted measured data with physical simulation data generated based on the radiative transfer model. Then, a systematic quality screening is carried out on this sample set to remove invalid samples that exceed numerical limits or have no physical meaning, ensuring the basic quality of the samples. Subsequently, based on the inversion requirements of satellite remote sensing data and the goal of phased training of the model, the qualified samples are processed in layers to form pre-training samples adapted for cross-satellite general learning, fine-tuning samples matching the characteristics of the target satellite and the study area, and extreme scenario samples to supplement extreme scenario coverage, providing hierarchical and targeted sample support for the subsequent phased training of the model.

[0019] By integrating measured data and physical simulation data to construct multi-type training sample sets, the limitations of single measured data in scene coverage can be effectively compensated, enriching the diversity and comprehensiveness of the samples. The quality screening operation on the sample sets can remove invalid samples from the source, ensuring the physical rationality and validity of the training data and avoiding interference from abnormal data on model training. Based on the satellite remote sensing inversion requirements and model training objectives, hierarchical processing is carried out to generate pre-training, fine-tuning, and extreme scene samples adapted to different training stages. This allows the model training to learn different features in stages and in a targeted manner. It can not only solve the problem of weak model generalization ability caused by insufficient measured samples, but also supplement the sample coverage of extreme scenes, greatly improving the matching degree between training samples and model stage training requirements. This can lay a solid data foundation for the subsequent model to achieve high-precision aerosol optical thickness inversion in complex scenes.

[0020] Step 120: Create a deep learning network model. The deep learning network model updates parameters through a total loss function. The total loss function is a weighted fusion function of data loss and inverse operator constraint loss. Data loss characterizes the deviation between the model's predicted value and the measured value, while inverse operator constraint loss characterizes the degree to which the model's predicted value conforms to the physical laws of aerosol radiative transfer.

[0021] Among them, the deep learning network model is a network model for aerosol optical thickness inversion built on a deep learning framework. It integrates functional modules such as feature extraction, physical constraints, and output correction, and can achieve accurate prediction of aerosol optical thickness through data learning and physical constraints. The total loss function is a comprehensive function used to measure the deviation of the model's prediction results. It is formed by weighted fusion of loss functions of different dimensions and is the core basis for guiding model parameter updates and optimizing the model's inversion accuracy. Parameter updates are performed during model training by adjusting the parameters of each layer of the network according to the deviation calculated by the loss function, through backpropagation and other methods, so that the model can improve its prediction accuracy. The process of the measured results continuously approaching the true value; the data loss is a loss function that characterizes the deviation between the model's predicted aerosol optical thickness and the actual ground observation value, reflecting the degree of fit between the model's prediction and the measured data; the inverse operator constraint loss is a loss function constructed based on the physical laws of aerosol radiative transfer, characterizing whether the model's predicted value conforms to the physical logic of atmospheric aerosol radiative transfer, and is used to constrain the physical rationality of the model output; the weighted fusion function is a function form that assigns corresponding weights to the data loss and the inverse operator constraint loss and then performs fusion calculation, which can achieve the dual optimization goals of data fitting accuracy and conformity to physical laws.

[0022] In this embodiment of the disclosure, a deep learning network model suitable for aerosol optical thickness inversion can be built first. To achieve dual optimization of model prediction accuracy and physical rationality, a total loss function consisting of data loss and inverse operator constraint loss is designed as the core basis for updating model parameters. The data loss is used to quantify the deviation between the model prediction value and the measured value, and the inverse operator constraint loss is used to quantify the fit between the model prediction value and the physical laws of aerosol radiation transfer. The two types of loss functions are integrated into a total loss function through weighted fusion. During the training process, the model uses this total loss function as the optimization target and drives the parameter update of each module of the network through backpropagation, so that the model always follows the physical laws of aerosol radiation transfer while learning the statistical laws of data.

[0023] By building a dedicated deep learning network model, an algorithmic framework for aerosol optical thickness inversion is provided. A weighted fusion total loss function is constructed by combining data loss and inverse operator constraint loss, and this function is used as the basis for updating model parameters. The data loss allows the model to learn the statistical correlation of measured data, improving the fit between the model's predictions and the actual values. The inverse operator constraint loss deeply integrates the physical laws of aerosol radiative transfer into the model training process, fundamentally constraining the model output to conform to physical logic and avoiding physicalally meaningless outliers. At the same time, the weighted fusion method can achieve synergistic optimization of data fitting and physical constraints, which can solve the problems of limited adaptability of traditional physical-driven algorithms, lack of physical rationality of pure data-driven models, and the inability of traditional hybrid methods to leverage synergistic advantages through simple superposition. This effectively improves the inversion accuracy and generalization ability of the model in complex surface and atmospheric scenarios. It can also solve the multiple solutions of the underdetermined inversion problem at the algorithm level and avoid the impact of error accumulation during data transformation on the inversion results.

[0024] Step 130: Using a cross-satellite transfer learning strategy, the deep learning network model is trained in stages, including pre-training, satellite-specific fine-tuning, and extreme scenario enhancement training, to obtain the trained aerosol optical thickness inversion model.

[0025] The cross-satellite transfer learning strategy is a learning method that relies on observation data from different satellite sensors to train the model. It first learns general features for aerosol retrieval based on non-target satellite data, and then adapts specific features based on target satellite data, achieving cross-satellite feature transfer and model adaptation, thus improving the model's retrieval capability for target satellite data. Pre-training is the initial stage of model training, using samples adapted for cross-satellite learning as training data to train the model's basic feature extraction layer, allowing the model to learn the general mapping relationship between aerosols and radiation signals, laying the foundation for subsequent targeted training. Satellite-specific fine-tuning is a targeted training stage based on pre-training, freezing the parameters of the trained general feature layer to adapt to... The target satellite samples serve as training data, training subsequent network layers and output modules of the model to adapt the model to the characteristics of the target satellite sensors and the atmospheric and surface features of the corresponding study area. Extreme scenario enhancement training is an intensive training phase to improve the model's extreme scenario inversion capability. It integrates extreme scenario samples and target satellite fine-tuning samples for model training, and optimizes model parameters in a targeted manner to enable the model to capture the characteristic patterns of extreme aerosol events. Phased training is a training method that divides model training into multiple progressive stages, including pre-training, satellite-specific fine-tuning, and extreme scenario enhancement training, based on the model's learning patterns and actual inversion needs. Each stage sets different training objectives and strategies to achieve a gradual improvement in the model's capabilities.

[0026] In this embodiment of the disclosure, a cross-satellite transfer learning strategy can be used as the core guide. Following the progressive logic of feature learning from general to specific and scenario adaptation from conventional to extreme, the deep learning network model is trained in stages. First, the model is pre-trained using pre-training samples to learn the general features of aerosol inversion across satellites. Then, the parameters of the general feature layer obtained from the pre-training are frozen. Satellite-specific fine-tuning is carried out using fine-tuning samples to adapt the model to the sensor characteristics of the target satellite and the atmospheric and surface features of the corresponding region. Finally, extreme scenario samples and fine-tuning samples are fused to carry out extreme scenario enhancement training to specifically optimize the model's inversion capability for extreme aerosol scenarios. Through multi-stage progressive training and parameter optimization, an aerosol optical thickness inversion model that adapts to the target satellite and can cover both conventional and extreme scenarios is finally obtained.

[0027] By employing a cross-satellite transfer learning strategy to implement phased training—including pre-training, satellite-specific fine-tuning, and extreme scenario enhancement training—the model can effectively utilize observational data resources from different satellites, addressing the problem of insufficient measured samples from a single satellite. This allows the model to first master general aerosol inversion features before adapting to the target satellite's specific characteristics, significantly improving the model's adaptability to remote sensing data from the target satellite. Simultaneously, extreme scenario enhancement training supplements feature learning for extreme scenarios, effectively addressing the issue of large inversion biases in non-training scenarios such as extreme pollution. This phased training approach makes the model's learning process more aligned with the actual needs of aerosol inversion. The progressive parameter optimization gives the model both general cross-satellite generalization capabilities and target satellite-specific inversion accuracy, while also enabling effective capture of extreme aerosol scenarios. Overall, this improves the model's inversion accuracy and generalization capabilities in complex atmospheric and surface scenarios, ensuring that the model can output stable and accurate aerosol optical thickness inversion results in both conventional and extreme scenarios.

[0028] Step 140: Perform preprocessing operations such as radiometric calibration, cloud detection, and geometric correction on the raw satellite remote sensing data. Input the preprocessed raw satellite remote sensing data into the trained aerosol optical thickness inversion model for inference calculation. Perform post-processing and quality classification on the model output results to obtain aerosol optical thickness inversion results used to assist atmospheric correction.

[0029] Radiometric calibration, which converts the digital quantization values ​​of raw satellite remote sensing data into physically meaningful radiance values, and then further calculates the top atmospheric reflectance, is a fundamental operation for eliminating satellite sensor errors and ensuring the quantitative accuracy of observation data. Specifically, the calibration accuracy can be optimized by using the sensor's built-in calibration coefficients and introducing a bias correction term: L = a×DN+b+ΔL (where a and b are calibration coefficients taken from the satellite payload manual, and ΔL is the bias correction term, obtained through linear regression fitting based on synchronous observation AERONET station data). Then, the TOA reflectance ρ = πLd² / (ES0 cosθs) is calculated by normalizing the solar irradiance (where d is the Earth-Sun distance correction factor, ES0 is the top atmospheric solar irradiance, and θs is the solar zenith angle). After calibration, the TOA reflectance must satisfy the RMSE (Root Mean Square) of the 6S simulated value. Square Error (root mean square error) ≤ 0.02; Cloud detection involves identifying and removing cloud pixels and affected pixels at cloud edges from satellite remote sensing data using specific feature recognition and threshold determination methods, thus avoiding interference from cloud cover on aerosol optical thickness inversion results; Geometric correction eliminates geometric distortions in satellite remote sensing data caused by sensor attitude, terrain, and other factors through coordinate system transformation and error correction methods, ensuring the spatial accuracy of remote sensing data; Inference calculation involves inputting preprocessed satellite remote sensing data into a trained model, which then automatically outputs aerosol optical thickness based on learned feature patterns and physical constraints. The process of aerosol optical thickness prediction results includes: post-processing, which optimizes the preliminary aerosol optical thickness inversion results output by the model to eliminate isolated outliers and improve the spatial continuity and data reliability of the inversion results; quality grading, which classifies and marks the reliability of the aerosol optical thickness inversion results according to specific evaluation indicators, providing a quality reference for the practical application of the inversion results; and atmospheric correction, which eliminates the influence of atmospheric scattering, absorption and other factors on satellite remote sensing observation signals and restores the true surface reflectance of remote sensing data. The aerosol optical thickness inversion results are an important input parameter for carrying out high-precision atmospheric correction.

[0030] In this embodiment of the present disclosure, a systematic preprocessing operation can be performed on the raw satellite remote sensing data, sequentially completing radiometric calibration, cloud detection, and geometric correction to eliminate various errors caused by the sensor itself, cloud cover, and geometric distortion, thereby obtaining effective satellite remote sensing data with quantitative accuracy, spatial accuracy, and no cloud interference. Then, the preprocessed effective data is input into the trained aerosol optical thickness inversion model, which performs inference calculations on aerosol optical thickness and outputs preliminary results. Finally, post-processing and quality grading operations are performed on the preliminary inversion results output by the model to optimize the data quality of the inversion results and classify the reliability level, ultimately obtaining aerosol optical thickness inversion results that can be directly used to assist atmospheric correction.

[0031] By performing preprocessing operations such as radiometric calibration, cloud detection, and geometric correction on the raw satellite remote sensing data, observation errors caused by various non-aerosol factors can be eliminated at the source, ensuring the accuracy and effectiveness of the input model data. This lays the data foundation for subsequent high-precision aerosol optical thickness inference calculations. Inputting the preprocessed effective data into the trained model for inference calculations can fully leverage the model's advantages in integrating data patterns and physical constraints, quickly outputting preliminary aerosol optical thickness results that conform to actual atmospheric conditions. Post-processing and quality grading of the model output results can further optimize the spatial continuity and data reliability of the inversion results. At the same time, it can provide a clear quality reference for the practical application of the inversion results, enabling the inversion results to accurately match the accuracy requirements of atmospheric correction. Overall, it achieves high-precision and high-reliability inversion of aerosol optical thickness, effectively improving the actual effect of inversion results in assisting atmospheric correction and solving the problem that the accuracy of atmospheric correction is limited by data errors and poor result quality in existing inversion results.

