A multi-parameter collaborative inversion method for aerosol, surface and leaf area index

Abstract

The application discloses a multi-parameter collaborative inversion method for aerosol, surface and leaf area index, and belongs to the technical field of land ecological monitoring. The application constructs a unified multi-task cascade collaborative inversion model, simultaneously inverts AOD, LSR and LAI, and changes a traditional serial processing flow; by jointly optimizing atmospheric parameters and surface vegetation parameters under unified observation constraints, information sharing and collaborative constraints between different physical variables are realized, thereby effectively improving consistency and overall precision of multi-parameter inversion results; based on a machine learning inversion method, complex atmospheric radiation transmission calculation and multi-step iterative processing processes in a traditional method are avoided, and calculation efficiency is significantly improved; by collaboratively modeling atmospheric and vegetation parameters, influence of aerosol inversion error on LAI inversion results is effectively weakened, and strong robustness is achieved under complex atmospheric conditions. The application is suitable for high-frequency inversion application at a global or regional scale, and has good engineering practicability.

Description

Technical Field

[0001] This invention relates to terrestrial ecological monitoring technology, specifically to a multi-parameter collaborative inversion method for aerosols, land surface and leaf area index. Background Technology

[0002] Leaf area index (LAI) is an important parameter describing vegetation canopy structure and plays a crucial role in terrestrial ecosystem monitoring, carbon cycle research, and climate change analysis. Utilizing satellite remote sensing to obtain spatiotemporal distribution information of LAI has become a major technical approach for global-scale vegetation monitoring.

[0003] During remote sensing observations, the radiation signals received by satellite sensors are the result of the interaction between the Earth's surface and the atmosphere. These signals include both reflectance information from the surface vegetation canopy and atmospheric effects caused by scattering and absorption by aerosols and gas molecules. Therefore, in vegetation parameter inversion, it is usually necessary to first correct for atmospheric effects to achieve effective separation of surface and atmospheric signals.

[0004] In existing technologies, LAI retrieval based on satellite remote sensing typically employs a sequential processing flow. This involves first retrieving atmospheric parameters such as aerosol optical depth (AOD), then performing atmospheric correction based on these parameters to obtain land surface reflectance (LSR), and finally using LSR to retrieve LAI. Therefore, the accuracy of LAI retrieval largely depends on the accuracy of AOD retrieval and atmospheric correction.

[0005] However, this type of serial processing method is prone to introducing stepwise error propagation in practical applications. That is, AOD inversion errors and residuals generated during atmospheric correction are propagated along the processing flow and amplified during surface reflectance calculation, thus significantly affecting LAI inversion results. This is especially true in areas with rapid aerosol changes or heavy pollution, where complex atmospheric conditions significantly increase AOD inversion uncertainty, leading to larger surface reflectance estimation errors and reducing the reliability of LAI inversion results. Furthermore, limited by atmospheric correction accuracy and observational conditions, existing LAI products typically require screening and synthesizing multi-day observational data, selecting only observational data with clear skies and minimal aerosol influence to generate products. This limits the temporal resolution of most LAI satellite products to an 8-day scale, making it difficult to obtain continuous daily-scale LAI information, thereby weakening the ability to characterize rapidly changing vegetation processes.

[0006] Therefore, there is an urgent need for a new technical solution that can achieve the coordinated inversion of atmospheric aerosol parameters, surface parameters and leaf area index under a unified framework, while reducing the error propagation effect, improving the accuracy and stability of LAI inversion under complex atmospheric conditions, and improving the temporal resolution of the inversion results while ensuring accuracy, so as to support global vegetation dynamic monitoring on a daily scale. Summary of the Invention

[0007] To address the problems of atmospheric correction error propagation, accuracy degradation under complex aerosol conditions, and limited temporal resolution in existing LAI inversion methods, this invention proposes a multi-parameter collaborative inversion method for aerosols, land surface, and leaf area index. Under a unified inversion framework, it achieves joint inversion of AOD, LSR, and LAI. By combining a multi-task machine learning strategy, atmospheric aerosol parameters and land surface vegetation parameters are collaboratively optimized under the same observation constraints, thereby effectively reducing the propagation of atmospheric correction errors to LAI inversion and improving the inversion accuracy and stability under complex atmospheric conditions.

