A method, apparatus, and storage medium for multi-parameter, multi-band joint remote sensing inversion of aerosol optical properties

Abstract

This disclosure provides a method, apparatus, and storage medium for multi-parameter, multi-band joint remote sensing inversion of aerosol optical properties, relating to the fields of Earth observation and remote sensing technology. The method includes acquiring observational data of multi-band atmospheric top-layer reflectivity from satellites and ground-based observation station aerosol parameter ground truth data; constructing dynamic observation features and static background features based on the observational data and multi-source auxiliary data, respectively; using the aerosol parameter ground truth data as training labels, inputting the dynamic observation features into a first-stage learning inversion model for multi-task joint inversion, outputting initial prediction results for aerosol multi-parameters, and generating predictive statistical features; fusing the static background features and predictive statistical features into a second-stage learning inversion model for correction, outputting the inversion results for aerosol multi-parameters. This addresses the shortcomings of existing technologies, such as single inversion parameters, susceptibility to missing measurements and spatial fragmentation under bright or complex surfaces, and insufficient generalization ability of single-data-driven models.

Description

Technical Field

[0001] This invention relates to the field of Earth observation and remote sensing technology, and in particular to a method, apparatus and storage medium for multi-parameter multi-band joint remote sensing inversion of aerosol optical properties. Background Technology

[0002] Aerosols have a significant impact on radiation balance and air quality health. Currently, large-scale information is obtained through satellite remote sensing, but existing technologies have obvious limitations: First, mainstream satellite operational products typically only provide 550 nm aerosol optical depth (AOD), lacking simultaneous acquisition of key parameters such as multi-band AOD, single scattering albedo (SSA), asymmetry factor (AF) of the scattering phase function, and fine-mode / coarse-mode AOD. Relying solely on AOD is insufficient to distinguish aerosol types and sources (wildfire smoke, dust, urban pollution, etc.), and it is also difficult to conduct radiation effect and absorption characteristic analysis, rapid event tracking, etc. Although the ground-based AERONET (AErosolRObotic NETwork) solar photometer can provide high-precision, multi-band AOD, and, under certain conditions, invert SSA and AF, its sparse stations and insufficient spatial coverage make it difficult to characterize large-scale distributions. Secondly, physical model-based inversion methods (such as the dark target method) are sensitive to surface reflectance assumptions and are prone to failure in high-albedo bare land and cloud edges, resulting in a large amount of missing data and noise, making it difficult to monitor aerosol changes under extreme events. Finally, machine learning / deep learning can learn nonlinear mappings from multi-source data and reduce the dependence on accurate surface / atmosphere priors and iterative calculations. However, conventional data-driven solutions use a single model, such as using only Gradient Boosting Decision Tree (GBDT) or deep networks for single-parameter modeling, or using only deep models to predict directly from observations. However, relying solely on dynamic observation features is insufficient to characterize regional differences (such as topography, underlying surface, and anthropogenic emission intensity). Using only deep networks or only tree models cannot simultaneously take into account nonlinear fitting, robustness, and multi-parameter joint information sharing, thus affecting cross-site and cross-time generalization.

[0003] In summary, existing technologies have significant drawbacks, such as the inability to identify types and assess radiation effects due to the reliance on single inversion parameters, the susceptibility to missing data and spatial fragmentation under bright or complex surfaces, and the insufficient generalization ability of data-driven single-model schemes. These limitations restrict the application of satellite remote sensing in aerosol scientific research and emergency monitoring of extreme events. Summary of the Invention

[0004] In view of the many shortcomings of the existing technology, this disclosure provides a method, device and storage medium for multi-parameter multi-band joint remote sensing inversion of aerosol optical properties. Based on apparent reflectance data observed by satellite passive remote sensing (sun-synchronous or geostationary orbit), it realizes multi-task joint inversion of the optical properties of aerosols on land, and outputs a multi-parameter set covering aerosol concentration, spectral characteristics, absorption and scattering characteristics, morphological information and particle size modes, thereby improving the monitoring capability of aerosol extreme events (such as wildfires, dust storms and haze).

[0005] The first aspect of this disclosure provides a multi-parameter, multi-band joint remote sensing inversion method for aerosol optical properties, comprising the following steps: acquiring observational data of atmospheric top-layer reflectivity from satellite multi-band and ground-based observation station aerosol parameter ground-based data, and performing spatiotemporal matching between the observational data and the ground-based aerosol parameter ground-based data; constructing dynamic observation features and static background features based on the observational data and multi-source auxiliary data, wherein the multi-source auxiliary data includes atmospheric reanalysis datasets, aerosol climate datasets, topographic elevation, and human activity proxy variables; using the ground-based aerosol parameter ground-based data as training labels, inputting the dynamic observation features into a first-stage learning inversion model for strong nonlinear fitting, outputting initial prediction results for multiple aerosol characteristic parameters, and calculating statistics on the output of the first-stage learning inversion model to generate predicted statistical features; fusing the static background features and predicted statistical features, and inputting them into a second-stage learning inversion model for residual structure and long-range dependency learning to correct the initial prediction results, and outputting inversion results for multiple aerosol characteristic parameters.

[0006] Furthermore, in some embodiments, the method further includes: calculating derived indices for source analysis or event identification based on the inversion results of multiple aerosol characteristic parameters, wherein the derived indices include at least one of visible light aerosol optical thickness, fine mode fraction, combustion aerosol index, dust index, and urban aerosol index.

[0007] In some embodiments, acquiring observational data of satellite multi-band atmospheric top-layer reflectivity includes: using original satellite images of aerosol-sensitive bands in satellite multi-band atmospheric top-layer reflectivity observations as a dynamic input data source; performing geographic correction on the original satellite images based on the satellite's geographic coordinates to ensure that the pixel positions of the original satellite images are registered with the real scene; removing invalid pixels from the original satellite images using cloud masks and / or snow masks; removing pixels from the original satellite images whose solar zenith angle and observed zenith angle are both greater than a preset standard angle, and calculating the scattering angle; performing time consistency processing on the original satellite images while retaining the original data type to obtain observational data of satellite multi-band atmospheric top-layer reflectivity.