[0032] The method for AOD-assisted atmospheric correction based on deep learning provided by this invention, by constructing a multi-type training sample set that integrates measured and physical simulation data and performing quality screening and stratification, can compensate for the deficiency of insufficient spatiotemporal matching sample coverage and effectively solve the problem of large inversion bias in extreme and non-training scenarios of data-driven models. By creating a deep learning network model that integrates the total loss function of data loss and inverse operator constraint loss, the physical laws of aerosol radiative transfer are deeply integrated into the model parameter update process. This avoids the dependence of physical-driven algorithms on specific surface assumptions and solves the accuracy problem on heterogeneous and bright surfaces. The descent problem can fundamentally constrain the model output to conform to physical laws, avoiding the generation of outliers without physical meaning. At the same time, it can overcome the shortcomings of simple superposition in traditional hybrid methods, effectively solve the multiple solutions of underdetermined inverse problems, and avoid the problem of data transformation error accumulation. By carrying out phased training through cross-satellite transfer learning strategies, the generalization ability and adaptability to extreme scenarios of the model can be further improved. Combined with satellite remote sensing data preprocessing and inversion result postprocessing and quality classification, high-precision and high-generalization inversion of aerosol optical thickness under complex surface and atmospheric scenarios can be achieved, which greatly improves the actual effect of AOD inversion results in assisting atmospheric correction.

[0033] Furthermore, as a refinement and extension of the specific implementation of the above embodiments, and to fully illustrate the implementation of this embodiment, this embodiment also provides another method for AOD-assisted atmospheric correction based on deep learning inversion, such as... Figure 2 As shown, the method includes: Step 210: Construct a multi-type training sample set that integrates measured data and physical simulation data. Perform quality screening and stratification on the multi-type training sample set to obtain pre-training samples, fine-tuning samples and extreme scenario samples adapted to satellite remote sensing data.

[0034] In specific application scenarios, when constructing a multi-type training sample set that integrates measured data and physical simulation data, it is first necessary to clarify the generation rules of core physical quantities and geometric parameters: the aerosol optical thickness (AOD) ranges from 0.01 to 5.0, covering clean atmosphere to extreme pollution scenarios, and adopts logarithmic uniform distribution sampling to avoid redundancy of low AOD samples; geometric parameters such as solar zenith angle, observation zenith angle, and relative azimuth angle are sampled according to uniform distribution to simulate real observation conditions. Surface reflectance was calculated using a pixel decomposition method, decomposing surface pixels into six endmembers: vegetation, soil, water, snow and ice, building or artificial surfaces, and desert. The reflectance of each endmember in the four AGRI bands was obtained based on the ASTER 2.0 spectral library. The surface reflectance of each pixel was calculated by weighting it using endmember abundance (the sum of the abundances of the six endmembers is 1), with the formula: ρ_surface = f_veg×ρ_veg + f_soil×ρ_soil + f_water×ρ_water + f_ice×ρ_ice + f_urban×ρ_urban +f_desert×ρ_desert, where ρ_surface is the final surface reflectance of a single pixel; f_veg is the vegetation endmember abundance; ρ_veg is the reflectance of the vegetation endmember in the corresponding band; f_soil is the soil endmember abundance; ρ_soil is the soil endmember reflectance in the corresponding band; f_water is the water endmember abundance; ρ_water is the water endmember reflectance in the corresponding band; f_ice is the snow and ice endmember abundance; ρ_ice is the snow and ice endmember reflectance in the corresponding band; f_urban is the building or artificial surface endmember abundance; ρ_urban is the building or artificial surface endmember reflectance in the corresponding band; f_desert is the desert endmember abundance; ρ_desert is the desert endmember reflectance in the corresponding band.

[0035] For embodiments of this disclosure, such as Figure 3As shown, the process can begin with the collection and preprocessing of measured data, including satellite remote sensing Level 1B data, ground-based measured AOD data, and atmospheric auxiliary data. Radiometric calibration, cloud detection, and geometric correction are performed on the satellite data, while outlier removal and spatiotemporal matching are applied to the ground-based measured data to obtain spatiotemporally matched measured samples. Then, an aerosol radiative transfer positive operator is constructed based on the 6S radiative transfer model, incorporating surface reflectance calculated using multiple atmospheric models, aerosol models, and the surface endmember decomposition method. Initial physical simulation data is generated through operator operations, and random noise consistent with satellite observation characteristics is added to enhance the data, resulting in physical simulation data. Finally, the measured samples and physical simulation data are fused to form an initial multi-type training sample set. This set is then subjected to quality screening, removing invalid samples with AOD values ​​and satellite top-space reflectance exceeding preset ranges, and filtering parameter combinations without physical meaning during the physical simulation process. This yields a multi-type training sample set of fused measured and physical simulation data with complete quality control. Accordingly, step 210 in the embodiment may include: Step 210-1: Collect satellite remote sensing Level 1B data, ground-measured AOD data, and atmospheric auxiliary data.

[0036] Among them, satellite remote sensing Level 1B data consists of raw observation data acquired by satellite sensors after radiometric calibration and geometric coarse correction. It is mainly expressed in digital quantization values ​​and includes radiation observation information at the top of the atmosphere. It is the basic remote sensing data source for aerosol optical thickness inversion. Ground-based measured AOD data consists of measured aerosol optical thickness values ​​acquired at the site through observation equipment such as ground-based solar photometers. These are valid data after being screened by quality labeling and serve as real reference data for verifying and training aerosol optical thickness inversion models. Atmospheric auxiliary data consists of various atmospheric-related data used to assist in satellite remote sensing data processing, aerosol radiative transfer simulation, and model construction. These data include key parameters such as ground object spectral information, atmospheric precipitable water, and atmospheric model type, providing atmospheric background support for aerosol optical thickness inversion.

[0037] Step 210-2: Based on atmospheric auxiliary data, perform preprocessing operations such as radiometric calibration, cloud detection, and geometric correction on the satellite remote sensing Level 1B data. Based on the preprocessed satellite remote sensing Level 1B data, perform outlier removal and spatiotemporal matching processing on the ground measured AOD data to obtain spatiotemporally matched measured samples.

[0038] The preprocessing operation consists of a series of standardized processing operations performed on Level 1B satellite remote sensing data. Through radiometric calibration, cloud detection, and geometric correction, various errors and interferences in the data are eliminated, ensuring quantitative accuracy, spatial accuracy, and validity, laying the foundation for subsequent data processing and model training. Outlier removal involves filtering out invalid data with unqualified quality labels or values ​​exceeding reasonable ranges from ground-based measured AOD data according to preset quality evaluation standards, retaining valid measured data with real observational value. Spatiotemporal matching processing involves associating preprocessed satellite remote sensing data with ground-based measured AOD data based on preset matching rules in time and space dimensions, establishing a one-to-one correspondence between satellite observation data and ground-based measured data at corresponding locations and times. The spatiotemporal matching measured sample is a comprehensive sample formed by associating ground-based measured AOD data after outlier removal with preprocessed Level 1B satellite remote sensing data within the same time and spatial range. This sample is the core sample reflecting actual atmospheric observation characteristics during model training.

[0039] In this embodiment of the disclosure, atmospheric auxiliary data can be used as an important support. First, preprocessing operations such as radiometric calibration, cloud detection, and geometric correction are performed on the satellite remote sensing Level 1B data in sequence. The preprocessing effect is optimized by using relevant parameters in the atmospheric auxiliary data to eliminate various data errors caused by the satellite sensor itself, cloud cover, and geometric distortion, so as to obtain accurate and effective satellite remote sensing observation data. Then, using the preprocessed satellite remote sensing Level 1B data as a reference, data cleaning is performed on the ground measured AOD data. Outlier removal is performed according to the established standards to filter out invalid measured data. Then, following the temporal and spatial matching rules, the ground measured AOD data after outlier removal is correlated and fused with the preprocessed satellite remote sensing data in the corresponding temporal and spatial range, and finally a temporal and spatial matching measured sample with both satellite observation information and ground real measured data is formed.

[0040] By performing radiometric calibration, cloud detection, and geometric correction preprocessing on Level 1B satellite remote sensing data supported by atmospheric auxiliary data, systematic errors and external interference in the satellite remote sensing data can be effectively eliminated, improving the quantitative and spatial accuracy of satellite observation data. This provides a high-quality remote sensing data foundation for subsequent data association and matching. Using the preprocessed satellite remote sensing data as a reference to remove outliers from ground-based measured AOD data can ensure the validity and authenticity of ground-based measured data from the source, avoiding the impact of invalid data on sample quality. Spatiotemporal matching processing can achieve precise correlation between satellite remote sensing data and ground-based measured data in the temporal and spatial dimensions, allowing the resulting measured samples to simultaneously reflect the satellite observation characteristics of aerosols and the actual ground observation results. This effectively compensates for the limitations of single data and provides real and accurate measured data support for subsequently building a high-quality training sample set and improving the model inversion accuracy. At the same time, it makes the sample data more consistent with the spatiotemporal characteristics of actual atmospheric observations, laying a key data foundation for the model to learn the true radiometric characteristics of aerosols.

[0041] Step 210-3: Construct an aerosol radiative transfer positive operator based on the 6S radiative transfer model. Using atmospheric model parameters from atmospheric auxiliary data as a reference, integrate multiple atmospheric models, multiple aerosol models, and the surface reflectance calculated using the surface endmember decomposition method into the aerosol radiative transfer positive operator. Generate initial physical simulation data through operator operations. Add random noise that conforms to satellite observation characteristics to the initial physical simulation data to complete the simulation data enhancement and obtain the final physical simulation data.

[0042] Among them, the 6S radiative transfer model is a radiative transfer model used to calculate the effects of atmospheric scattering and absorption of solar radiation. It can accurately achieve a forward mapping from atmospheric physical parameters to radiation signals observed by satellite, and is a core tool for physical simulation in aerosol optical thickness inversion. The aerosol radiative transfer positive operator is a forward mapping operator built based on the radiative transfer model, which can realize the quantitative calculation of atmospheric top reflectivity observed by satellite sensors from physical parameters such as aerosols, atmosphere, surface, and geometry. Atmospheric model parameters are parameters that characterize the vertical distribution characteristics of atmospheric elements such as temperature, humidity, and air pressure under different atmospheric environments, and are important basic parameters for radiative transfer calculation. Atmospheric models are standardized atmospheric vertical structure types classified according to different latitudes and seasons, which can reflect the atmospheric radiative transfer characteristics of different regions and time periods. Aerosol models are standardized aerosol types classified based on aerosol optical properties, which can reflect different... The study investigates the radiative scattering and absorption patterns of similar aerosols. The surface endmember decomposition method decomposes complex surface pixels into several single-type surface endmembers, and calculates the comprehensive surface reflectance of pixels by weighting endmember abundance, thus accurately obtaining surface reflectance parameters. The initial physical simulation data consists of aerosol-related simulated observation data directly generated from various physical parameters through aerosol radiative transfer positive operator operations, without considering satellite observation errors. Random noise is noise added to simulate the accidental errors generated during satellite sensor observations, and its characteristics match those of actual satellite observation errors. Simulation data enhancement involves adding noise to the initial physical simulation data that conforms to actual observation characteristics, making the simulation data closer to real satellite observation data. The final physical simulation data is aerosol optical thickness-related simulated data that combines physical conformity with satellite observation characteristics after simulation data enhancement.

[0043] In this embodiment of the disclosure, an aerosol radiative transfer positive operator can be constructed based on the 6S radiative transfer model. First, atmospheric model parameters in atmospheric auxiliary data are used as important references, and multiple atmospheric models and multiple aerosol models are incorporated into the parameter system of the positive operator. At the same time, the accurate surface reflectance is calculated by the surface endmember decomposition method and incorporated into it to improve the physical parameter support of the positive operator. Then, various atmospheric, aerosol, surface and related geometric parameters are input into the aerosol radiative transfer positive operator. The initial physical simulation data that can reflect the physical laws of aerosol radiative transfer is generated through the forward operation of the operator. Finally, in order to make the simulation data more consistent with the actual satellite observation, random noise that conforms to the satellite observation characteristics is added to the initial physical simulation data to carry out simulation data enhancement processing, and finally the final physical simulation data with both physical rationality and observational authenticity is obtained.

[0044] By constructing an aerosol radiative transfer positive operator based on the 6S radiative transfer model and incorporating surface reflectance calculated using multiple atmospheric models, aerosol models, and the surface endmember decomposition method, the generated initial physical simulation data strictly adheres to the physical laws of aerosol radiative transfer. This ensures the physical rationality of the simulation data and provides physical data support that closely reflects actual atmospheric conditions for subsequent model training. By incorporating various model parameters with atmospheric model parameters as a reference, the simulation data can cover diverse atmospheric, aerosol, and surface scenarios, enriching the scenario diversity of the simulation data. Furthermore, random noise consistent with satellite observation characteristics is added to the initial physical simulation data. Data augmentation makes the final physical simulation data closer to the actual satellite observation data, which can effectively reduce the deviation between pure physical simulation and actual observation, avoid overfitting problems caused by data deviation in model training, and the final physical simulation data not only has the support of physical laws, but also fits the actual observation characteristics and covers diverse scenarios. It can effectively supplement the insufficient coverage of the measured data. After being fused with the measured data, a high-quality training sample set can be constructed, which lays a solid physical data foundation for improving the model's inversion accuracy and generalization ability. At the same time, it can also effectively reduce the impact of error accumulation during the data conversion process on the aerosol optical thickness inversion results.