[0008] The multi-parameter synergistic inversion method for aerosols, land surface and leaf area indices of the present invention includes the following steps:

[0009] 1) Obtain the raw data:

[0010] We obtain satellite observation data, meteorological data, and auxiliary product data; the auxiliary product data includes multi-band aerosol optical thickness, surface reflectance, and leaf area index.

[0011] 2) Data processing:

[0012] a) Clear-sky data processing:

[0013] Radiometric correction was applied to the satellite observation data, and the normalized vegetation index and aerosol free vegetation index were calculated. Cloud, ice and snow pixels were removed from the radiometrically corrected satellite observation data to obtain clear-sky satellite observation data.

[0014] b) Quality control procedures:

[0015] The auxiliary product data is processed for quality control to remove low-quality and abnormal data, thus obtaining quality control auxiliary product data.

[0016] 3) Construct the dataset for training the model:

[0017] The input dataset is obtained by matching the clear-sky satellite observation data and meteorological data for temporal and spatial consistency; the output dataset is obtained by matching the quality control auxiliary product data for temporal and spatial consistency; the input dataset and the output dataset are then matched for temporal and spatial consistency to form a dataset for training the model.

[0018] 4) Constructing the inversion model:

[0019] A multi-task cascaded collaborative inversion model is constructed, comprising: an input layer, a shared feature extraction layer, a branch prediction layer, and an output layer. The branch prediction layer includes three distinct branches: the AOD branch prediction layer, the LSR branch prediction layer, and the LAI branch prediction layer. The output of the shared feature extraction layer is connected to the inputs of the three branches of the branch prediction layer, and the outputs of the three branches are connected to the output layer. Furthermore, the outputs of the AOD and LSR branch prediction layers are also connected to the inputs of the LAI branch prediction layer. Clear-sky satellite observation data and meteorological data serve as input features, while quality control auxiliary product data serve as output features. The shared feature extraction layer extracts common information from the input features and the output features. The input features with the common information extracted are then input to the corresponding branch prediction layers. The AOD branch prediction layer and the LSR branch prediction layer predict aerosol optical thickness and surface reflectance, respectively, and input them to the LAI branch prediction layer. The LAI branch prediction layer uses the input features combined with the aerosol optical thickness and surface reflectance predicted by the AOD branch prediction layer and the LSR branch prediction layer to perform cascaded refinement, predicting the cascaded refined leaf area index, thereby achieving the joint constraint of "shared features, branch prediction and cascaded refinement".

[0020] 5) Training the inversion model:

[0021] Using the dataset constructed in step 3), the multi-task cascaded collaborative inversion model is trained to obtain the trained multi-task cascaded collaborative inversion model.

[0022] 6) Applying the inversion model:

[0023] After obtaining satellite observation data and meteorological data, perform clear sky data processing according to step 2), and perform time and space consistency matching according to step 3), input the data into the trained multi-task cascaded collaborative inversion model to obtain aerosol optical thickness, surface reflectivity and leaf area index through collaborative inversion.

[0024] In step 1), the multi-band satellite observation data includes multi-band apparent reflectance of the atmospheric top and its corresponding observation geometric parameters, altitude, and land cover type; the meteorological data includes atmospheric state parameters. The apparent reflectance of the atmospheric top and its corresponding observation geometric parameters, altitude, and land cover type are obtained from satellite observations; among the atmospheric state parameters, this invention only uses surface air pressure, total water vapor column, and total ozone column, provided by meteorological reanalysis data products; the multi-band aerosol optical thickness is derived from satellite aerosol inversion products; the surface reflectance is obtained from satellite surface inversion products; and the leaf area index is obtained from land surface vegetation fusion products. The observation geometric parameters include the cosine values ​​of the observed zenith angle, solar zenith angle, and relative azimuth angle.

[0025] In step 2)a), the Normalized Difference Vegetation Index (NDVI) and the Aerosol Free Vegetation Index (AFRI) are calculated based on the radiation-corrected satellite observation data.