[0008] In some embodiments, the dynamic observation features include at least the multi-band atmospheric top-layer reflectivity and corresponding observation geometric parameters from the observation data, as well as the dynamic background field of aerosol optical thickness and absorbing gas information provided by the atmospheric reanalysis dataset. The observation geometric parameters include the solar zenith angle, the observation zenith angle, and the scattering angle. The dynamic background field of aerosol optical thickness serves as a priori background input during the first-stage inversion model training process, assisting the model in learning the inversion relationship of aerosol optical thickness. The static background features include spatial and temporal variables, topographic elevation, human activity proxy variables, and climate background field information such as single-scattering albedo and / or scattering phase function asymmetry factor provided by the aerosol climate dataset.

[0009] In some embodiments, the first-stage learning inversion model adopts a gradient boosting decision tree model and runs in a multi-task regression mode to perform joint inversion of multiple aerosol optical property parameters and output the initial prediction results of the corresponding multiple aerosol optical property parameters respectively. The multiple aerosol property parameters include at least: multi-band aerosol optical thickness, fine-mode aerosol optical thickness and / or coarse-mode aerosol optical thickness, multi-band single-scattering albedo and / or scattering phase function asymmetry factor.

[0010] In some embodiments, the second-stage learning inversion model is a neural network model based on a self-attention mechanism, which uses a mean squared error loss function and a graphics processor for model training, and includes residual connection layers and normalization layers for stable training.

[0011] The second aspect of this disclosure provides a multi-parameter, multi-band joint remote sensing inversion device for aerosol optical properties, comprising: a training data acquisition module, used to acquire observational data of satellite multi-band atmospheric top-layer reflectivity and ground-based observation station aerosol parameter ground-based data, and to perform spatiotemporal matching between the observational data and the ground-based aerosol parameter ground-based data; a feature construction module, used to construct dynamic observation features and static background features based on observational data and multi-source auxiliary data, wherein the multi-source auxiliary data includes atmospheric reanalysis datasets, aerosol climate datasets, topographic elevation, and human activity proxy variables; and a first-stage learning inversion module, the second... The first-stage learning inversion module uses the true aerosol parameter data as training labels, inputs dynamic observation features into the first-stage learning inversion model for strong nonlinear fitting, outputs initial prediction results for multiple aerosol characteristic parameters, and calculates statistics on the output of the first-stage learning inversion model to generate predicted statistical features. The second-stage learning inversion module uses the true aerosol parameter data as training labels, fuses static background features with predicted statistical features, and inputs them into the second-stage learning inversion model to learn residual structure and long-range dependence to correct the initial prediction results, and outputs inversion results for multiple aerosol characteristic parameters.

[0012] A third aspect of this disclosure provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the aforementioned first aspect.

[0013] A fourth aspect of this disclosure provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to perform the aforementioned first aspect.

[0014] The fifth aspect of this disclosure provides a computer program product including a computer program that, when executed by a processor, implements the first aspect described above.

[0015] Due to the adoption of the above technical solution, the beneficial effects of this disclosure include: 1. This disclosure adopts a two-stage learning inversion structure. The first-stage learning inversion model first robustly fits the dynamic strong nonlinear relationship. The second-stage learning inversion model then uses static / background features for residual correction and information fusion. Compared with a single model, it can reduce overfitting, improve cross-time / cross-site generalization, and has smaller errors and higher stability. 2. This disclosure adopts multi-task joint inversion. Since there is physical correlation between AOD, fine mode / coarse mode AOD, SSA, and AF, multi-task shared representation can improve the invertibility of weakly sensitive parameters to a certain extent, and output a more complete aerosol characterization to support aerosol type identification and source resolution. 3. This application introduces atmospheric reanalysis dataset and aerosol climate dataset as prior background input models, which makes the model more robust. That is, under conditions of observation noise, geometric changes or sparse training samples, the background field can provide reasonable constraints for the model and improve the prediction ability when ground-based observations are lacking. 4. The consistency between the 550nm AOD and ground-based measurements of this application is significantly higher than that of current official products, and it maintains a more stable correlation and smaller deviation, making it more suitable for monitoring aerosols under extreme events; 5. For bright surfaces and complex underlying surfaces, this application adopts a two-stage model data-driven fusion method, which can reduce the fragmentation / void of traditional dark-target (DT) algorithms, produce a more continuous AOD field, and facilitate the monitoring of plume / dust transport. Attached Figure Description

[0016] The above and other objects, features and advantages of this disclosure will become clearer from the following description of embodiments with reference to the accompanying drawings, in which: Figure 1 A flowchart illustrating a multi-parameter, multi-band joint remote sensing inversion method for aerosol optical properties according to an embodiment of the present disclosure is shown. Figure 2 The flowchart of AeroCatTrans, a multi-band, multi-parameter aerosol inversion model integrating machine learning CatBoost and deep learning Transformer models, is illustrated schematically according to an embodiment of the present disclosure. Figure 3 This schematic diagram illustrates a multi-parameter, multi-band joint remote sensing inversion device for aerosol optical properties according to an embodiment of the present disclosure. Figure 4 A schematic diagram of an electronic device according to an embodiment of the present disclosure is shown. Detailed Implementation

[0017] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0018] The following description, with reference to the accompanying drawings, describes an embodiment of a method, apparatus, and storage medium for multi-parameter, multi-band joint remote sensing inversion of aerosol optical properties.

[0019] Figure 1 A flowchart illustrating the multi-parameter, multi-band joint remote sensing inversion method for aerosol optical properties according to an embodiment of the present disclosure is shown.

[0020] like Figure 1 As shown, the method includes the following steps 101-104.

[0021] Step 101: Obtain the observation data of the multi-band atmospheric top reflectivity from the satellite and the true value data of the aerosol parameters from the ground-based observation station, and perform spatiotemporal matching between the observation data and the true value data of the aerosol parameters.