[0045] Step 210-4: Fuse the spatiotemporally matched measured samples with the physical simulation data to obtain an initial multi-type training sample set.

[0046] Accordingly, when performing quality screening and stratification on multiple types of training sample sets to obtain pre-training samples, fine-tuning samples, and extreme scenario samples adapted to satellite remote sensing data, step 210 of the embodiment may include: Step 210-5: Perform quality screening on the multi-type training sample set. The quality screening includes removing invalid samples whose AOD value and satellite top reflectivity exceed the preset range, and filtering parameter combination samples that have no physical meaning in the physical simulation process.

[0047] The quality screening process involves systematically cleaning the multi-type training sample sets after fusion. By setting data ranges and physical rationality judgment rules, invalid and abnormal samples are removed from the sample sets, ensuring the effectiveness and physical compliance of the training samples. The preset range is a reasonable numerical range for AOD values ​​and satellite top-sky reflectivity, defined based on the actual atmospheric observation laws of aerosol optical thickness inversion and the physical characteristics of satellite remote sensing data. This range serves as a quantitative standard for determining the validity of a sample. Invalid samples are those whose AOD values ​​and satellite top-sky reflectivity in the multi-type training sample sets exceed the reasonable numerical range and cannot truly reflect the actual situation of atmospheric aerosols and satellite observations. Samples with parameter combinations that lack physical meaning are simulation samples where the parameter combinations do not conform to objective physical laws such as atmospheric radiation transfer and satellite observations, resulting in distorted simulation results or calculation failures.

[0048] By conducting quality screening on various types of training sample sets, including numerical verification and physical rationality checks, invalid samples with AOD values ​​and satellite top-sky reflectivity exceeding preset ranges are directly removed. This ensures the authenticity and rationality of the training samples from a numerical perspective, preventing abnormal numerical samples from interfering with the normal learning process of the model. At the same time, filtering out parameter combinations that have no physical meaning during the physical simulation process ensures the effectiveness of the simulation samples from a physical law perspective. This ensures that the training sample set as a whole conforms to the objective laws of atmospheric radiation transfer and satellite observation, effectively improving the overall quality of the training sample set. This provides a high-quality and reliable sample foundation for subsequent model training, avoiding problems such as overfitting, poor generalization ability, or outputting physically meaningless results due to poor-quality samples.

[0049] Step 210-6: For the samples that have been screened out for quality, perform band response correction based on the characteristics of satellite sensors, and construct a pre-training sample set adapted for cross-satellite learning based on the corrected samples.

[0050] Among them, band response correction is a correction process carried out to address the differences in band response functions of different satellite sensors. By using function mapping, band matching of observation data between different satellite sensors is achieved, so that the observation data of different satellites have unified band response characteristics. Cross-satellite learning is a learning method that uses observation data from different satellite sensors to train the model. This allows the model to learn without relying on the general aerosol radiation characteristics of a specific satellite, thus achieving model adaptation across different satellite sensors. The pre-training sample set is a sample set that has been processed by quality screening and band response correction to meet the needs of cross-satellite learning. It is used for general feature learning in the initial stage of the model and lays the foundation for subsequent targeted training.

[0051] In this embodiment of the disclosure, based on the valid samples after quality screening, standardized band response correction processing is carried out on the samples according to the band response characteristics of different satellite sensors. This achieves unified matching of observation data from different satellites in the band dimension, eliminates the observation data deviation caused by differences in the band response of satellite sensors, and then integrates and organizes all the samples after band response correction to construct a pre-training sample set that can adapt to cross-satellite learning needs and support the model to learn general aerosol radiation characteristics.

[0052] By correcting the band response of the samples after quality screening based on the characteristics of satellite sensors, the bias in observation data caused by differences in band response functions of different satellite sensors can be effectively eliminated. This achieves band standardization matching of multi-source satellite data, allowing observation data from different satellites to have unified feature representation. Based on the corrected samples, a pre-training sample set adapted for cross-satellite learning can be constructed, providing high-quality sample support for model pre-training that covers multiple satellites and is free from band bias. This allows the model to learn the general aerosol radiation characteristics that do not depend on specific satellite sensors during the pre-training stage, effectively improving the model's cross-satellite adaptability. At the same time, it can also enrich the data source of pre-training samples and solve the problem of insufficient samples from a single satellite.

[0053] Step 210-7: Based on the atmospheric and surface characteristics of the target study area, the samples after quality screening are screened, and a fine-tuned sample set adapted to the target satellite is constructed based on the screened samples.

[0054] The target study area is the core application area for aerosol optical thickness inversion, and is the geographical range for which the model is specifically adapted. This area has unique atmospheric composition, aerosol types, and land cover characteristics. The adapted target satellite is a satellite sensor that matches the remote sensing observation needs of the target study area. The model needs to specifically learn the observation characteristics of this satellite to improve the accuracy of regional inversion. The fine-tuning sample set is a sample set selected and constructed based on the atmospheric and surface characteristics of the target study area and adapted to the observation characteristics of the target satellite. It is used for targeted optimization training of the model based on pre-training.

[0055] In this embodiment of the disclosure, based on the valid samples after quality screening, sample screening rules are formulated in combination with the unique atmospheric and surface characteristics of the target study area. According to the rules, samples that can accurately reflect the distribution pattern of atmospheric aerosols and the characteristics of surface cover in the region are selected from the valid samples. At the same time, it is ensured that the selected samples match the observation characteristics of the target satellite. Then, the selected samples are integrated and organized to construct a fine-tuned sample set that can adapt to the observation needs of the target satellite and fit the atmospheric and surface characteristics of the target study area, providing targeted training data support for the satellite-specific fine-tuning stage of the model.

[0056] By screening samples after quality rejection based on atmospheric and surface characteristics of the target study area, the constructed fine-tuned sample set can accurately match the actual atmospheric aerosol and surface characteristics of the target area, effectively improving the matching degree between the samples and the inversion requirements of the region. Based on this sample, a fine-tuned sample set adapted to the target satellite is constructed, allowing the model to specifically learn the observation characteristics of the target satellite and the atmospheric and surface patterns of the target area during the satellite-specific fine-tuning stage. This achieves a precise transition of the model from general feature learning to specific scene adaptation, effectively solving the problem of insufficient adaptability of the model when inverting in a specific area. It significantly improves the accuracy of aerosol optical thickness inversion based on remote sensing data from the target satellite in the target study area, while making the model's inversion results more consistent with the actual atmospheric conditions of the target area, further enhancing the practical value of the model in specific application scenarios.

[0057] Step 210-8: Generate simulated samples of extreme aerosol events through generative adversarial networks, perform physical consistency checks on the simulated samples of extreme aerosol events and screen valid samples, and construct an extreme scenario sample set based on the screened valid samples.

[0058] Among them, Generative Adversarial Networks (GANs) are deep learning models that generate simulated data through adversarial training between a generator and a discriminator. They can learn the feature distribution of real data and generate simulated samples that are highly similar to real data features, effectively supplementing sample data for scarce scenarios. Extreme aerosol event simulation samples are generated by GANs to simulate aerosol optical thickness in extreme aerosol pollution scenarios such as sandstorms and severe smog where measured samples are insufficient. These samples cover various atmospheric and surface features under extreme pollution conditions. Physical consistency verification is based on the physical laws of aerosol radiative transfer and measured data from extreme scenarios. Based on the characteristics, multi-dimensional verification rules are formulated to verify the rationality of the generated simulation samples, ensuring that the simulation samples conform to objective physical laws and actual observation characteristics. Valid samples are extreme aerosol event simulation samples that have passed the physical consistency verification and meet the requirements in terms of feature distribution and physical law conformity, and can truly reflect the aerosol observation characteristics of extreme scenarios. The extreme scenario sample set is a sample set constructed by integrating the valid simulation samples of extreme aerosol events that have passed the physical consistency verification. It is specifically used for extreme scenario enhancement training of the model to improve the model's ability to invert extreme aerosol scenarios.

[0059] In this embodiment of the disclosure, generative adversarial networks can be used to learn the characteristic distribution patterns of real samples of extreme aerosol events, generating simulated samples of extreme aerosol events covering different types of extreme pollution and different surface features. Subsequently, based on the physical laws of aerosol radiative transfer and the observation characteristics of extreme scenarios, scientific verification rules are formulated to conduct a comprehensive physical consistency verification of the generated simulated samples, strictly screening out valid samples that conform to physical laws and are consistent with the observation characteristics of actual extreme scenarios. Finally, all valid samples that pass the verification are integrated and organized to construct an extreme scenario sample set that can accurately reflect the characteristics of extreme aerosol events, providing targeted sample support for the extreme scenario enhancement training of the model.

[0060] By generating simulated samples of extreme aerosol events using generative adversarial networks (GANs), we can effectively compensate for the incomplete coverage and insufficient quantity of measured samples in extreme pollution scenarios. This enriches the types and quantities of samples in extreme scenarios. Performing physical consistency checks on the generated simulated samples can ensure the validity and rationality of the samples from the perspective of physical laws, avoiding interference from unrealistic simulated samples in model training. Constructing an extreme scenario sample set based on the selected effective samples can provide high-quality and highly relevant sample support for the model's extreme scenario enhancement training, allowing the model to learn the characteristic patterns of extreme aerosol events in a targeted manner. This effectively improves the model's accuracy and feature capture capability for aerosol optical thickness inversion in extreme scenarios such as sandstorms and severe smog, and solves the problem of significant inversion bias in non-trained extreme scenarios for traditional data-driven models.

[0061] Step 220: Create a deep learning network model. The deep learning network model updates parameters through a total loss function. The total loss function is a weighted fusion function of data loss and inverse operator constraint loss. Data loss characterizes the deviation between the model's predicted value and the measured value, while inverse operator constraint loss characterizes the degree to which the model's predicted value conforms to the physical laws of aerosol radiative transfer.

[0062] Among them, the data loss L_data adopts the L1 loss, and the formula is as follows: ( To calculate the maximum AOD, y_i represents the ground-measured AOD; the inverse operator constraint loss L_inv is the mean square error loss with band sensitivity weights, calculated by the inverse operator constraint module using the aerosol radiative transfer positive operator to determine the deviation between the simulated atmospheric top reflectivity and the satellite-observed atmospheric top reflectivity, and optimized using the band sensitivity weights; the total loss function is the weighted sum of the data loss and the inverse operator constraint loss multiplied by their respective balance coefficients, updated by gradient backpropagation to update the parameters of each module in the aerosol optical thickness inversion model. Specifically, the balance coefficients can be determined through grid search for weighted coefficient optimization (search range: α∈[0.5,2.0] (step size 0.1), β∈[0.1,0.5] (step size 0.1)), with the final optimization result being α=1.0 and β=0.3. The total loss formula is obtained as: L_total = αL_data + βL_inv, which is automatically differentiated and updated using the PyTorch framework.

[0063] In this embodiment of the disclosure, a dedicated deep learning network model can be built to meet the accuracy and physical rationality requirements of satellite remote sensing AOD inversion. To ensure that the model strictly follows the physical laws of aerosol radiative transfer while learning the statistical laws of data, a total loss function consisting of data loss and inverse operator constraint loss is designed and used as the sole optimization objective for updating model parameters. The data loss is used to measure the deviation between the model's predicted values ​​and the ground-based measured values, while the inverse operator constraint loss is used to measure the deviation between the physical observation signal corresponding to the model's predicted values ​​and the actual satellite observation signal. By assigning appropriate balance coefficients to the two types of losses, weighted fusion is achieved. During the training process, the model aims to minimize this total loss function and continuously updates the parameters of each module of the network through backpropagation, ultimately obtaining a deep learning network model that combines data fitting accuracy and physical law conformity.