[0026] Because the radiative effects of aerosols differ significantly under cloudy and cloudless conditions, and because snow and ice surfaces have high reflectivity, they can easily interfere with AOD inversion, thus affecting the accuracy of subsequent LSR and LAI inversions. Therefore, before performing aerosol, surface, and leaf area index inversions, it is necessary to identify and remove cloud, ice, and snow pixels from satellite observation data to ensure the physical consistency of the input data and the reliability of the inversion results.

[0027] In step 3), data independent of the training dataset is used as the test dataset.

[0028] In step 4), the shared feature extraction layer extracts common information from all variables of the output features from the input features to characterize the coupling relationship between the atmosphere, land surface, and vegetation. Different branch prediction layers predict AOD, LSR, and LAI-related variables separately to preserve the differences between different physical quantities. Simultaneously, the LAI branch prediction layer explicitly incorporates AOD and LSR prediction results through cascaded refinement, enabling LAI inversion to utilize atmospheric and surface information for correction and optimization.

[0029] In step 5), during the training process, the input features are standardized, and the model hyperparameters are searched using a multi-objective optimization algorithm. An early stopping strategy is also used to prevent the model from overfitting.

[0030] Advantages of this invention:

[0031] (1) This invention uses satellite observation data to construct a unified multi-task cascaded collaborative inversion model, which simultaneously inverts aerosol optical thickness (AOD), land surface reflectance (LSR) and leaf area index (LAI) in the same model, changing the traditional serial processing flow based on "aerosol inversion - atmospheric correction - vegetation inversion"; by jointly optimizing atmospheric parameters and land surface vegetation parameters under unified observation constraints, information sharing and collaborative constraints between different physical variables are realized, thereby effectively improving the consistency and overall accuracy of multi-parameter inversion results.

[0032] (2) The present invention adopts a machine learning-based inversion method, and realizes end-to-end parameter prediction by constructing a multi-task learning model, which avoids the complex atmospheric radiation transfer calculation and multi-step iterative processing process in traditional methods, and significantly improves the computational efficiency. Under the premise of ensuring inversion accuracy, it can quickly process large-scale remote sensing data, and is suitable for high-frequency inversion applications at global or regional scales, and has good engineering practicality.

[0033] (3) The present invention has strong robustness under complex atmospheric conditions, especially in regions with high or rapidly changing aerosol optical thickness, and can still maintain stable inversion performance. By co-modeling atmospheric and vegetation parameters, the influence of aerosol inversion error on LAI inversion results is effectively reduced, thereby significantly improving the spatial consistency and temporal stability of LAI products under high AOD conditions.

[0034] (4) Based on the collaborative inversion method proposed in this invention, AOD, LSR and LAI datasets with a time resolution of day can be constructed. Compared with traditional products that rely on multi-day synthesis, this method significantly improves the temporal continuity and observation frequency of LAI data, thereby better supporting the dynamic monitoring and analysis of rapidly changing vegetation processes, and providing more timely data support for terrestrial ecosystem research and climate change assessment. Attached Figure Description

[0035] Figure 1 This is a flowchart of the multi-parameter synergistic inversion method for aerosols, land surface and leaf area indices of the present invention. Detailed Implementation

[0036] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0037] The multi-parameter collaborative inversion method for aerosols, land surface, and leaf area indices in this embodiment, such as... Figure 1 As shown, it includes the following steps:

[0038] 1) Obtain the raw data:

[0039] The system obtains multi-band satellite observation data, meteorological data, and auxiliary product data. The satellite observation data includes multi-band apparent reflectance of the top of the atmosphere and its corresponding observation geometric parameters, altitude, and land cover type. The meteorological data includes atmospheric state parameters. The auxiliary product data includes multi-band aerosol optical thickness, surface reflectance, and leaf area index.