[0022] In some embodiments of this disclosure, obtaining observational data of satellite multi-band atmospheric top-layer reflectivity includes: using original satellite images of aerosol-sensitive bands (preferably 0.45-2.2 µm) from satellite multi-band atmospheric top-layer reflectivity observations as a dynamic input data source; performing geographic correction on the original satellite images according to the satellite's geographic coordinates to ensure that the pixel positions of the original satellite images are registered with the real scene; removing invalid pixels from the original satellite images using cloud masks and / or snow masks; removing pixels from the original satellite images whose solar zenith angle and observed zenith angle are both greater than a preset standard angle (preferably 80°) and calculating the scattering angle; performing time consistency processing on the original satellite images while retaining the original data type to obtain observational data of satellite multi-band atmospheric top-layer reflectivity.

[0023] Among them, Top-of-Atmosphere Reflectance (TOA) is used for the reflectance signal received by the satellite at the top of the atmosphere, which includes the combined effects of the surface and the atmosphere (aerosols, gas absorption); Solar Zenith Angle (SZA), Observational Zenith Angle (VZA), and Scattering Angle (SCA) are used to describe the observation geometry and affect radiative transfer.

[0024] It should be noted that, in one application scenario of this disclosure, for the current next-generation geostationary operational environmental satellites (GOES), the existing algorithm only provides single-band AOD through radiative transfer methods, and there is a significant diurnal variation bias, which limits the characterization of high temporal resolution aerosol characteristics.

[0025] In the aforementioned application scenarios, the core observational data originates from the Advanced Baseline Imager (ABI) aboard the GOES-16 (formerly known as GOES-East) geostationary satellite. The ABI is a multispectral sensor covering a broad spectral range from visible to infrared, significantly enhancing the monitoring capabilities of atmospheric aerosol load. Furthermore, since its launch, GOES-16 has undergone multiple reconfigurations of its satellite observation strategy, establishing a comprehensive scanning mode covering different regions and providing various temporal frequencies and spatial resolutions to meet diverse application requirements.

[0026] Furthermore, a TOA image with a GOES-16 / ABI panchromatic disk Level 1.0 resolution and a near-zenith resolution of approximately 2 km was selected as the data source. The aim was to acquire raw radiometric information covering multiple bands from visible to infrared light, providing a rich spectral basis for retrieving parameters such as multi-band AOD, SSA, AF, and fine / coarse mode AOD.

[0027] In some embodiments, performing time-consistency processing on the original satellite imagery while preserving the original data type includes: using hourly data for the imagery and preserving the original NetCDF (Network Common Data Form) format without any spatial interpolation or resampling, thereby maintaining data consistency and reducing computational costs.

[0028] In some embodiments, to select suitable observation conditions, the corresponding cloud and snow mask products of NWC SAF are used, and high air mass geometry cases are excluded, i.e., cases where both the solar zenith angle and the observed zenith angle are greater than or equal to 80° are excluded. Simultaneously, observation geometry parameters are used to characterize satellite observation conditions and effectively separate aerosol signals.

[0029] To verify the performance advantages of the disclosed inversion method, official operational level 2.0 aerosol products from GOES16 / ABI were simultaneously acquired as a direct comparison benchmark. This product is based on physical processes and supported by a look-up table (LUT) for multispectral inversion. Methodologically, it closely follows the MODIS / VIIRS tradition for dark targets and ocean multichannel inversion, representing the operational level of current technology. For fair comparison, only enhanced hourly official 550nm AOD products were used, preserving all quality markers of the land inversion and performing rigorous colocation matching with ground-based observation ground truth within spatial (±2 km) and temporal (±5 min) windows to form sample pairs for accuracy verification and comparative analysis.

[0030] In the embodiments of this disclosure, the true value data of aerosol parameters measured on the ground are used as reference data for training and validating the two-stage learning inversion model.

[0031] In one implementation, the Level 2.0 multi-band AOD of AERONET 3.0 (AErosol Robotic Network) is used as the training target for the model's multi-band AOD inversion task. The multi-bands include 440nm, 500nm, 675nm, 870nm, and 1020nm. AERONET sites directly measure aerosol extinction using solar photometers, which are less affected by the Earth's surface, thus providing higher-precision AOD. The uncertainty of the Level 2.0 visible band is approximately 0.01. To facilitate comparison with official AOD products, the model's multi-band output interpolation is converted to a 550nm AOD using multi-band (440nm, 500nm, 675nm, 870nm, 1020nm) polynomial interpolation.

[0032] Furthermore, the 500 nm fine-mode and coarse-mode AOD provided by the AERONET SDA Level 2.0 product were collected as training targets for the model's fine-mode and coarse-mode AOD inversion tasks, respectively.

[0033] Furthermore, aerosol scattering and absorption correlation products obtained from almucantar inversion, including SSA and AF at 440 nm, 675 nm, 870 nm, and 1020 nm, were used as training targets for the model's multi-band SSA and AF inversion tasks, respectively. It should be noted that, due to the Level 2.0 quality assurance requirement of AOD > 0.4 at 440 nm for the inversion products, many sites lack sufficient records to generate adequate samples. To achieve a balance between data quality and usability, this disclosure uses Level 1.5 quality-controlled inversion products for SSA and AF, with screening criteria including: AOD > 0.2 at 550 nm, SSA > 0.7, sky measurement fitting residual < 10%, and effective hours per day > 4.

[0034] Step 102: Based on the observation data and multi-source auxiliary data, construct dynamic observation features and static background features respectively.

[0035] The multi-source auxiliary data include atmospheric reanalysis datasets, aerosol climate datasets, topographic elevations, and proxy variables for human activities.

[0036] In some embodiments, the method further includes acquiring multi-source auxiliary data; wherein, the atmospheric reanalysis dataset provides the dynamic background field of AOD and the absorbing gases; the absorbing gases include the total ozone column concentration (TCO3) and the total water vapor column concentration (TCPW); the aerosol climate dataset provides the SSA / AF climatological background field, which is used as prior constraints / background information in the SSA / AF inversion stage; the human activity proxy variable includes population distribution (WorldPop), which is an effective proxy variable for emission intensity, and is used to extract fine modal information in the fine modal AOD / coarse modal AOD (fAOD / cAOD) inversion stage.