[0064] When creating a deep learning network model, the calculation rule for the inverse operator constraint loss (L_inv) is: input model predicts the aerosol optical thickness ( _i), auxiliary physical parameters (θ_i = [aerosol type, surface reflectivity, atmospheric model, solar zenith angle, observed zenith angle]), will Inputting _i and θ_i into the pre-constructed 6S positive operator T, the reflectivity T of the simulated top of the atmosphere (TOA) is calculated. _i, θ_i); Introduce band sensitivity weights w_λ determined based on radiative transfer simulation (e.g., w470=0.35, w650=0.25, w860=0.3, w1610=0.1, and the sum of all band weights is 1), according to the formula L_inv = (1 / N)×Σ[Σ(w_λ×(T_λ( Calculate the mean square error (N=32) using the formula _i,θ_i) - x_λ,i))]², where T_λ( _i,θ_i) represents the simulated top-of-atmosphere (TOA) reflectivity calculated by the 6S radiative transfer positive operator for the λ-th band, and x_λ,i represents the actual TOA reflectivity observed by the satellite (λ represents the four bands). Finally, the gradient of L_inv is used to update the parameters of the feature extraction backbone network through backpropagation, forcing the model to learn a mapping relationship that conforms to the radiative transfer law.

[0065] Step 230: Input the pre-trained samples corrected by band response into the deep learning network model. Based on the pre-set optimizer and learning rate scheduling strategy, train the feature extraction backbone network pre-sequence layer of the deep learning network model with the total loss function as the optimization objective. Update the parameters of the feature extraction backbone network pre-sequence layer through gradient backpropagation of the total loss function, learn the cross-satellite general mapping relationship between aerosols and radiation signals, and complete the cross-satellite pre-training.

[0066] Among them, the feature extraction backbone network pre-sequence layer is the core network layer in the deep learning network that undertakes the basic feature extraction task. It is the key part of the model learning the underlying general features of the data. The cross-satellite general mapping relationship is the common correlation between aerosol physical characteristics and satellite radiation observation signals without relying on the characteristics of specific satellite sensors. Mastering this relationship allows the model to adapt to remote sensing observation data from different satellites. Cross-satellite pre-training is the initial training of the model's feature extraction backbone network pre-sequence layer using cross-satellite adapted samples as training data. The goal is to enable the model to learn the general laws of aerosols and radiation signals and achieve cross-satellite feature transfer capability.

[0067] In this embodiment of the disclosure, pre-trained samples after band response correction can be used as training data input into the deep learning network model. First, an optimized optimizer and learning rate scheduling strategy are preset for model training, clarifying the optimization method and learning rate adjustment rules for updating model parameters. Then, the total loss function, which is a weighted fusion of data loss and inverse operator constraint loss, is used as the sole optimization objective to conduct targeted training on the pre-sequence layer of the feature extraction backbone network of the model. During the training process, the parameters of the pre-sequence layer of the feature extraction backbone network are continuously adjusted and updated through backpropagation of the gradient of the total loss function, allowing the model to gradually learn and master the cross-satellite universal mapping relationship between aerosols and radiation signals until the model training index converges, thus completing the entire cross-satellite pre-training process.

[0068] By inputting pre-trained samples corrected for band response into the model, the differences in band response between different satellite sensors can be eliminated, providing standardized, high-quality training data for cross-satellite pre-training. Pre-set adaptive optimizers and learning rate scheduling strategies make model parameter updates more efficient and learning rate adjustments more in line with training patterns, ensuring stable convergence of the pre-training process. The backbone network pre-sequence layers are extracted with the total loss function as the optimization objective, and parameters are updated through gradient backpropagation. This allows the model to learn the statistical correlation between aerosols and radiation signals while always following the physical laws of aerosol radiation transfer, effectively learning the cross-satellite general mapping relationship between the two.

[0069] Step 240: Freeze the parameters of the pre-trained feature extraction backbone network pre-layers, input the fine-tuning samples into the deep learning network model, train the feature extraction backbone network post-layers and output correction module with the total loss function as the optimization target, and update the parameters of the feature extraction backbone network post-layers and output correction module through backpropagation of the gradient of the total loss function to complete the satellite-specific fine-tuning.

[0070] Among them, the feature extraction backbone network post-sequence layer is a network layer in the deep learning network that inherits the basic features of the preceding layers and is responsible for learning the fine features of the specific scene. It is the key part to realize the model's adaptation learning for specific satellite and regional features. The output correction module is a functional module at the end of the deep learning network used to optimize and adjust the initial output of the model. It can realize the range mapping of output values, adaptive scaling and outlier correction, so that the model output is more in line with the needs of actual application scenarios. Satellite-specific fine-tuning is a targeted model training carried out on the basis of cross-satellite pre-training, based on the sensor characteristics of the target satellite and the features of the corresponding research area. By optimizing the network post-sequence layer and the output correction module, the model can accurately adapt to the remote sensing observation data of the target satellite.

[0071] In this embodiment of the disclosure, the parameters of the deep learning network model that has completed cross-satellite pre-training can be locked first, and the parameters of the preceding layers of its feature extraction backbone network can be frozen to retain the learned cross-satellite general mapping relationship between aerosols and radiation signals. Then, the fine-tuned samples adapted to the target satellite are input into the model. The total loss function, which is a weighted fusion of data loss and inverse operator constraint loss, is still used as the training optimization target. Targeted training is carried out specifically for the subsequent layers of the feature extraction backbone network and the output correction module of the model. During the training process, the relevant parameters of the subsequent layers of the feature extraction backbone network and the output correction module are continuously updated through backpropagation of the gradient of the total loss function, so that the model can gradually learn and adapt to the sensor characteristics of the target satellite and the atmospheric and surface characteristics of the corresponding region until the training index converges, thus completing the entire satellite-specific fine-tuning process.

[0072] By freezing the parameters of the pre-trained feature extraction backbone network's preceding layers, the model's learned cross-satellite general aerosol features can be effectively preserved, preventing subsequent targeted training from damaging the foundation of general features. Fine-tuned samples are input into the model, and the subsequent layers and output correction module are trained with the total loss function as the optimization target. This allows the model to learn the observation characteristics of the target satellite and the atmospheric and surface patterns of the corresponding region while adhering to physical laws, achieving accurate learning from general features to scene-specific features. Gradient backpropagation updates the parameters of the subsequent layers and output correction module, making the model's parameter optimization more aligned with the target satellite's inversion requirements. The completed satellite-specific fine-tuning gives the model accurate adaptability to target satellite remote sensing data, significantly improving the model's aerosol optical thickness inversion accuracy under target satellite observation data. Simultaneously, the optimization of the output correction module makes the model output more consistent with actual application scenarios, further enhancing the practicality and reliability of the model's inversion results.

[0073] Step 250: Input the extreme scene samples and fine-tuned samples into the deep learning network model, unfreeze the post-sequence parameters of the feature extraction backbone network, and add a Dropout regularization layer to a specified layer of the feature extraction backbone network to suppress overfitting. Use the total loss function as the optimization objective, and combine the unfrozen post-sequence parameters of the feature extraction backbone network with the Dropout regularization layer to carry out training. Update the corresponding parameters through backpropagation of the gradient of the total loss function to complete the extreme scene enhancement training of the deep learning network model.

[0074] Among them, the post-sequence parameters of the feature extraction backbone network are the network parameters of the later layers of the feature extraction backbone network in the deep learning network. They are responsible for learning the fine features of the data and scene-specific features, and are key parameters for enabling the model to adapt to specific scenes. The Dropout regularization layer is a network layer in deep learning used to suppress model overfitting. By randomly discarding the output of some neurons during training, it avoids the model from over-relying on local features and improves the model's generalization ability. Extreme scene enhancement training is a model reinforcement training carried out by fusing extreme scene samples and fine-tuned samples on the basis of satellite-specific fine-tuning. It specifically optimizes the model's feature capture ability for extreme aerosol scenes and improves the model's inversion accuracy in extreme scenes.

[0075] In this embodiment, extreme scenario samples and fine-tuned samples can be fused and used as training data input to complete a satellite-specific fine-tuning deep learning network model. First, the parameters of the feature extraction backbone network of the model are adjusted, and some of the subsequent parameters are unfrozen to retain the target satellite-specific features already learned by the model and support the adaptive learning of extreme scenario features. At the same time, Dropout regularization layers are added to designated layers of the feature extraction backbone network to avoid overfitting problems in extreme scenario sample training. The total loss function, which is a weighted fusion of data loss and inverse operator constraint loss, is still used as the optimization target for training. The model is trained by combining the unfrozen subsequent parameters and the Dropout regularization layer. During the training process, the unfrozen subsequent parameters of the feature extraction backbone network and related network layer parameters are continuously updated through backpropagation of the gradient of the total loss function, allowing the model to gradually learn and capture the feature patterns of extreme aerosol scenarios until the training indicators converge, thus completing the extreme scenario enhancement training of the entire deep learning network model.

[0076] By mixing extreme scenario samples with fine-tuned samples and inputting them into the model, the model learns extreme scenario features based on the target satellite's specific features, achieving a fusion learning of features from both normal and extreme scenarios. Unfreezing some features and extracting subsequent parameters from the backbone network not only preserves the target satellite observation characteristics already adapted to the model but also reserves parameter optimization space for learning extreme scenario features. Adding Dropout regularization layers to specified layers can effectively suppress overfitting problems that may occur during extreme scenario training, improving the model's generalization ability to extreme scenarios. Using the total loss function as the optimization objective and updating the corresponding parameters through gradient backpropagation, the model always follows the physical laws of aerosol radiative transfer while learning extreme scenario features, avoiding the output of physically meaningless outliers. The completed extreme scenario enhancement training enables the model to accurately capture the characteristic patterns of extreme aerosol events, significantly improving the model's aerosol optical thickness inversion accuracy under extreme pollution scenarios.

[0077] Step 260: In each stage of pre-training, satellite-specific fine-tuning, and extreme scenario enhancement training, monitor the inversion accuracy index of the deep learning network model in real time, and monitor the changing trends of the total loss function, data loss, and inversion operator constraint loss. If the inversion accuracy index does not decrease for a consecutive preset number of rounds and the total loss function, data loss, and inversion operator constraint loss all tend to stabilize without decreasing, then trigger the early stop strategy to terminate the training of the corresponding stage. After pre-training, satellite-specific fine-tuning, and extreme scenario enhancement training are all completed and each stage of training has converged, the trained aerosol optical thickness inversion model is obtained.

[0078] Among them, the inversion accuracy index is a quantitative indicator used to measure the degree of fit between the model's predicted aerosol optical thickness and the true value, and is the core basis for evaluating the model training effect; the change trend is the numerical change trend of the total loss function, data loss, inverse operator constraint loss and inversion accuracy index as the number of training rounds progresses during the model training process, reflecting the learning and convergence status of the model; the early stopping strategy is a training termination rule formulated to avoid model overfitting and improve training efficiency. When the model training index reaches the preset convergence condition, the corresponding stage of the training process is automatically terminated; training convergence is the state in which the values ​​of the inversion accuracy index and various loss functions no longer change significantly with the number of training rounds, tend to be stable and reach the preset training target during the model training process.

[0079] In this embodiment of the disclosure, the inversion accuracy index of the model can be monitored in real time throughout the three stages of pre-training, satellite-specific fine-tuning, and extreme scenario enhancement training of the deep learning network model. At the same time, the numerical change trends of the total loss function, data loss, and inverse operator constraint loss are monitored simultaneously. A unified training judgment rule is set. When it is detected that the inversion accuracy index no longer decreases for a preset number of consecutive rounds, and the values ​​of the total loss function, data loss, and inverse operator constraint loss all tend to be stable and do not decrease, the early stop strategy is immediately triggered and the model training of the current stage is terminated. After the three stages of pre-training, satellite-specific fine-tuning, and extreme scenario enhancement training are completed and each reaches the training convergence state, the aerosol optical thickness inversion model that has completed the entire training process can be obtained.

[0080] By monitoring the inversion accuracy metrics and the changing trends of various loss functions in real time at each stage of model training, the learning status and training effect of the model can be accurately grasped throughout the process. Convergence nodes of model training can be identified in a timely manner, and early stopping strategies can be triggered under preset conditions to terminate the corresponding training stage. This can effectively avoid overfitting caused by overtraining, ensure the model's generalization ability, and significantly improve the overall model training efficiency, reducing meaningless training time. At the same time, it is required that the training at each stage converges before obtaining the final inversion model, ensuring that the training effect of the model in each stage of pre-training, dedicated fine-tuning, and extreme enhancement reaches the preset target. This allows the model to master cross-satellite general features, adapt to the characteristics of the target satellite, and accurately capture extreme scene features. The final aerosol optical thickness inversion model has stable inversion accuracy and good generalization ability, and can output accurate and reliable aerosol optical thickness inversion results in various scenarios.