[0040] Multispectral primary data products from the MERSI sensor of the domestically produced Fengyun-3 D satellite were used to obtain multi-band apparent reflectance of the top of the atmosphere in the 470 nm, 550 nm, 650 nm, 865 nm, 1640 nm, and 2130 nm bands, along with their corresponding observational geometric parameters, altitude, and land cover type, constituting satellite observation data. Atmospheric state parameters were provided by meteorological reanalysis data products from the European Centre for Medium-Range Weather Forecasts Reanalysis (ERA5), constituting meteorological data. This invention only uses surface air pressure, total water vapor column, and total ozone column. Satellite aerosol inversion products from the Moderate Resolution Imaging Spectroradiometer (MODIS) yielded multi-band aerosol optical thickness in the 470 nm, 550 nm, 650 nm, and 2130 nm bands. MODIS satellite surface inversion products yielded data in the 470 nm, 550 nm, 650 nm, and 2130 nm bands. Multi-band surface reflectance in the nm band is obtained from the Global Land Surface Satellite Leaf Area Index Product (GLASS LAI), which constitutes auxiliary product data; the observation geometric parameters include the cosine values ​​of the observation zenith angle, solar zenith angle, and relative azimuth angle;

[0041] 2) Data processing:

[0042] a) Clear-sky data processing:

[0043] Radiometric correction was performed on satellite observation data in the 470 nm, 550 nm, 650 nm, 865 nm, 1640 nm, and 2130 nm bands. This radiometric correction employed existing, mature calibration methods. In this embodiment, cross-calibration with MODIS sensor observation data was used to obtain stable and reliable atmospheric top reflectance data. Based on the radiometrically corrected satellite observation data, a vegetation index calculation method based on multispectral band reflectance combinations was employed to calculate the normalized vegetation index (NDI) constructed using the shortwave infrared band. And the Aerosol Free Vegetation Index (AFRI); in this embodiment, they are calculated using the following formulas:

[0044]

[0045]

[0046] in, and These represent the radiometrically corrected apparent reflectances of the top of the atmosphere at 865 nm, 1640 nm, and 2130 nm, respectively; the normalized vegetation index (NDI) constructed based on the shortwave infrared band in this invention... Compared with traditional systems built on red and near-infrared bands Distinguish between them:

[0047]

[0048] Where, ρ 650 The apparent reflectance of the atmospheric top at 650 nm after radiation correction; in comparison, the normalized vegetation index constructed based on the shortwave infrared band. MERSI observations are more stable and less affected by aerosol variations; meanwhile, under other satellite observation conditions, traditional methods based on red and near-infrared bands are used. The definition is also feasible; AFRI reduces atmospheric impact by introducing aerosol-sensitive bands to correct vegetation signals.

[0049] Because the radiation effects of aerosols differ significantly under cloudy and cloudless conditions, and because snow and ice surfaces have high reflectivity, they can easily interfere with AOD inversion, thus affecting the accuracy of subsequent LSR and LAI inversions. Therefore, before inversion, it is necessary to identify and remove cloud and snow pixels from satellite observation data, normalized vegetation index, and aerosol free vegetation index to ensure the physical consistency of the input data and the reliability of the inversion results, and obtain clear-sky satellite observation data.

[0050] b) Quality control procedures:

[0051] Quality control processing of auxiliary product data involves using quality flags provided in the auxiliary product data to identify and eliminate low-quality and abnormal data. Specifically, this involves using the quality flags provided for each product to assess data quality, eliminating low-quality and abnormal data, and retaining only data within a reasonable physical range with a high-quality flag, thereby obtaining quality-controlled auxiliary product data. The reasonable physical range can be defined based on the physical meaning of each parameter, including but not limited to:

[0052] For aerosol optical thickness (AOD) in each band, retain data with values ​​in the range of 0 to 5;

[0053] For the surface reflectance (LSR) of each band, retain the data with values ​​in the range of 0 to 1;

[0054] For leaf area index (LAI), retain data with values ​​in the range of 0 to 10;

[0055] The quality label indicates excellent data, specifically selecting data from each product's quality label that signifies high quality or reliability; for example:

[0056] For MODIS aerosols and surface products, select pixels that indicate "high-quality inversion" in the Quality Assurance (QA) label;

[0057] Through the above quality control process, data with cloud contamination, inversion failure, or observational anomalies can be effectively removed, thereby improving the reliability of training samples and the stability of the inversion model.