[0037] In one implementation, the MERRA2 (ModernEra Retrospective analysis for Research and Applications, Version 2) reanalysis dataset is collected to provide aerosol-related meteorological fields such as AOD, AE, SSA, ozone, and water vapor. Specifically, hourly average aerosol diagnostics, including 550 nm AOD, 470–870 nm Ångström exponents (AE), and 550 nm SSA calculated from total aerosol-scattered AOD, are obtained as auxiliary inputs to the learning inversion model. The data is globally covered, hourly averaged, two-dimensional grid with an original spatial resolution of 0.5° latitude × 0.625° longitude.

[0038] It should be noted that MERRA2 is a current global atmospheric reanalysis produced by NASA's Global Modeling and Assimilation Office (GMAO) for the satellite era, using the GEOS 5.12.4 assimilation system. This system assimilates data from MODIS, MISR, AERONET, and other satellites and in-situ observations, with aerosol assimilation using the GOCART module. Overall, MERRA2 performs satisfactorily, with a typical bias of approximately ~0.108–0.159 and a root mean square error (RMSE) of approximately ~0.108–0.159 compared to ground-based measurements. However, uncertainties may be greater in sparsely observed regions.

[0039] Furthermore, due to the significant uncertainties remaining in the MERRA2 reanalysis, the monthly-scale SSA and AF climate fields provided by the MACv3 (MaxPlanck Aerosol Climatology v3) monthly aerosol climatology were supplemented as auxiliary inputs to the learning inversion model, with a data spatial resolution of 1°×1°. MACv3 also provides merged visible light (550 nm) AOD and absorbed AOD fields separated by fine mode (radius <0.5 µm) and coarse mode (radius >0.5 µm), and includes the effective radius (fRE) of the fine mode. It should be noted that this merging was adjusted based on the regional monthly background map combined with local ground-based solar photometer statistics. MACv2 approximately distributes about 50% of the total AOD to each mode at 550 nm, and assigns about 70% of the total AAOD (0.0072) to the fine mode. Compared with MACv2, MACv3 shows significant improvements in accuracy and temporal coverage.

[0040] Furthermore, this disclosure also uses the hourly instantaneous total ozone (TCOZ, unit: Dobson) and total column water content (TCPW, unit: kg / m³) of MERRA2. 2 For example, `inst1_2d_asm_Nx` (resolution 0.5°×0.625°) was used as an auxiliary input to the learning inversion model to characterize the absorption effects of ozone and water vapor on satellite multispectral signals. An annual-scale 90 m digital elevation model (DEM) from the Space Shuttle Radar Altimetry Mission (SRTM) was introduced to describe surface topography, helping to capture the spatial distribution and vertical variability of aerosols, thereby improving inversion accuracy and robustness. To separate the fine model information from the total aerosol load, annual 1 km resolution WorldPop unconstrained population distribution data was also incorporated into the model.

[0041] It should be noted that, in specific implementation, the AOD prior can be replaced with other reanalysis / model products, such as CAMS (Copernicus Atmospheric Monitoring Service Reanalysis), NAAPS (Naval Aerosol Analysis and Prediction System), etc.; the SSA / AF background can be replaced with other climatological or model monthly mean fields, and this disclosure does not impose any restrictions.

[0042] In some embodiments, the dynamic observation features include at least the multi-band TOA and observation geometry in the observation data, the AOD dynamic background field provided by the atmospheric reanalysis dataset, and the absorbing gas, wherein the observation geometry includes the solar zenith angle, the observation zenith angle, and the scattering angle; the AOD dynamic background field is used as a priori background input when the first-stage learning inversion model performs the AOD inversion task.

[0043] In some embodiments, the static background features include at least spatial and temporal variables, topographic elevation, human activity proxy variables, and the SSA / AF climate background field provided by the aerosol climate dataset. The spatial variables include latitude and longitude / regional coding, which can be further constructed using three-dimensional spherical coordinates to capture spatial location relationships. The temporal variables include hours and days of year, which can be encoded using spiral triangular sequence vectors to preserve the periodicity of time. The SSA / AF climate background field serves as the prior background input for the second-stage learning inversion model, providing background information for the SSA / AF inversion task.

[0044] It should be noted that, in specific implementation, features such as land surface type / vegetation index / BRDF parameters, meteorological field including boundary layer height, wind speed, and humidity can be introduced to enrich the model's input information.

[0045] Step 103: Using the true value data of aerosol parameters as training labels, the dynamic observation features are input into the first-stage learning inversion model for strong nonlinear fitting, and the initial prediction results of multiple aerosol characteristic parameters are output. The statistics of the output of the first-stage learning inversion model are calculated to generate predictive statistical features.

[0046] It should be noted that the method disclosed herein provides a novel ensemble deep learning framework that uses ground-based measured aerosol parameter ground truth data as training labels to jointly invert a series of key aerosol properties based on satellite observation data. This framework employs a two-stage ensemble design, coupling a first-stage learning inversion model capable of handling tabular nonlinearity and a second-stage learning inversion model capable of modeling long-range dependencies, to respectively optimize the nonlinear fitting of dynamic features and the modeling of static background and long-range dependencies. Patterned aerosol information provided by atmospheric reanalysis datasets and aerosol climate datasets is also incorporated into the inversion process as background and auxiliary indicators, thereby enhancing the model's robustness and generalization ability.

[0047] In some embodiments, the first-stage learning inversion model adopts a gradient boosting decision tree model and runs in a multi-task regression mode to perform joint inversion of multiple aerosol optical property parameters and output the initial prediction results of the corresponding multiple aerosol optical property parameters respectively. The multiple aerosol property parameters include at least: multi-band AOD, fine mode and / or coarse mode AOD, multi-band SSA and / or AF.