[0081] Specifically, when using a cross-satellite transfer learning strategy to conduct phased training, such as Figure 4As shown, the pre-trained sample set constructed in the early stage can be input into the deep learning network model with the completed architecture design. Cross-satellite pre-training can be carried out according to the preset parameters. The training indicators are monitored until convergence to obtain the pre-trained feature extraction backbone network. Then, the fine-tuning sample set is input into the pre-trained model, the parameters of the preceding layers of the backbone network are frozen, and the training parameters are set according to the target satellite adaptation requirements. Satellite-specific fine-tuning is carried out until the indicators converge, and a preliminary training model adapted to the target satellite is obtained. Subsequently, the extreme scene sample set and the fine-tuning sample set are mixed and input into the preliminary training model. The parameters of the subsequent parts of the backbone network are unfrozen, a Dropout layer is added to suppress overfitting, and targeted training parameters are set. Extreme scene enhancement training is carried out. The core indicators such as total loss and data loss are monitored throughout the process until the model converges. Finally, an aerosol optical thickness inversion model that has completed cross-satellite pre-training, satellite-specific fine-tuning, and extreme scene enhancement training is obtained.

[0082] Step 270: Perform preprocessing operations such as radiometric calibration, cloud detection, and geometric correction on the raw satellite remote sensing data. Input the preprocessed raw satellite remote sensing data into the trained aerosol optical thickness inversion model for inference calculation. Perform post-processing and quality classification on the model output results to obtain aerosol optical thickness inversion results used to assist atmospheric correction.

[0083] The aerosol optical thickness inversion model includes a feature extraction backbone network, an inverse operator constraint module, and an output correction module. The feature extraction backbone network consists of the first 49 convolutional layers based on a modified ResNet-50, containing 1×1, 3×3, and 5×5 multi-scale convolutional kernel layers and residual blocks. It is responsible for extracting multi-scale aerosol and radiation signal features from satellite remote sensing data, providing core feature support for subsequent inversion. The inverse operator constraint module incorporates aerosol radiation transfer positive operator and a band sensitivity weight calculation unit, used to compare the model's prediction results with the physical laws of aerosol radiation transfer, constraining the model output to have physical rationality. The output correction module includes a Sigmoid activation function layer, a dynamic scaling coefficient library, and an anomaly filtering logic layer, used to perform interval mapping, adaptive scaling, and outlier correction on the initial model output, making the output results more consistent with practical application scenarios. The radiative transfer positive operator is a forward mapping operator built based on the radiative transfer model. It can quantitatively calculate the reflectivity of the top of the atmosphere observed by satellite sensors, from physical parameters such as aerosols, atmosphere, surface, and geometry. It is the core carrier of physical constraints. The band sensitivity weight calculation unit is used to calculate the sensitivity of different satellite bands to aerosol changes, providing a basis for weight allocation for the inverse operator constraint module and improving the accuracy of physical constraints. The sigmoid activation function layer maps the model output to the 0 to 1 interval through nonlinear transformation, ensuring the numerical rationality of the aerosol optical thickness inversion results. The dynamic scaling coefficient library stores the output scaling coefficients under different scenarios, which are used to adaptively adjust the model output and improve the adaptability of the inversion results under different scenarios. The anomaly filtering logic layer identifies and filters outliers in the model output according to preset rules, improving the reliability of the inversion results.

[0084] For embodiments of this disclosure, when inputting preprocessed raw satellite remote sensing data into a trained aerosol optical thickness inversion model for inference calculation, step 270 may include: Step 270-1: Based on the preprocessed satellite remote sensing raw data, extract basic reflectivity features, band combination features, atmospheric geometric auxiliary features and radiation-derived features, fuse and construct multi-dimensional input features adapted to the aerosol optical thickness inversion model, and input the multi-dimensional input features into the feature extraction backbone network of the aerosol optical thickness inversion model.

[0085] Among them, the basic reflectivity feature is the atmospheric top reflectivity feature directly obtained from preprocessed satellite remote sensing data, reflecting the basic information of surface and atmospheric radiation signals received by satellite sensors; the band combination feature is a combination feature formed by performing ratio, difference, and normalization operations on reflectivity data from different satellite observation bands, used to enhance the distinction between aerosols and the surface; the atmospheric geometric auxiliary feature is an auxiliary feature composed of satellite observation geometric parameters, including solar zenith angle, observation zenith angle, relative azimuth angle, etc., which are important geometric constraints on the aerosol radiation transmission process; the radiation-derived feature is a derived feature further calculated based on the basic radiation observation data, used to characterize the radiation variation laws such as atmospheric scattering and absorption, and improve the model's ability to perceive aerosol signals; the multidimensional input feature is a comprehensive input feature constructed by fusing the basic reflectivity feature, band combination feature, atmospheric geometric auxiliary feature, and radiation-derived feature, which has the ability to represent information in multiple dimensions and scales. The multidimensional input features can be 15-dimensional, including: basic reflectance features (four-band TOA reflectance), band combination features (470nm / 650nm, 860nm / 1610nm ratio), auxiliary features (solar zenith angle, observed zenith angle, relative azimuth angle, atmospheric precipitable water), and derived features (reflectance slope, atmospheric correction index).

[0086] In this embodiment of the disclosure, preprocessed satellite remote sensing raw data can be used as a basis to extract basic reflectivity features, band combination features, atmospheric geometric auxiliary features, and radiation-derived features. These multiple features are effectively fused to construct multidimensional input features that are compatible with the aerosol optical thickness inversion model. The constructed multidimensional input features are then input into the feature extraction backbone network of the aerosol optical thickness inversion model, providing complete and rich information support for the subsequent feature learning and inference calculation of the model.

[0087] By extracting various types of features from preprocessed satellite remote sensing data and fusing them to construct multidimensional input features, we can fully explore the multidimensional information related to aerosols in satellite observation data, enrich the representation dimensions of input data, and improve the model's adaptability to complex surface and atmospheric scenes. Inputting multidimensional input features into the feature extraction backbone network can provide the model with more comprehensive and accurate input information, effectively enhancing the model's ability to extract and learn aerosol signals, improving the model's feature capture accuracy and generalization ability under different surface and atmospheric conditions, and thus ensuring the accuracy and stability of subsequent aerosol optical thickness inversion results.

[0088] Step 270-2: The multidimensional input features are reduced and fused by the 1×1 convolutional kernel layer in the feature extraction backbone network, the local radiation features are captured by the 3×3 convolutional kernel layer, and the global aerosol distribution features are captured by the 5×5 convolutional kernel layer. After the feature depth extraction and gradient transfer are completed by the residual block, the deep fused features are output.

[0089] The network consists of several layers: a 1×1 convolutional kernel layer, a 3×3 convolutional kernel layer, and a 5×5 convolutional kernel layer. The 1×1 kernel layer performs operations on the input features using a 1×1 kernel, primarily compressing the feature dimension and fusing features from different channels, simplifying computation while preserving key feature information. The 3×3 kernel layer serves as the local feature capture module in the feature extraction backbone network, performing convolution operations on the dimensionality-reduced and fused features using a 3×3 kernel, focusing on capturing radiation signal features within local regions and uncovering local aerosol distribution patterns. Finally, the 5×5 kernel layer serves as the global feature capture module in the feature extraction backbone network, performing operations on the features using a 5×5 kernel. It covers a wider feature region to capture the global aerosol distribution characteristics and overall radiation variation patterns; the residual block is a deep feature enhancement module in the feature extraction backbone network, which transmits gradients through skip connections, effectively alleviating the gradient vanishing problem in deep network training, while realizing deep feature extraction and enhancement, and improving feature representation capabilities; the deep fusion feature is a comprehensive feature that integrates local radiation features, global aerosol distribution features and multi-dimensional input features after being processed by each module of the feature extraction backbone network, and has rich information representation capabilities, providing core feature support for subsequent model inference.

[0090] In this embodiment of the disclosure, the multidimensional input features can be used as a basis for the feature extraction backbone network. First, the multidimensional input features are reduced in dimensionality by a 1×1 convolutional kernel layer in the network to compress the feature dimensions and fuse feature information from different channels. Then, the processed features are input into a 3×3 convolutional kernel layer to capture radiation signal features in local areas. Subsequently, the local features are input into a 5×5 convolutional kernel layer to further capture aerosol distribution features in the global range. Finally, the residual block is used to achieve deep feature extraction and effective gradient transfer, alleviating the gradient vanishing problem in deep networks. After the sequential processing of the above modules, the final output is a deep fusion feature with both local and global features and rich information, providing high-quality feature support for subsequent inverse operator constraints and output correction.

[0091] By using 1×1 convolutional kernel layers to reduce and fuse multidimensional input features, computational complexity can be effectively simplified. Simultaneously, key information from multidimensional input features is integrated, avoiding redundant feature interference. 3×3 and 5×5 convolutional kernel layers capture local radiation features and global aerosol distribution features respectively, achieving comprehensive feature capture from local to global perspectives and ensuring the completeness of feature information. The application of residual blocks effectively alleviates the gradient vanishing problem in deep network training, ensuring the effectiveness of feature depth extraction and the stability of gradient propagation. The final deep fused feature output combines local and global information, possessing powerful feature representation capabilities. This effectively improves the feature extraction accuracy and efficiency of the feature extraction backbone network, providing high-quality core feature support for subsequent inverse operator constraint verification and output correction, thereby enhancing the inversion accuracy and generalization ability of the entire aerosol optical thickness inversion model.

[0092] Step 270-3: The deep fusion feature input-output correction module first maps the feature output to the [0,1] interval through the Sigmoid activation function layer. Then, based on the land surface type and atmospheric mode corresponding to the preprocessed satellite remote sensing raw data, it calls the dynamic scaling coefficient library to complete adaptive linear scaling. After the anomaly filtering logic layer corrects the output value that exceeds the reasonable range of the scene, the preliminary aerosol optical thickness prediction value is obtained.

[0093] Among them, the land surface type refers to the land cover category of the area corresponding to the preprocessed satellite remote sensing raw data, which is an important reference for determining the reasonable range of aerosol optical thickness and scaling factor; the atmospheric model is a standardized type that characterizes the atmospheric vertical structure and radiative transfer characteristics of a specific area, providing a core reference for calling the dynamic scaling factor; the dynamic scaling factor library is a coefficient storage unit built into the output correction module, storing scaling factors under different combinations of land surface type and atmospheric model, used to achieve adaptive adjustment of the output value; adaptive linear scaling is based on the land surface type and atmospheric model corresponding to the current data, calling the corresponding coefficient in the dynamic scaling factor library to linearly adjust the mapped output value, so that the output value fits the actual range of the specific scene; the preliminary aerosol optical thickness prediction value is the aerosol optical thickness prediction result obtained after sequential processing by each unit of the output correction module, which is the basis for subsequent post-processing and quality classification.

[0094] In this embodiment of the disclosure, the deep fusion feature output from the feature extraction backbone network can be input to the output correction module. First, the output of the deep fusion feature is nonlinearly transformed by the Sigmoid activation function layer in the module to map it to a fixed interval. Then, combined with the land surface type and atmospheric model corresponding to the preprocessed satellite remote sensing raw data, the appropriate scaling factor is called from the dynamic scaling factor library to adaptively linearly scale the mapped output value so that the output value fits the actual reasonable range of the current scene. Finally, the scaling output value is verified by the anomaly filtering logic layer to correct abnormal outputs that exceed the reasonable range of the scene. After the above series of processing, a preliminary aerosol optical thickness prediction value that conforms to the actual scene and is numerically reasonable is obtained.

[0095] By mapping the feature output to a fixed interval through the Sigmoid activation function layer, the numerical rationality of the initial output values ​​can be effectively guaranteed, avoiding initial outputs that exceed the theoretical range. Based on the surface type and atmospheric model, a dynamic scaling coefficient library is used to complete adaptive linear scaling, enabling the predicted values ​​to accurately adapt to the actual characteristics of different scenarios and improve the scenario adaptability of the prediction results. The anomaly filtering logic layer corrects abnormal output values, further eliminating unreasonable prediction results and ensuring the reliability of the initial prediction values. The synergistic effect of a series of processing steps in the output correction module can effectively correct model output deviations, making the initial aerosol optical thickness prediction values ​​both consistent with numerical laws and in line with the needs of actual scenarios. This provides high-quality basic data for subsequent post-processing and quality grading, thereby improving the inversion accuracy and reliability of the entire aerosol optical thickness inversion model.

[0096] Step 270-4: Extract the auxiliary physical parameters corresponding to the preprocessed satellite remote sensing raw data, input the preliminary aerosol optical thickness prediction value and the auxiliary physical parameters into the inverse operator constraint module, complete the physical consistency verification through the aerosol radiation transmission positive operator, optimize the verification results by combining the band sensitivity weight calculation unit, and output the aerosol optical thickness inference calculation result that conforms to the physical law of aerosol radiation transmission.

[0097] Among them, the physical consistency verification is a process of combining the preliminary predicted value with auxiliary physical parameters through the aerosol radiation transmission positive operator, comparing the consistency of the calculation result with the satellite observation signal, and determining whether the preliminary predicted value conforms to the physical laws of aerosol radiation transmission. The aerosol optical thickness inference calculation result is the aerosol optical thickness prediction result that conforms to the physical laws of aerosol radiation transmission after the physical consistency verification and optimization are completed by the inverse operator constraint module. It is the core foundation of the final output of the model.