[0058] 3) Construct the dataset for training the model:

[0059] The input dataset is obtained by matching the clear-sky satellite observation data and meteorological data for temporal and spatial consistency; the output dataset is obtained by matching the quality control auxiliary product data for temporal and spatial consistency; the input dataset and the output dataset are then matched for temporal and spatial consistency to form a dataset for training the model.

[0060] 4) Constructing the inversion model:

[0061] A multi-task cascaded collaborative inversion model is constructed, comprising: an input layer, a shared feature extraction layer, a branch prediction layer, and an output layer. The input layer is connected to the input of the shared feature extraction layer. The output of the shared feature extraction layer is connected to the input of the three branches of the branch prediction layer, and the outputs of the three branches are connected to the output layer. Furthermore, the outputs of the AOD and LSR branch prediction layers are also connected to the input of the LAI branch prediction layer. Temporally and spatially consistent clear-sky satellite observation data and meteorological data are used as input features, and temporally and spatially consistent quality control auxiliary products are used as output features.

[0062] The shared feature extraction layer is used to extract multi-source coupled information from the input features that can simultaneously characterize atmospheric state, surface reflectance properties, and vegetation canopy structure. The shared information is output to the branch prediction layer in the form of a shared feature vector. The shared information includes, but is not limited to:

[0063] i) Information about atmospheric scattering and absorption characteristics implied in the apparent reflectance of the top of the atmosphere;

[0064] ii) The modulation relationship between satellite observation geometric parameters and the radiative transfer process, including the influence of solar zenith angle, observation zenith angle and relative azimuth angle on reflectivity;

[0065] iii) The spectral response relationship between the apparent reflectance of the top of the atmosphere in different bands is used to reflect the differences in aerosol type and surface reflectance;

[0066] iv) Information on vegetation cover and growth status represented by the two vegetation indices, NDVI and AFRI.

[0067] v) Information on the influence of atmospheric state parameters used in this invention on atmospheric radiative transfer processes;

[0068] vi) Constraint information on atmospheric radiation and surface reflection characteristics caused by altitude and land cover type;

[0069] The above information together constitutes the coupled feature representation of the atmosphere-land surface-vegetation system, providing a unified feature basis for subsequent multi-task collaborative inversion;

[0070] The branch prediction layer comprises three distinct branches: the AOD branch prediction layer, the LSR branch prediction layer, and the LAI branch prediction layer. Shared feature vectors containing common information are extracted and input into the respective branch prediction layers. These different branch prediction layers then predict AOD, LSR, and LAI-related variables to preserve the differences between various physical quantities. The AOD and LSR branch prediction layers predict aerosol optical thickness and surface reflectance, respectively, and these predictions are input into the LAI branch prediction layer. The LAI branch prediction layer uses the input features combined with the aerosol optical thickness and surface reflectance predicted by the AOD and LSR branch prediction layers for cascaded refinement, resulting in a cascaded refined leaf area index. The specific process includes the following steps:

[0071] i) First, the shared feature vector output by the shared feature extraction layer is concatenated with the aerosol optical thickness obtained by the AOD branch prediction layer and the surface reflectance obtained by the LSR branch prediction layer to form an extended feature vector.

[0072] ii) Next, the extended feature vector is input into the LAI branch prediction layer. The relationship between AOD and LSR and the original input features is jointly calculated through the nonlinear mapping in the multi-layer neural network, so as to learn the influence mechanism of aerosols on surface reflection and vegetation signals in radiative transfer.

[0073] iii) Further, within the LAI branch prediction layer, the features input to the LAI branch prediction layer are reconstructed and weighted, and the predicted LAI is output. This enables the multi-task cascaded collaborative inversion model to adaptively adjust the contribution weights of AOD and LSR to LAI inversion according to different atmospheric conditions. Feature reconstruction refers to performing high-dimensional nonlinear mapping on the input features through the multi-task cascaded collaborative inversion model, converting the input features into feature representations that are more conducive to LAI prediction. Weighted combination refers to learning the importance of different features through network parameters to achieve adaptive allocation of the contribution weights of AOD, LSR, and common information.