[0048] Specifically, the first-stage learning inversion model employs a tree-based CatBoost framework to perform strong nonlinear fitting on the dynamic observation features of the input, outputting initial predictions with multiple parameters / bands, and generating statistical features for use in the second stage to avoid information loss. As a high-performance gradient boosting decision tree (GBDT) framework, CatBoost excels in handling multi-task outputs and heterogeneous inputs (especially categorical features), making it suitable for this stage.

[0049] In some embodiments, the hyperparameters of the first-stage learning inversion model are set as follows: multi-task root mean square error (RMSE) loss function, tree depth 6, early stopping 20 rounds, maximum iterations 2000, learning rate 0.1, and the model is trained using a central processing unit (CPU).

[0050] It should be noted that in specific implementations, the CatBoost model can be replaced by models that can handle tabular nonlinearity, such as XGBoost, LightGBM, Random Forest, and ExtraTrees, and this disclosure does not impose any restrictions.

[0051] In some embodiments, the maximum, minimum, and mean values ​​of the multiple outputs / multiple output vectors of the first-stage learning inversion model are calculated and used as part of the input to the second-stage learning inversion model for error correction. In specific implementations, the connection method between the two stages of learning inversion can be replaced by: the input to the second-stage learning inversion model using the full output vectors of the first-stage learning inversion model, or by connecting them in the form of residuals. For example, the final output of the overall integrated deep learning framework is the output of the first-stage learning inversion model and the prediction residuals of the second-stage learning inversion model; this disclosure does not limit this.

[0052] Step 104: The static background features and the predicted statistical features are fused and input into the second-stage learning inversion model to learn the residual structure and long-range dependence in order to correct the initial prediction results and output the inversion results of multiple aerosol characteristic parameters.

[0053] In some embodiments, the second-stage learning inversion model is a neural network model based on a self-attention mechanism, which uses the mean squared error (MSE) loss function and a graphics processing unit (GPU) for model training, and includes residual connection layers and normalization layers to stabilize training.

[0054] Specifically, the second-stage learning inversion model employs a Transformer decoder structure to efficiently extract latent aerosol features from the multi-dimensional input space and reduce the risk of overfitting. The decoder consists of stacked layers, each typically including a masked multi-head self-attention module, an encoder-decoder cross-attention module, and a feedforward network; each sub-layer uses residual connections and layer normalization. Key hyperparameters, including the number of layers, neurons, attention heads, and the size of the feedforward hidden layers, are selected through trial and error to achieve a balance between model capacity and computational cost. The ReLU activation function is used between hidden layers to improve training efficiency and promote sparse feature representation. The training process uses MSE as the loss function and is optimized on GPUs using the CUDA computing architecture to ensure stable gradient propagation.

[0055] It should be noted that, in specific implementations, architectures capable of modeling long-range dependencies, such as Encoder-Decoder Transformer, TemporalFusion Transformer, LSTM / GRU+Attention, and MLP-Mixer, can be used to replace the Transformer decoder, and this disclosure does not impose any restrictions.

[0056] In some embodiments, the method further includes: calculating derived indices for source analysis or event identification based on the inversion results of multiple aerosol characteristic parameters, wherein the derived indices include at least one of visible light aerosol optical thickness, fine mode fraction, combustion aerosol index, dust index, and urban aerosol index.

[0057] In some embodiments, based on the inversion results of multiple aerosol characteristic parameters, the calculation of derived indices for source apportionment or event identification includes: interpolating the AOD of multiple bands to obtain a 550 nm AOD, for example, using 440 and 675 nm interpolated via the Ångström index (AE); calculating the fine mode fraction based on fAOD and cAOD, using the formula FMF=fAOD / (fAOD+cAOD); calculating the combustion aerosol index based on SSA and AOD or based on band ratios, including BBI_SSA and BBI_RAOD; calculating the dust index DUI based on AE and SSA; and constructing the urban aerosol index UII based on SSA and AE. Constructing indices such as wildfire / dust / urban industrial indices enables source-related identification and change monitoring, improving the multi-parameter diagnostic capabilities for extreme events.

[0058] Specifically, the calculation formulas for BBI_SSA, BBI_RAOD, DUI, and UII are shown in Formulas 1-4 below:

[0059]

[0060]

[0061]

[0062] Wherein, BBI_SSA and BBI_RAOD are combustion aerosol indices calculated using SSA and RAOD; RAOD is the ratio of AOD in two bands (preferably 1020nm and 440nm); DUI is the dust index; and UII is the urban aerosol index. For the combustion aerosol index, K and m are slope parameters controlling the steepness of the curve (preferably both set to 10), and c1 and c2 are threshold parameters, 1.7 and 0.96 respectively; for the dust index... , , For urban aerosol index , , .

[0063] In the comprehensive implementation of the above embodiments, see Figure 2 This paper presents a multi-band, multi-parameter aerosol inversion model that integrates machine learning CatBoost and deep learning Transformer models, named... AeroCatTrans A multi-task approach was employed to jointly retrieve aerosol properties hourly based on GOES16 / ABI imagery. Five key aerosol properties were retrieved simultaneously: multi-band AODλ, fine and coarse AOD at 500 nm, SSAλ, and AFλ, where λ represents wavelength in nanometers. AERONET measurements were used as the training target, and the modeling is shown in Equation 5 below.

[0064] (Formula 5) Formula description AeroCatTrans Simultaneously invert the AOD (Aspect-Oriented Discharge) at five bands (440, 500, 675, 870, 1020 nm), the fine and coarse AOD (fAOD, cAOD) at 500 nm, and the SSA (Short-Side Array) and AF (Aspect-Oriented Array) at four bands (440, 500, 675, 870, 1020 nm) from multi-source input features, corresponding to... Figure 2 The Outputs section is shown. AeroCatTrans Input such as Figure 2 As shown in the Inputs section, the first stage of CatBoost inputs includes TOAR (reflectance at the top of the atmosphere) for multiple shortwave bands, Angles (geometric angles), Gases (absorbent gas columns), and the topographic elevation model (Model). The second stage of Transformer inputs consist of CatBoost input statistics, including maximum, minimum, and mean values ​​(max, min, mean), as well as static input.