[0098] In this embodiment of the present disclosure, auxiliary physical parameters can be extracted from the preprocessed satellite remote sensing raw data. Then, the preliminary aerosol optical thickness prediction value obtained by the output correction module and the extracted auxiliary physical parameters are input together into the inverse operator constraint module. The aerosol radiation transfer positive operator built into the module combines the two to complete the physical consistency verification, and determines whether the preliminary prediction value conforms to the objective physical laws of aerosol radiation transfer. At the same time, the sensitivity weight of different satellite bands is calculated by the band sensitivity weight calculation unit. The physical consistency verification result is optimized and adjusted using this weight, and finally, the aerosol optical thickness inference calculation result that conforms to the physical laws of aerosol radiation transfer and has both accuracy and rationality is output.

[0099] By extracting auxiliary physical parameters corresponding to the preprocessed satellite remote sensing raw data, comprehensive physical support can be provided for the physical consistency verification of the inverse operator constraint module, ensuring that the verification process conforms to the actual atmospheric and surface physical characteristics. The preliminary predicted values ​​and auxiliary physical parameters are input into the inverse operator constraint module, and the physical consistency verification is completed in conjunction with the aerosol radiative transfer positive operator. This can constrain the physical rationality of the model output from the root, avoiding abnormal prediction results without physical meaning. The optimization of the verification results by the band sensitivity weight calculation unit can further improve the accuracy of physical constraints, making the verification results more in line with the actual characteristics of satellite observation. The final output aerosol optical thickness inference calculation results strictly follow the physical laws of aerosol radiative transfer, which can effectively improve the accuracy and reliability of the model inversion results, provide high-quality core data for subsequent post-processing and quality classification, and further enhance the practicality and generalization ability of the entire inversion model.

[0100] Accordingly, when post-processing and quality grading the model output to obtain aerosol optical thickness inversion results for assisting atmospheric correction, step 270 of the embodiment may include: Step 270-5: Apply Gaussian filtering to the aerosol optical thickness inference calculation results output by the aerosol optical thickness inversion model to perform spatial smoothing, and obtain the smoothed aerosol optical thickness inversion results.

[0101] Among them, Gaussian filtering is a smoothing method based on Gaussian function. By performing a weighted average operation on the surrounding pixel values, it smooths the data, effectively eliminating isolated outliers and improving the spatial continuity of the data. Spatial smoothing is a spatial dimension optimization process for the aerosol optical thickness inference calculation results. It eliminates spatial noise and isolated outliers in the results through specific filtering methods, making the inversion results more consistent in spatial distribution and more in line with the actual aerosol distribution law. The smoothed aerosol optical thickness inversion result is an aerosol optical thickness inversion result that has been spatially smoothed by Gaussian filtering, eliminating spatial noise and isolated outliers, and improving spatial continuity and data reliability.

[0102] In this embodiment of the disclosure, the spatial smoothing process can be carried out by Gaussian filtering on the aerosol optical thickness inference calculation results output by the aerosol optical thickness inversion model. By performing a weighted average operation on each pixel and its surrounding pixel values ​​in the inference calculation results using a Gaussian function, isolated outliers and spatial noise in the results are eliminated, the spatial distribution continuity of the inversion results is optimized, and finally a smoothed aerosol optical thickness inversion result with coherent spatial distribution and more reliable data is obtained.

[0103] By applying Gaussian filtering to spatially smooth the aerosol optical thickness inference calculation results, isolated outliers and spatial noise in the inference calculation results can be effectively eliminated, solving the problem of inconsistent spatial distribution of the inversion results, improving the spatial continuity and data uniformity of the inversion results, while preserving the true spatial distribution characteristics of aerosol optical thickness and avoiding feature distortion caused by smoothing. This makes the smoothed inversion results more consistent with the actual spatial distribution of aerosols, which can further improve the reliability and practicality of aerosol optical thickness inversion results, and provide higher quality basic data for subsequent quality classification and auxiliary atmospheric correction.

[0104] Step 270-6: Derive the confidence level of the aerosol optical thickness inversion result based on the inverse operator constraint loss value output by the inverse operator constraint module. The smaller the inverse operator constraint loss value, the higher the confidence level of the corresponding inversion result.

[0105] Among them, the inverse operator constraint loss value is output by the inverse operator constraint module and is used to quantify the degree of fit between the aerosol optical thickness inference calculation results and the physical laws of aerosol radiative transfer. It is the core indicator for measuring the physical rationality of the inversion results. The confidence level is a quantitative indicator used to characterize the reliability of the aerosol optical thickness inversion results. It reflects the degree of agreement between the inversion results and the actual atmospheric aerosol conditions. The higher the confidence level, the stronger the reliability of the inversion results. The aerosol optical thickness inversion results are the aerosol optical thickness prediction results that conform to physical laws and have good spatial continuity after being verified and optimized by the inverse operator constraint module and smoothed by Gaussian filtering.

[0106] In this embodiment of the disclosure, the inverse operator constraint loss value output by the inverse operator constraint module can be used as the core basis. By using preset derivation rules and combining the magnitude of the inverse operator constraint loss value, the confidence level corresponding to the aerosol optical thickness inversion result is derived. Following the principle that the smaller the inverse operator constraint loss value, the higher the degree of conformity between the inversion result and the physical law, the confidence level of the inversion result is determined, and finally the confidence level corresponding to each inversion result is obtained, providing a core reference basis for subsequent quality classification.

[0107] By deriving the confidence level of the inversion results based on the constraint loss value of the inverse operator, a quantitative correlation between the physical rationality and reliability of the inversion results can be established. This provides a clear basis for judging the reliability of the inversion results. The rule that the smaller the constraint loss value of the inverse operator, the higher the confidence level, can accurately reflect the degree of agreement between the inversion results and the actual atmospheric conditions. It can effectively distinguish the reliability levels of different inversion results, providing scientific and accurate core support for subsequent quality classification. At the same time, it allows users to intuitively judge the usability of the inversion results, further improving the practicality and credibility of aerosol optical thickness inversion results, and also providing an important reference for the application and optimization of subsequent inversion results.

[0108] Step 270-7: Perform multi-level quality classification on the aerosol optical thickness inversion results according to the confidence level, and mark the aerosol optical thickness inversion result regions of each level.

[0109] The multi-level quality grading is a classification method that divides the aerosol optical thickness inversion results into multiple reliability levels based on different confidence intervals, used to distinguish the reliability of the inversion results; the level labeling is the processing operation that marks the regions where the aerosol optical thickness inversion results of different levels are located after multi-level quality grading, making it easy to distinguish them intuitively and for subsequent applications; the confidence level is a quantitative indicator that characterizes the reliability of the aerosol optical thickness inversion results and is the core judgment basis for multi-level quality grading. The higher the value, the stronger the reliability of the inversion results; the aerosol optical thickness inversion results are the aerosol optical thickness prediction results that conform to the physical laws of aerosol radiative transfer and have good spatial continuity after a series of optimization processes.

[0110] In this embodiment of the disclosure, the confidence level of the derived aerosol optical thickness inversion result is used as the core judgment criterion. Different confidence level intervals are preset to correspond to different quality levels. Based on this correspondence, the aerosol optical thickness inversion result is classified into multiple quality levels, and the inversion result is divided into different reliability levels. After the classification is completed, the aerosol optical thickness inversion result area corresponding to each level is clearly marked to ensure that the inversion results of different levels can be intuitively distinguished, providing a clear reliability reference for the subsequent application of the inversion result.

[0111] By classifying the inversion results into multiple quality levels based on confidence levels, the reliability of the inversion results can be accurately distinguished, allowing for clear grading of inversion results with different levels of reliability. The grading labels make the regions of different grades of inversion results easily identifiable, facilitating the selection of appropriate grades of inversion results according to application needs. This effectively improves the application relevance and convenience of the inversion results, while also providing a clear direction for subsequent optimization of the inversion results. This further enhances the practicality and application value of aerosol optical thickness inversion results, ensuring that the inversion results can be better adapted to practical application scenarios such as assisted atmospheric correction.

[0112] Step 270-8: Convert the quality-graded aerosol optical thickness inversion results with completed grade markings into GeoTIFF format in the WGS-84 coordinate system, and output the corresponding quality label map and inversion accuracy report to obtain aerosol optical thickness inversion results for auxiliary atmospheric correction.

[0113] Among them, the WGS-84 coordinate system is a globally universal geographic coordinate system used to unify geospatial location information and ensure that the spatial positioning of the inversion results is accurate and universal; the GeoTIFF format is an image file format that supports geospatial information and can store the numerical information and spatial coordinate information of the inversion results, which facilitates subsequent geographic data processing and application; the quality label map is an image that intuitively presents the distribution of quality levels in different regions of the aerosol optical thickness inversion results, clearly showing the spatial distribution of different reliability levels; the inversion accuracy report is a report that records the accuracy indicators, error analysis and quality level statistics of the aerosol optical thickness inversion results, which is used to explain the accuracy level and reliability of the inversion results.

[0114] In this embodiment of the disclosure, the quality-graded aerosol optical thickness inversion results with completed grade marking can be converted to a GeoTIFF format in the WGS-84 coordinate system to ensure that the inversion results have a unified spatial positioning and a standard file format. At the same time, the corresponding quality identification map is output simultaneously to intuitively present the quality grade distribution of each region, as well as an inversion accuracy report containing accuracy indicators, error analysis, etc., to comprehensively explain the accuracy and quality of the inversion results, and finally obtain aerosol optical thickness inversion results that can be directly used to assist atmospheric correction.

[0115] Converting the inversion results to GeoTIFF format in the WGS-84 coordinate system ensures accurate spatial positioning and global applicability, facilitating compatibility with other geospatial data and meeting the format and coordinate system requirements of subsequent atmospheric correction and other practical applications. The output quality label map visually displays the quality level distribution of the inversion results, facilitating quick assessment of the reliability of inversion results in different regions. The inversion accuracy report clearly presents the accuracy level and error of the inversion results, providing a comprehensive accuracy reference for the application of the inversion results. The synergistic effect of these three elements makes the final aerosol optical thickness inversion results standardized, visualized, and traceable, significantly improving the practicality and adaptability of the inversion results and efficiently supporting subsequent atmospheric correction work.

[0116] In specific application scenarios, as a preferred approach, the process of verifying the performance of the aerosol optical thickness inversion results may also be included, such as... Figure 5As shown, the IOR-ResNet model (aerosol optical thickness inversion model), independent validation data, and comparison algorithm results can be used as inputs to construct an independent validation dataset containing independent validation sets, extreme scenario validation sets, and aerosol type validation sets. Based on this dataset, core indicators, extreme scenario indicators, and aerosol type-specific indicators are calculated. Subsequently, comparative experiments such as internal comparison, external comparison, efficiency comparison, and stability verification are carried out. The evaluation indicators of the inversion results are compared with the results of traditional aerosol optical thickness inversion algorithms. At the same time, long-term stability verification of the aerosol optical thickness inversion model is carried out. If all indicators of the inversion results are better than the corresponding indicators of the traditional algorithm, and the fluctuation range of the indicators of the long-term stability verification of the model is within the preset threshold, then the aerosol optical thickness inversion result is determined to have passed the performance verification.

[0117] Accordingly, the implementation steps may include: constructing an independent validation dataset, which includes spatiotemporal matching test samples not involved in model training, extreme scenario validation samples, and aerosol type validation samples; calculating the core evaluation indicators, extreme scenario-specific evaluation indicators, and aerosol type-specific evaluation indicators for the aerosol optical thickness inversion results, whereby the core evaluation indicators include the coefficient of determination, mean deviation, and root mean square error; comparing the aerosol optical thickness inversion results with the results of traditional aerosol optical thickness inversion algorithms based on the core evaluation indicators, extreme scenario-specific evaluation indicators, and aerosol type-specific evaluation indicators, and simultaneously performing long-term stability verification on the aerosol optical thickness inversion model. If all indicators of the aerosol optical thickness inversion results are superior to the corresponding indicators of the traditional aerosol optical thickness inversion algorithm, and the fluctuation range of the long-term stability verification indicators of the aerosol optical thickness inversion model is within a preset threshold, then the aerosol optical thickness inversion results are determined to have passed performance verification. The preset threshold is a standard numerical range used to determine the long-term stability of the model; if the stability indicators fluctuate within this range, it indicates that the model is stable in long-term operation.