[0074] After the above steps, the final output is the leaf area index prediction result after cascaded refinement.

[0075] By explicitly incorporating AOD and LSR prediction results through cascaded refinement in the LAI branch prediction layer, LAI inversion can utilize atmospheric and surface information for correction and optimization, thereby achieving the joint constraint of "shared features, branch prediction and cascaded refinement".

[0076] 5) Training the inversion model:

[0077] Using the dataset constructed in step 3), the multi-task cascaded collaborative inversion model was trained. During the training process, the input features were standardized, and the model hyperparameters were searched using a multi-objective optimization algorithm. An early stopping strategy was adopted to prevent the model from overfitting, resulting in a well-trained multi-task cascaded collaborative inversion model. The training data consisted of sample data from 2020 to 2022, and the test data consisted of independent sample data from 2023.

[0078] The Huber loss function is used as the basic loss term for each task to improve the model's robustness to outliers. Meanwhile, bias and slope constraints are introduced in the AOD branch prediction layer to reduce systematic bias in the prediction results and enhance the consistency between predicted and true values, thereby reducing the propagation of AOD errors to LAI inversion. Furthermore, cascaded refinement is employed in the LAI branch prediction layer to improve the stability and convergence efficiency of LAI inversion.

[0079] Furthermore, a multi-objective optimization method is used to search and optimize the model hyperparameters, including the number of network layers, the number of hidden layer neurons, the learning rate, the batch size, and the regularization parameter; by comprehensively evaluating the errors of multiple inversion objectives on the validation set, the optimal combination of model parameters is selected.

[0080] 6) Applying the inversion model:

[0081] New apparent reflectance of the top of the atmosphere and its corresponding observational geometric parameters, altitude and land cover type were obtained using MERSI observations. Atmospheric state parameters were obtained from meteorological reanalysis data products. Clear sky data were processed according to step 2), and temporal and spatial consistency matching was performed. The data were then input into a trained multi-task cascaded collaborative inversion model to obtain aerosol optical thickness, surface reflectance and leaf area index through collaborative inversion.

[0082] Step 2a) Cloud, ice, and snow pixel screening process is as follows:

[0083] (1) Identify the effective observation area: Only retain pixels with effective land cover type, and remove water body pixels (land cover type code 0 or 17), snow and ice pixels (code 15), and invalid land cover code pixels (pixels with code greater than 17).

[0084] (2) Perform validity judgment of apparent reflectance at the top of the atmosphere: For any pixel, if the reflectance of any band exceeds the preset physical reasonable range of reflectance (less than the lower limit or greater than the upper limit), the pixel is judged as an invalid pixel and removed; wherein, in this embodiment, the physical reasonable range of reflectance can be set between 0 and 1; further, for outlier removal, the lower limit threshold can be set to 0.001 to avoid noise interference or invalid observation data from participating in subsequent processing;

[0085] (3) Cloud identification based on visible light band reflectance: When the reflectance of the 470 nm band is greater than 0.4, the pixel is identified as a cloud-polluted pixel;

[0086] (4) Identify cloud pollution using spatial consistency features: Calculate the standard deviation of reflectance in the neighborhood window for the 550 nm and 1380 nm bands. When the local standard deviation of any band is greater than 0.0025, the pixel is identified as a cloud pollution pixel.

[0087] (5) Further cloud identification based on shortwave infrared band threshold: When the reflectance of the 1380 nm band is greater than 0.015, the pixel is identified as a cloud-polluted pixel to enhance the ability to identify thin clouds in the upper atmosphere.

[0088] (6) Snow and ice pixel removal based on snow and ice index: Calculate the normalized differential snow index. When the index is greater than 0.4, the corresponding pixel is identified as a snow and ice pixel and removed.

[0089] (7) Water body removal based on near-infrared index: Calculate the near-infrared normalized difference index. When the index is less than 0, the corresponding pixel is identified as a water body pixel and removed.