[0065] Furthermore, in multi-task joint inversion AeroCatTrans The specific considerations for input and output are as follows: In the multi-band AOD inversion task, the 440, 500, 675, 870, and 1020 nm AOD values ​​provided by AERONET were used as training targets, denoted as... The first stage of CatBoost input includes the reflectance of five shortwave bands of the hourly GOES-16 / ABI top atmosphere, expressed as... TOAR. Five shortwave bands were selected, preferably in the range of 0.45–2.2 µm, as these bands are highly sensitive to the extinction of the total aerosol load. Since reflectivity is strongly influenced by solar and observation geometry, the input variables were scaled using the cosine of the solar zenith angle and the observed zenith angle. The scattering angle was also included because it is an important indicator of satellite signal path and aerosol extinction characteristics, especially for characterizing the aerosol scattering phase function. The solar zenith angle, observed zenith angle, and scattering angle are respectively expressed as… SZ VZ and SCA .

[0066] In addition, to provide background priors and improve inversion, 550 nm AOD of hourly MERRA2 is added to the CatBoost input, denoted as MEAOD The total column ozone and the total column precipitable water column, along with other absorbent gases, are also used as inputs to correct for their effects on... TOAR The absorption effect is mitigated, improving the model's sensitivity and accuracy for AOD in multispectral signals. Total column ozone and total column precipitation are expressed as follows: TCOZ and TCPW CatBoost models the system using the aforementioned dynamic features in the first stage. Subsequently, static auxiliary features, including terrain elevation, are incorporated. DEM Spatiotemporal variables ST The statistical outputs of CatBoost (such as maximum, minimum, and mean, denoted as...) This information is then fed into the second-stage Transformer. The Transformer's task is to correct the residuals after CatBoost prediction and further refine the CatBoost prediction structure using static information, outputting the final inversion result.

[0067] Correspondingly, the two-stage modeling of CatBoost and Transformer in multi-band AOD inversion is shown in Equations 6 and 7 below:

[0068] (Formula 6) (Formula 7) In the fine-mode and coarse-mode AOD inversion task, the 500 nm fine-mode and coarse-mode AOD measured by AERONET are used as the training targets of AeroCatTrans, denoted as . The input variables for the first stage of CatBoost are the same as those for multi-band AOD inversion, referring to Formula 2. Subsequently, the statistical outputs of CatBoost for fine-mode and coarse-mode AOD are calculated. The second-stage Transformer is input. In addition to using the static features mentioned in multi-band AOD inversion, the Transformer also incorporates population distribution (POP) as an important indicator variable of anthropogenic emissions. The second-stage Transformer model is shown in Equation 8: (Formula 8) It should be noted that the fine modality score (FMF) is not used as a joint training target because its numerical range differs greatly from that of the coarse modality score (AOD). FMF can be derived after obtaining the fine modality and coarse modality scores (AOD).

[0069] In the multi-band SSA and AF inversion task, the training target is the SSA and AF in the 440, 675, 870, and 1020 nm bands provided by the quality-controlled AERONET Level 1.5 inversion product, denoted as The input to the first stage of CatBoost includes scaled input. TOAR Geometric angles ( SZ VZ and SCA ), TCOZ and TCPW and time variables T Compared to the aforementioned inversion tasks, MERRA-2 AOD was excluded as it does not provide AF and its single-band SSA is insufficient to significantly improve multi-band SSA inversion. Subsequently, the statistical outputs of CatBoost for multi-band SSA and AF were compared. The second-stage Transformer is input. It provides background information for SSA and AF inversion. The static inputs of the second-stage Transformer, besides... DEM , ST In addition, monthly SSA and AF climatological data from MACv3 were also introduced, denoted as MASSA and MAAF Correspondingly, the two-stage modeling of CatBoost and Transformer in multi-band SSA and AF inversion tasks is shown in Equations 9 and 10 below: (Formula 9)

[0070] Further, refer to Figure 2 Spatial distribution of model parameters in CatBoost and Transformer neural network sections. AeroCatTrans The model training included: the first stage, CatBoost, was implemented on a CPU using the CatBoost library in Python, with hyperparameters set as follows: multi-task RMSE loss, tree depth N=6, early stopping 20 epochs, maximum epochs=2000, and learning rate Lr=0.1. The second stage, Transformer, employed a Transformer decoder structure to efficiently extract latent aerosol features from the multi-dimensional input space and reduce overfitting risk. The Transformer decoder consisted of stacked layers, each typically including a multi-head attention module, an encoder-decoder cross-attention module, and a feedforward network; each sub-layer employed residual connections and layer normalization (Add&Norm); it also included an input embedding layer, a fully connected layer, and a linear output layer. Figure 2 As shown, the key hyperparameters are set as follows: Encode layer number = 2, Neuron number = 128, Head number = 4, and Dropout size of the feedforward hidden layer = 0.01. ReLU activation is used between hidden layers to improve training efficiency and promote sparse feature representation. The training process uses mean squared error (MSE) as the loss function and is performed on a GPU (CUDA) using the CUDA computing architecture.

[0071] Further, refer to Figure 2 In the ten-fold crossing validation (CV) section, ten-fold cross-validation is used to verify the multi-band, multi-parameter aerosol inversion model provided in this disclosure. AeroCatTrans Objective evaluation was conducted. The validation process involved dividing all samples into ten subsets, iteratively training on nine subsets (90%), and validating on the remaining subset (10%) to obtain outoffold predictions as a baseline. To further examine the model's spatiotemporal generalization ability without ground observations, ten-fold cross-validation by site and by hour was also performed. This involved first grouping the samples by site location or observation hour, then dividing the grouping results into ten folds, and repeating the same training / validation process. Figure 2As shown, the grouping results include sample CV, time CV grouped by hour, and spatial CV grouped by station. These are then divided into ten subsets D1-D10. Training is iteratively trained on nine subsets (90%) and Testing is performed on the remaining subset (10%) to obtain ten validation results E1-E10 and their mean Mean(E).