[0118] In summary, the deep learning-based method for AOD inversion-assisted atmospheric correction provided in this application compensates for insufficient spatiotemporal matching sample coverage by constructing a multi-type training sample set that integrates measured and physical simulation data and performing quality screening and stratification. This effectively addresses the problem of large inversion bias in extreme and non-training scenarios caused by data-driven models. Furthermore, by creating a deep learning network model that integrates data loss and inverse operator constraint loss into a total loss function, the physical laws of aerosol radiative transfer are deeply integrated into the model parameter update process. This avoids the dependence of physical-driven algorithms on specific surface assumptions and improves their accuracy on heterogeneous and bright surfaces. The descent problem can fundamentally constrain the model output to conform to physical laws, avoiding the generation of outliers without physical meaning. At the same time, it can overcome the shortcomings of simple superposition in traditional hybrid methods, effectively solve the multiple solutions of underdetermined inverse problems, and avoid the problem of data transformation error accumulation. By carrying out phased training through cross-satellite transfer learning strategies, the generalization ability and adaptability to extreme scenarios of the model can be further improved. Combined with satellite remote sensing data preprocessing and inversion result postprocessing and quality classification, high-precision and high-generalization inversion of aerosol optical thickness under complex surface and atmospheric scenarios can be achieved, which greatly improves the actual effect of AOD inversion results in assisting atmospheric correction.

[0119] Furthermore, as Figure 1 and Figure 2 The specific implementation of the method shown in this embodiment provides a device for AOD-assisted atmospheric correction based on deep learning inversion, such as... Figure 6 As shown, the device may include: a construction module 61, a creation module 62, a training module 63, and a processing module 64.

[0120] Module 61 can be used to construct a multi-type training sample set that integrates measured data and physical simulation data, perform quality screening and stratification on the multi-type training sample set, and obtain pre-training samples, fine-tuning samples and extreme scenario samples that are adapted to satellite remote sensing data. Module 62 can be used to create deep learning network models. The deep learning network models update parameters through the total loss function, which is a weighted fusion function of data loss and inverse operator constraint loss. Data loss characterizes the deviation between the model's predicted value and the measured value, while inverse operator constraint loss characterizes the degree to which the model's predicted value conforms to the physical laws of aerosol radiative transfer. Training module 63 can be used to perform phased training of deep learning network models by adopting cross-satellite transfer learning strategies, including pre-training, satellite-specific fine-tuning, and extreme scenario enhancement training, to obtain a trained aerosol optical thickness inversion model. The processing module 64 can be used to perform preprocessing operations such as radiometric calibration, cloud detection, and geometric correction on the raw satellite remote sensing data. The preprocessed raw satellite remote sensing data is input into the trained aerosol optical thickness inversion model for inference calculation. The model output results are post-processed and quality graded to obtain aerosol optical thickness inversion results used to assist atmospheric correction.

[0121] In some embodiments of this application, the construction module 61 is specifically used to collect satellite remote sensing Level 1B data, ground-measured AOD data, and atmospheric auxiliary data; based on the atmospheric auxiliary data, preprocessing operations such as radiometric calibration, cloud detection, and geometric correction are performed on the satellite remote sensing Level 1B data; based on the preprocessed satellite remote sensing Level 1B data, outlier removal and spatiotemporal matching processing are performed on the ground-measured AOD data to obtain spatiotemporally matched measured samples; an aerosol radiative transfer positive operator is constructed based on the 6S radiative transfer model, using atmospheric model parameters in the atmospheric auxiliary data as a reference, and multiple atmospheric models, multiple aerosol models, and the surface reflectance calculated using the surface endmember decomposition method are integrated into the aerosol radiative transfer positive operator; initial physical simulation data is generated through operator operations; random noise that conforms to satellite observation characteristics is added to the initial physical simulation data to complete the simulation data enhancement and obtain the final physical simulation data; the spatiotemporally matched measured samples are fused with the physical simulation data to obtain an initial multi-type training sample set.

[0122] In some embodiments of this application, the construction module 61 can also be used to perform quality screening on multi-type training sample sets. The quality screening includes removing invalid samples whose AOD values ​​and satellite top reflectivity exceed a preset range, filtering parameter combination samples that have no physical meaning in the physical simulation process; for the samples after quality screening, performing band response correction based on satellite sensor characteristics, and constructing a pre-training sample set adapted for cross-satellite learning based on the corrected samples; screening the samples after quality screening according to the atmospheric and surface characteristics of the target study area, and constructing a fine-tuned sample set adapted to the target satellite based on the screened samples; generating extreme aerosol event simulation samples through a generative adversarial network, performing physical consistency verification on the extreme aerosol event simulation samples and screening valid samples, and constructing an extreme scenario sample set based on the screened valid samples.

[0123] In some embodiments of this application, the inverse operator constraint loss is a mean square error loss with band sensitivity weights, calculated by the inverse operator constraint module using the aerosol radiative transfer positive operator to determine the deviation between the simulated atmospheric top reflectivity and the satellite-observed atmospheric top reflectivity, and optimized using the band sensitivity weights. The total loss function is a weighted sum of the data loss and the inverse operator constraint loss multiplied by their respective balance coefficients, and is used to update the parameters of each module in the aerosol optical thickness inversion model through gradient backpropagation. Correspondingly, the training module 63 can be specifically used to input pre-trained samples corrected for band response into the deep learning model. The learning network model, based on a pre-set optimizer and learning rate scheduling strategy, trains the pre-sequence layers of the feature extraction backbone network of the deep learning network model with the total loss function as the optimization objective. The parameters of the pre-sequence layers are updated through backpropagation of the gradient of the total loss function, learning the cross-satellite universal mapping relationship between aerosols and radiation signals, thus completing cross-satellite pre-training. The pre-trained parameters of the pre-sequence layers are frozen, and fine-tuned samples are input into the deep learning network model. The post-sequence layers of the feature extraction backbone network and the output correction module are trained with the total loss function as the optimization objective. The parameters are then updated through backpropagation of the gradient of the total loss function. Backpropagation is used to update the parameters of the subsequent layers of the feature extraction backbone network and the parameters of the output correction module, completing satellite-specific fine-tuning. Extreme scenario samples and fine-tuned samples are mixed and input into the deep learning network model. Some of the subsequent parameters of the feature extraction backbone network are unfrozen, and Dropout regularization layers are added to designated layers of the feature extraction backbone network to suppress overfitting. The total loss function is used as the optimization objective. Training is conducted using the unfrozen partial parameter set of the feature extraction backbone network and the Dropout regularization layer. The corresponding parameters are updated through backpropagation of the gradient of the total loss function, completing the extreme scenario fine-tuning of the deep learning network model. Scene-enhanced training: In each stage of pre-training, satellite-specific fine-tuning, and extreme scene-enhanced training, the inversion accuracy index of the deep learning network model is monitored in real time. At the same time, the changing trends of the total loss function, data loss, and inversion operator constraint loss are monitored. If the inversion accuracy index does not decrease for a preset number of consecutive rounds and the total loss function, data loss, and inversion operator constraint loss all tend to stabilize without decreasing, an early stopping strategy is triggered to terminate the training of the corresponding stage. After pre-training, satellite-specific fine-tuning, and extreme scene-enhanced training are all completed and the training of each stage has converged, the trained aerosol optical thickness inversion model is obtained.

[0124] In some embodiments of this application, the aerosol optical thickness inversion model includes a feature extraction backbone network, an inverse operator constraint module, and an output correction module. The feature extraction backbone network consists of the first 49 convolutional layers based on a modified ResNet-50, including 1×1, 3×3, and 5×5 multi-scale convolutional kernel layers and residual blocks. The inverse operator constraint module incorporates an aerosol radiative transfer positive operator and a band sensitivity weight calculation unit. The output correction module includes a Sigmoid activation function layer, a dynamic scaling coefficient library, and an anomaly filtering logic layer. The processing module 64 is specifically used to extract basic reflectance features, band combination features, atmospheric geometric auxiliary features, and radiative derivative features based on preprocessed satellite remote sensing raw data, and to fuse and construct multidimensional input features adapted to the aerosol optical thickness inversion model. These multidimensional input features are then input into the feature extraction backbone network of the aerosol optical thickness inversion model. The multidimensional input features are then dimensionality-reduced and fused through the 1×1 convolutional kernel layer in the feature extraction backbone network. A 3×3 convolutional kernel layer captures local radiation features, and a 5×5 convolutional kernel layer captures global aerosol distribution features. After feature depth extraction and gradient transfer are completed by the residual block, the deep fusion features are output. The deep fusion features are input to the output correction module. First, the feature output is mapped to the [0,1] interval through the Sigmoid activation function layer. Then, based on the land surface type and atmospheric model corresponding to the preprocessed satellite remote sensing raw data, the dynamic scaling coefficient library is called to complete adaptive linear scaling. The output value that exceeds the reasonable range of the scene is corrected by the anomaly filtering logic layer to obtain the preliminary aerosol optical thickness prediction value. The auxiliary physical parameters corresponding to the preprocessed satellite remote sensing raw data are extracted. The preliminary aerosol optical thickness prediction value and the auxiliary physical parameters are input together into the inverse operator constraint module. The physical consistency verification is completed by the aerosol radiation transmission positive operator. The verification result is optimized by the band sensitivity weight calculation unit, and the aerosol optical thickness inference calculation result that conforms to the physical law of aerosol radiation transmission is output.

[0125] In some embodiments of this application, the processing module 64 can be further used to perform spatial smoothing processing on the aerosol optical thickness inference calculation results output by the aerosol optical thickness inversion model using Gaussian filtering to obtain smoothed aerosol optical thickness inversion results; derive the confidence level of the aerosol optical thickness inversion results based on the inverse operator constraint loss value output by the inverse operator constraint module, wherein the smaller the inverse operator constraint loss value, the higher the confidence level of the corresponding inversion results; perform multi-level quality classification on the aerosol optical thickness inversion results according to the confidence level, and mark the aerosol optical thickness inversion result regions of each level; convert the quality-classified aerosol optical thickness inversion results with completed level marking into GeoTIFF format in the WGS-84 coordinate system, and output the corresponding quality label map and inversion accuracy report to obtain aerosol optical thickness inversion results used to assist atmospheric correction.

[0126] In some embodiments of this application, such as Figure 7 As shown, the device may further include: a verification module 65; The verification module 65 can be used to construct an independent verification dataset, which includes spatiotemporal matching test samples, extreme scenario verification samples, and aerosol type verification samples that were not involved in model training. It calculates the core evaluation indicators, extreme scenario-specific evaluation indicators, and aerosol type-specific evaluation indicators for the aerosol optical thickness inversion results. The core evaluation indicators include the coefficient of determination, mean deviation, and root mean square error. Based on these indicators, the aerosol optical thickness inversion results are compared with those of the traditional aerosol optical thickness inversion algorithm. Simultaneously, the long-term stability of the aerosol optical thickness inversion model is verified. If all indicators of the aerosol optical thickness inversion results are superior to the corresponding indicators of the traditional aerosol optical thickness inversion algorithm, and the fluctuation range of the long-term stability verification indicators of the aerosol optical thickness inversion model is within a preset threshold, then the aerosol optical thickness inversion results are deemed to have passed performance verification.

[0127] It should be noted that other corresponding descriptions of the functional units involved in the device for AOD-assisted atmospheric correction based on deep learning inversion provided in this embodiment can be found in [reference needed]. Figure 1 and Figure 2 The corresponding descriptions in [the document] will not be repeated here.

[0128] Based on the above, Figure 1 and Figure 2 Accordingly, this embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the above-described method. Figure 1 and Figure 2 The method shown is an atmospheric correction method based on deep learning-based AOD inversion.

[0129] Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause an electronic device (such as personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of this application.

[0130] Based on the above, Figure 1 and Figure 2 The method shown, and Figure 6 , 7To achieve the above objectives, the present application also provides an electronic device, specifically a personal computer, tablet computer, server, or other network device, as shown in the virtual device embodiment. This device includes a storage medium and a processor; the storage medium stores a computer program; the processor executes the computer program to achieve the above-described objectives. Figure 1 and Figure 2 The method shown is an atmospheric correction method based on deep learning-based AOD inversion.

[0131] Optionally, the aforementioned physical devices may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB interfaces, card reader interfaces, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Wi-Fi interfaces), etc.

[0132] Those skilled in the art will understand that the physical device structure provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0133] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the aforementioned physical device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software in the information processing physical device.

[0134] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platform, or it can be implemented by hardware.