[0090] Finally, it should be noted that the purpose of disclosing the embodiments is to help further understand the present invention. However, those skilled in the art will understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection of the present invention is defined by the claims.

Claims

1. A multi-parameter synergistic inversion method for aerosols, land surface and leaf area indices, characterized in that, The multi-parameter collaborative inversion method includes the following steps: 1) Obtain the raw data: We obtain satellite observation data, meteorological data, and auxiliary product data; the auxiliary product data includes multi-band aerosol optical thickness, surface reflectance, and leaf area index. 2) Data processing: a) Clear-sky data processing: Clear-sky data processing is performed on satellite observation data to obtain clear-sky satellite observation data; b) Quality control procedures: The auxiliary product data is processed for quality control to remove low-quality and abnormal data, thus obtaining quality control auxiliary product data. 3) Construct the dataset: The dataset for training the model is constructed by matching the temporal and spatial consistency of clear-sky satellite observation data, meteorological data, and quality control auxiliary product data. 4) Constructing the inversion model: A multi-task cascaded collaborative inversion model is constructed, comprising: an input layer, a shared feature extraction layer, a branch prediction layer, and an output layer. The branch prediction layer includes: an AOD branch prediction layer, an LSR branch prediction layer, and a LAI branch prediction layer. The output of the shared feature extraction layer is connected to the inputs of the AOD, LSR, and LAI branch prediction layers, respectively. The outputs of the three branch prediction layers are connected to the output layer, and the outputs of the AOD and LSR branch prediction layers are also connected to the inputs of the LAI branch prediction layer. Clear-sky satellite observation data and meteorological data are used as input features, and quality control auxiliary product data are used as output features. 5) Training the inversion model: Using the dataset constructed in step 3), the multi-task cascaded collaborative inversion model is trained to obtain the trained multi-task cascaded collaborative inversion model. 6) Applying the inversion model: Satellite observation data and meteorological data are obtained, processed, and input into a trained multi-task cascaded collaborative inversion model to obtain auxiliary product data through collaborative inversion.

2. The multi-parameter collaborative inversion method according to claim 1, characterized in that, In step 1), satellite observation data includes the apparent reflectance of the top of the atmosphere and its corresponding observation geometric parameters, altitude, and land cover type; meteorological data includes atmospheric state parameters.

3. The multi-parameter collaborative inversion method according to claim 1, characterized in that, In step 2), the satellite observation data is radiometrically corrected. Based on the radiometrically corrected satellite observation data, the normalized vegetation index and aerosol free vegetation index are calculated. Cloud, ice and snow pixels are removed from the radiometrically corrected satellite observation data to obtain clear sky satellite observation data.

4. The multi-parameter collaborative inversion method according to claim 1, characterized in that, In step 3), the clear-sky satellite observation data and meteorological data are matched for temporal and spatial consistency to obtain the input dataset; the quality control auxiliary product data are matched for temporal and spatial consistency to obtain the output dataset; the input dataset and the output dataset are matched for temporal and spatial consistency to form the dataset.

5. The multi-parameter collaborative inversion method according to claim 1, characterized in that, In step 4), the shared feature extraction layer extracts common information from the input features and the output features. The input features with the common information extracted are then input to the AOD, LSR, and LAI branch prediction layers, respectively. The AOD and LSR branch prediction layers predict aerosol optical thickness and surface reflectance, respectively, and input them to the LAI branch prediction layer. The LAI branch prediction layer uses the input features combined with the aerosol optical thickness and surface reflectance predicted by the AOD and LSR branch prediction layers to perform cascaded refinement, predicting the cascaded refined leaf area index, and performing joint constraints of shared features, branch prediction, and cascaded refinement.

6. The multi-parameter collaborative inversion method according to claim 5, characterized in that, In step 6), satellite observation data and meteorological data are processed for clear sky data according to step 2), and temporal and spatial consistency matching is performed according to step 3). The data are then input into the trained multi-task cascaded collaborative inversion model to obtain aerosol optical thickness, surface reflectivity and leaf area index through collaborative inversion.

Images