[0072] Standard metrics were used to evaluate all aerosol parameters: correlation coefficient (R), root mean square error (RMSE), and mean absolute error (MAE). The reliability of multiband AOD was also assessed by comparing it to the expected error (EE) rectangle of the MODIS Deep Blue over land, defined as ±(0.05 + 20% × AODλ); and to the more stringent Global Climate Observation System (GCOS) requirements, defined as ±max(0.03, 10% × AODλ). For SSA and AF, the GCOS criterion was also used for further evaluation, defined as ±max(0.03, 10% × SSA / AFλ).

[0073] Furthermore, the model output parameters are expanded to include joint inversion of AE, AAOD, FMF, effective radius, etc., or direct output of event recognition indicators. Figure 2 Showing AeroCatTrans Examples of the final output inversion results and examples of tracking extreme events such as wildfires, dust storms, and haze.

[0074] Furthermore, the data results for verifying the model's accuracy using the aforementioned ten-fold cross-validation strategy are as follows: For multi-band AOD retrieval, AeroCatTrans generally demonstrates high accuracy: based on approximately 162,000 ground-based observations, the sample-based 10-fold cross-validation correlation coefficient (CV-R) for each band (440-1020 nm) is 0.92, with a fitting slope of 0.86-0.87, and the 1020 nm AOD exhibits the lowest correlation. Uncertainty is low, with RMSE ranging from 0.031 to 0.108 and MAE from 0.017 to 0.046, with the maximum uncertainty occurring in the 440 nm band. Furthermore, over 90% of the retrieval results fall within the error envelope (EE), and approximately 60% meet the GCOS requirements. Although the sensitivity and numerical dynamic range of 1020 nm decreases for aerosols, the proportion of this long band within the EE and GOES envelopes remains close to 100% and 90%, respectively. At over 80% of the sites, the inversion of each band showed high correlation coefficients (R>0.7) and low errors (RMSE<0.01), further demonstrating the reliability of AeroCatTrans in diurnal multiband AOD inversion.

[0075] In terms of fAOD and cAOD inversion, the model also demonstrated good overall accuracy: compared with approximately 126,000 ground-based observations, the sample-based CVR for fAOD and cAOD were 0.92 and 0.88, respectively, and the RMSE for fAOD and cAOD were 0.091 and 0.020, respectively. Furthermore, over 92% (fAOD) and 98% (cAOD) of the inversions were within the EE envelope, and 73% (fAOD) and 93% (cAOD) met the GCOS criteria, respectively, indicating that cAOD has higher reliability. At the site level, over 80% (fAOD) and 70% (cAOD) of the sites achieved high correlation (R>0.7) and low error (RMSE<0.01), demonstrating that the model can reliably extract fine-scale aerosol properties based on multispectral observations and population distribution information, which is helpful in assessing anthropogenic aerosol loads and disturbances.

[0076] For regions or periods lacking ground-based observations, temporal and spatial cross-validation (CV) was performed on the predictive capabilities of AeroCatTrans. In the temporal CV, even when data from certain hours were excluded for testing, the model still robustly predicted hourly AOD during the 00:00-23:00 UTC period. Specifically, the temporal CVR for each band was 0.91-0.92, and the RMSE was 0.032-0.109, approximating the sample-based validation performance; moreover, 91%-98% of the predictions fell within the EE (Extended Estimation), and 62%-86% met the more stringent GCOS (Gas Conformity Standard) requirements. This indicates that the model possesses strong hourly, multi-band AOD prediction capabilities even in periods without ground-based observations. The temporal CV results for fAOD and cAOD were similar: CVRs were 0.91 and 0.87, respectively, and RMSEs were 0.092 and 0.022, respectively, with over 90% (fAOD) and 70% (cAOD) of the predictions meeting the EE and GCOS criteria. Spatial CV shows that the model can still predict multi-band AOD well during spatial extrapolation, with spatial CVR ranging from 0.82 to 0.88 and RMSE ranging from 0.045 to 0.130. 85%-95% of the predictions fall within the EE envelope, and 53%-80% meet the GCOS range, indicating that the model has robust spatial prediction capabilities in areas without ground-based observations. Spatial CVRs are 0.88 (fAOD) and 0.78 (cAOD), with RMSEs of 0.109 and 0.028, respectively. Furthermore, over 80% (fAOD) and 60% (cAOD) of the predictions fall within the EE and GCOS ranges, indicating a high ability to spatially extrapolate fine-scale model information.

[0077] For hourly SSA retrieval, the sample-based CVR is at a moderate level (>0.5) for most hours, especially in the 440nm band. Specifically, based on 7,273 hourly ground-based observations, the sample-based CVR for each band is 0.59-0.71, and the RMSE is 0.025-0.039. In terms of practicality, the proportion within the EE and GOES envelopes is at a moderate level (48%-64%). Compared with the sample-based CV, the model's time-domain CVR is 0.60-0.72 for most hours, showing improved accuracy; the RMSE is 0.023-0.038, which is relatively lower; the proportion meeting the GCOS standard is 48%-67%, showing improvement; and the spatial CV-R for each band is 0.33-0.51, and the RMSE is 0.030-0.048. These results indicate that SSA measurements themselves have significant uncertainties and obvious intra-diurnal variations, reflecting the inherent difficulty of retrieving this parameter.

[0078] In multi-band AF inversion, the model significantly outperformed SSA inversion, with minimal degradation in temporal and spatial validation performance. Even in unobserved areas, the long-band CV-R remained around 0.85, demonstrating excellent robustness and generalization ability. Specifically, based on 7,293 ground-based observations, the sample-based CVR remained stable and high, ranging from 0.89 to 0.92, while the RMSE remained low, ranging from 0.020 to 0.034. Furthermore, over 80% of the stations exhibited both high correlation (R > 0.7) and low error (RMSE < 0.12) in AF inversion, proving the robustness of this method in AF inversion. The hourly time-based CV of AF remained relatively stable during the 00:00–23:00 UTC period, showing only a slight decrease but still maintaining high correlation (R = 0.86–0.91) and low error (RMSE = 0.026–0.034), indicating strong predictive ability of the model for AF during periods of sparse observation. In the airspace CV, the model's ability to spatially extrapolate AF is also satisfactory. In the long band, it can still achieve the highest airspace CVR of about 0.85 and the lowest RMSE of 0.029, indicating that it can provide relatively reliable multi-band AF estimates even in areas lacking ground stations.