[0135] This invention, through constructing a multi-type training sample set that integrates measured and physical simulation data and performing quality screening and stratification, can compensate for the deficiency of insufficient spatiotemporal matching sample coverage and effectively solve the problem of large inversion bias in extreme and non-training scenarios of data-driven models. By creating a deep learning network model that integrates the total loss function of data loss and inverse operator constraint loss, the physical laws of aerosol radiative transfer are deeply integrated into the model parameter update process. This not only avoids the dependence of physical-driven algorithms on specific surface assumptions and solves the problem of accuracy degradation on heterogeneous and bright surfaces, but also addresses the root cause of the problem. The bundled model output conforms to physical laws, avoiding the generation of outliers without physical meaning. At the same time, it can overcome the shortcomings of simple superposition in traditional hybrid methods, effectively solve the multiple solutions of underdetermined inverse problems, and avoid the problem of data transformation error accumulation. By carrying out phased training through cross-satellite transfer learning strategies, the generalization ability and adaptability to extreme scenarios can be further improved. Combined with satellite remote sensing data preprocessing and inversion result postprocessing and quality classification, high-precision and high-generalization inversion of aerosol optical thickness under complex surface and atmospheric scenarios can be achieved, which greatly improves the actual effect of AOD inversion results in assisting atmospheric correction.

[0136] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application. Those skilled in the art will understand that the modules in the apparatus of the embodiment can be distributed within the apparatus of the embodiment as described, or can be modified to be located in one or more apparatuses different from this embodiment. The modules of the above-described embodiment can be combined into one module, or further divided into multiple sub-modules.

[0137] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of any particular implementation scenario. The above disclosures are merely a few specific implementation scenarios of this application; however, this application is not limited thereto, and any variations conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. A method for AOD-assisted atmospheric correction based on deep learning inversion, characterized in that, include: Construct a multi-type training sample set that integrates measured data and physical simulation data, and perform quality screening and stratification on the multi-type training sample set to obtain pre-training samples, fine-tuning samples and extreme scenario samples adapted to satellite remote sensing data; A deep learning network model is created, wherein the deep learning network model updates its parameters through a total loss function, which is a weighted fusion function of data loss and inverse operator constraint loss. The data loss characterizes the deviation between the model's predicted value and the measured value, and the inverse operator constraint loss characterizes the degree to which the model's predicted value conforms to the physical laws of aerosol radiative transfer. A cross-satellite transfer learning strategy was adopted to perform phased training on the deep learning network model, including pre-training, satellite-specific fine-tuning, and extreme scenario enhancement training, to obtain a trained aerosol optical thickness inversion model. Preprocessing operations such as radiometric calibration, cloud detection, and geometric correction are performed on the raw satellite remote sensing data. The preprocessed raw satellite remote sensing data is then input into the trained aerosol optical thickness inversion model for inference calculation. The model output results are post-processed and quality graded to obtain aerosol optical thickness inversion results used to assist atmospheric correction.

2. The method according to claim 1, characterized in that, The construction of a multi-type training sample set that integrates measured data and physical simulation data includes: Collect satellite remote sensing Level 1B data, ground-measured AOD data, and atmospheric auxiliary data; Based on the atmospheric auxiliary data, radiometric calibration, cloud detection, and geometric correction preprocessing operations are performed on the satellite remote sensing Level 1B data. Based on the preprocessed satellite remote sensing Level 1B data, outlier removal and spatiotemporal matching processing are performed on the ground measured AOD data to obtain spatiotemporally matched measured samples. An aerosol radiative transfer positive operator is constructed based on the 6S radiative transfer model. Taking the atmospheric model parameters in the atmospheric auxiliary data as a reference, multiple atmospheric models, multiple aerosol models, and the surface reflectance calculated by the surface endmember decomposition method are incorporated into the aerosol radiative transfer positive operator. Initial physical simulation data is generated through operator operations. Random noise that conforms to the characteristics of satellite observation is added to the initial physical simulation data to complete the simulation data enhancement and obtain the final physical simulation data. The spatiotemporal matching measured samples are fused with the physical simulation data to obtain an initial multi-type training sample set.

3. The method according to claim 1, characterized in that, The process of quality screening and stratification of the multi-type training sample sets yields pre-training samples, fine-tuning samples, and extreme scenario samples adapted to satellite remote sensing data, including: The multi-type training sample set is subjected to quality screening, which includes removing invalid samples whose AOD value and satellite top reflectivity exceed the preset range, and filtering parameter combination samples that have no physical meaning in the physical simulation process. For the samples that have passed the quality screening, band response correction is performed based on the characteristics of satellite sensors, and a pre-training sample set adapted for cross-satellite learning is constructed based on the corrected samples. The samples after quality screening are screened according to the atmospheric and surface characteristics of the target study area, and a fine-tuned sample set adapted to the target satellite is constructed based on the screened samples. Extreme aerosol event simulation samples are generated by generative adversarial networks. Physical consistency checks are performed on the extreme aerosol event simulation samples and valid samples are screened. An extreme scenario sample set is constructed based on the screened valid samples.

4. The method according to claim 1, characterized in that, The inverse operator constraint loss is a mean square error loss with band sensitivity weights. It is calculated by the inverse operator constraint module using the aerosol radiative transfer positive operator to determine the deviation between the simulated atmospheric top reflectivity and the actual atmospheric top reflectivity observed by the satellite, and then optimized using the band sensitivity weights. The total loss function is a weighted sum of the data loss and the inverse operator constraint loss multiplied by their respective balance coefficients. It is used to update the parameters of each module in the aerosol optical thickness inversion model through gradient backpropagation. The deep learning network model is trained in stages using a cross-satellite transfer learning strategy, including pre-training, satellite-specific fine-tuning, and extreme scenario enhancement training, to obtain a trained aerosol optical thickness inversion model, comprising: The pre-trained samples corrected by band response are input into the deep learning network model. Based on the pre-set optimizer and learning rate scheduling strategy, the feature extraction backbone network pre-sequence layer of the deep learning network model is trained with the total loss function as the optimization objective. The parameters of the feature extraction backbone network pre-sequence layer are updated through backpropagation of the gradient of the total loss function. The cross-satellite general mapping relationship between aerosols and radiation signals is learned, and cross-satellite pre-training is completed. Freeze the parameters of the pre-trained feature extraction backbone network pre-layer, input the fine-tuned sample into the deep learning network model, train the feature extraction backbone network post-layer and output correction module with the total loss function as the optimization target, update the parameters of the feature extraction backbone network post-layer and output correction module through gradient backpropagation of the total loss function, and complete the satellite-specific fine-tuning; The extreme scene samples and the fine-tuned samples are mixed and input into the deep learning network model. The post-processing parameters of the partial feature extraction backbone network are unfrozen, and a Dropout regularization layer is added to a specified layer of the feature extraction backbone network to suppress overfitting. The total loss function is used as the optimization objective. Training is carried out by combining the unfrozen post-processing parameters of the partial feature extraction backbone network with the Dropout regularization layer. The corresponding parameters are updated by backpropagation of the gradient of the total loss function to complete the extreme scene enhancement training of the deep learning network model. In each stage of the pre-training, satellite-specific fine-tuning, and extreme scenario enhancement training, the inversion accuracy index of the deep learning network model is monitored in real time. At the same time, the changing trends of the total loss function, the data loss, and the inversion operator constraint loss are monitored. If the inversion accuracy index does not decrease for a preset number of consecutive rounds and the total loss function, the data loss, and the inversion operator constraint loss all tend to stabilize without decreasing, an early stopping strategy is triggered to terminate the training of the corresponding stage. After the pre-training, satellite-specific fine-tuning, and extreme scenario enhancement training are all completed and the training of each stage has converged, the trained aerosol optical thickness inversion model is obtained.

5. The method according to claim 1, characterized in that, The aerosol optical thickness inversion model includes a feature extraction backbone network, an inverse operator constraint module, and an output correction module. The feature extraction backbone network consists of the first 49 convolutional layers based on a modified ResNet-50, including 1×1, 3×3, and 5×5 multi-scale convolutional kernel layers and residual blocks. The inverse operator constraint module incorporates an aerosol radiative transfer positive operator and a band sensitivity weight calculation unit. The output correction module includes a Sigmoid activation function layer, a dynamic scaling coefficient library, and an anomaly filtering logic layer. The step of inputting preprocessed satellite remote sensing raw data into the trained aerosol optical thickness inversion model for inference calculation includes: Based on the preprocessed satellite remote sensing raw data, basic reflectance features, band combination features, atmospheric geometric auxiliary features and radiation-derived features are extracted. Multi-dimensional input features adapted to the aerosol optical thickness inversion model are fused and constructed. The multi-dimensional input features are then input into the feature extraction backbone network of the aerosol optical thickness inversion model. The multidimensional input features are reduced in dimensionality and fused by the 1×1 convolutional kernel layer in the feature extraction backbone network, local radiation features are captured by the 3×3 convolutional kernel layer, and global aerosol distribution features are captured by the 5×5 convolutional kernel layer. After feature depth extraction and gradient transfer are completed by the residual block, the deep fused features are output. The deep fusion features are input into the output correction module. First, the feature output is mapped to the [0,1] interval through the Sigmoid activation function layer. Then, based on the land surface type and atmospheric mode corresponding to the preprocessed satellite remote sensing raw data, the dynamic scaling coefficient library is called to complete adaptive linear scaling. The output value that exceeds the reasonable range of the scene is corrected by the anomaly filtering logic layer to obtain the preliminary aerosol optical thickness prediction value. The auxiliary physical parameters corresponding to the preprocessed satellite remote sensing raw data are extracted. The preliminary aerosol optical thickness prediction value and the auxiliary physical parameters are input into the inverse operator constraint module. The physical consistency verification is completed by the aerosol radiation transmission positive operator. The verification result is optimized by the band sensitivity weight calculation unit, and the aerosol optical thickness inference calculation result that conforms to the physical law of aerosol radiation transmission is output.

6. The method according to claim 5, characterized in that, The post-processing and quality grading of the model output results yields aerosol optical thickness inversion results used to assist atmospheric correction, including: The aerosol optical thickness inference calculation results output by the aerosol optical thickness inversion model are spatially smoothed using Gaussian filtering to obtain smoothed aerosol optical thickness inversion results. The confidence level of the aerosol optical thickness inversion result is derived based on the inverse operator constraint loss value output by the inverse operator constraint module, wherein the smaller the inverse operator constraint loss value, the higher the confidence level of the corresponding inversion result. Based on the confidence level, the aerosol optical thickness inversion results are classified into multiple levels of quality, and the regions of the aerosol optical thickness inversion results at each level are marked with a level label. The quality-graded aerosol optical thickness inversion results with completed grade markings are converted into GeoTIFF format in the WGS-84 coordinate system. At the same time, the corresponding quality label map and inversion accuracy report are output to obtain aerosol optical thickness inversion results for assisting atmospheric correction.

7. The method according to claim 6, characterized in that, The method further includes performance verification of the aerosol optical thickness inversion results, including: Construct an independent validation dataset, which includes spatiotemporal matching test samples that did not participate in model training, extreme scenario validation samples, and aerosol type validation samples. The core evaluation indicators, extreme scenario-specific evaluation indicators, and aerosol type-specific evaluation indicators for the aerosol optical thickness inversion results are calculated. The core evaluation indicators include the coefficient of determination, average deviation, and root mean square error. Based on the core evaluation indicators, the extreme scenario-specific evaluation indicators, and the aerosol type-specific evaluation indicators, the aerosol optical thickness inversion results are compared with the results of the traditional aerosol optical thickness inversion algorithm. At the same time, the long-term stability of the aerosol optical thickness inversion model is verified. If all indicators of the aerosol optical thickness inversion results are better than the corresponding indicators of the traditional aerosol optical thickness inversion algorithm, and the fluctuation range of the indicators of the long-term stability verification of the aerosol optical thickness inversion model is within the preset threshold, then the aerosol optical thickness inversion results are determined to have passed the performance verification.

8. A device for AOD-assisted atmospheric correction based on deep learning inversion, characterized in that, include: The construction module is used to build a multi-type training sample set that integrates measured data and physical simulation data. The multi-type training sample set is subjected to quality screening and stratification to obtain pre-training samples, fine-tuning samples and extreme scenario samples adapted to satellite remote sensing data. A creation module is used to create a deep learning network model. The deep learning network model updates its parameters through a total loss function, which is a weighted fusion function of data loss and inverse operator constraint loss. The data loss characterizes the deviation between the model's predicted value and the measured value, and the inverse operator constraint loss characterizes the degree to which the model's predicted value conforms to the physical laws of aerosol radiative transfer. The training module is used to perform phased training on the deep learning network model using a cross-satellite transfer learning strategy, including pre-training, satellite-specific fine-tuning, and extreme scenario enhancement training, to obtain a trained aerosol optical thickness inversion model. The processing module is used to perform preprocessing operations such as radiometric calibration, cloud detection, and geometric correction on the raw satellite remote sensing data. The preprocessed raw satellite remote sensing data is then input into the trained aerosol optical thickness inversion model for inference calculation. The model output results are then post-processed and quality graded to obtain aerosol optical thickness inversion results used to assist atmospheric correction.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

10. An electronic device comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

Images