[0079] Furthermore, the deep ensemble model disclosed herein has been verified. AeroCatTrans It also outperforms models using CatBoost or Transformer alone in overall accuracy, with a correlation R improvement of 16%-18% for different aerosol parameters.

[0080] In summary, the accuracy verification of AeroCatTrans shows that AeroCatTrans achieves high accuracy in hourly, multi-band AOD, and fine and coarse model AOD inversion, and also demonstrates robustness even in multi-band SSA and AF inversion, which are inherently more difficult. For regions or periods lacking ground-based observations, rigorous time-domain and spatial cross-validation proves that the model has strong time-series prediction and spatial extrapolation capabilities, making it particularly suitable for monitoring in areas with sparse ground-based stations. The overall inversion accuracy of the CatBoost and Transformer deep integration architecture is significantly better than that of a single model, directly demonstrating that the method disclosed in this paper represents a clear technological advancement.

[0081] Figure 3 This schematically illustrates a multi-parameter, multi-band joint remote sensing inversion device for aerosol optical properties according to an embodiment of the present disclosure, such as... Figure 3 As shown, the device includes: The training data acquisition module 210 is used to acquire the observation data of the reflectivity of the upper atmosphere of the satellite multi-band atmosphere and the true value data of the aerosol parameters of the ground-based observation station, and to perform spatiotemporal matching between the observation data and the true value data of the aerosol parameters. The feature construction module 220 is used to construct dynamic observation features and static background features based on observation data and multi-source auxiliary data, respectively. The multi-source auxiliary data includes atmospheric reanalysis dataset, aerosol climate dataset, topographic elevation and human activity proxy variables. The first-stage learning inversion module 230 is used to use the true value data of aerosol parameters as training labels, input dynamic observation features into the first-stage learning inversion model for strong nonlinear fitting, output the initial prediction results of multiple aerosol characteristic parameters, and calculate the statistics of the output of the first-stage learning inversion model to generate predictive statistical features. The second-stage learning inversion module 240 is used to use the true value data of aerosol parameters as training labels, fuse static background features with predicted statistical features, and input them into the second-stage learning inversion model to learn the residual structure and long-range dependence, so as to correct the initial prediction results and output the inversion results of multiple aerosol characteristic parameters.

[0082] It should be noted that the specific implementation methods of the embodiments disclosed herein are different from those of the embodiments in this paper. Figure 1 The principle of the embodiments shown is the same, and will not be repeated here.

[0083] Figure 4 A schematic block diagram of an example electronic device 300 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0084] like Figure 4 As shown, device 300 includes a computing unit 301, which can perform various appropriate actions and processes based on a computer program stored in ROM (Read-Only Memory) 302 or a computer program loaded from storage unit 308 into RAM (Random Access Memory) 303. RAM 303 can also store various programs and data required for the operation of device 300. The computing unit 301, ROM 302, and RAM 303 are interconnected via bus 304. I / O (Input / Output) interface 305 is also connected to bus 304.

[0085] Multiple components in device 300 are connected to I / O interface 305, including: input unit 306, such as keyboard, mouse, etc.; output unit 307, such as various types of monitors, speakers, etc.; storage unit 308, such as disk, optical disk, etc.; and communication unit 309, such as network card, modem, wireless transceiver, etc. Communication unit 309 allows device 300 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0086] The computing unit 301 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 301 include, but are not limited to, CPUs (Central Processing Units), GPUs (Graphics Processing Units), various special-purpose AI (Artificial Intelligence) computing chips, various computing units running machine learning model algorithms, DSPs (Digital Signal Processors), and any suitable processor, controller, microcontroller, etc. The computing unit 301 performs the various methods and processes described above, such as the multi-parameter multi-band joint remote sensing inversion method for aerosol optical properties. For example, in some embodiments, the multi-parameter multi-band joint remote sensing inversion method for aerosol optical properties can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as storage unit 308. In some embodiments, part or all of the computer program can be loaded and / or installed on device 300 via ROM 302 and / or communication unit 309. When the computer program is loaded into RAM 303 and executed by computing unit 301, one or more steps of the method described above can be performed. Alternatively, in other embodiments, computing unit 301 can be configured by any other suitable means (e.g., by means of firmware) to perform the aforementioned multi-parameter, multi-band joint remote sensing inversion method for aerosol optical properties.

[0087] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, FPGAs (Field Programmable Gate Arrays), ASICs (Application-Specific Integrated Circuits), ASSPs (Application-Specific Standard Products), SOCs (System-on-Chips), CPLDs (Complex Programmable Logic Devices), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0088] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0089] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, RAM, ROM, EPROM (Electrically Programmable Read-Only Memory) or flash memory, optical fiber, CD-ROM (Compact Disc Read-Only Memory), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0090] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (Cathode-Ray Tube) or LCD (Liquid Crystal Display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0091] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include LANs (Local Area Networks), WANs (Wide Area Networks), the Internet, and blockchain networks.

[0092] Computer systems can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. A server can be a cloud server, also known as a cloud computing server or cloud host, a hosting product within the cloud computing service ecosystem, addressing the shortcomings of traditional physical hosts and VPS (Virtual Private Server, or simply "VPS") services, such as high management difficulty and weak business scalability. Servers can also be servers for distributed systems or servers incorporating blockchain technology.

[0093] It's important to note that artificial intelligence (AI) is the study of enabling computers to simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, and planning). It encompasses both hardware and software technologies. AI hardware technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, and big data processing. AI software technologies primarily include computer vision, natural language processing, machine learning / deep learning, big data processing, and knowledge graph technologies.

[0094] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0095] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A multi-parameter, multi-band joint remote sensing inversion method for aerosol optical properties, characterized in that, The method includes: Acquire observational data of multi-band atmospheric top reflectivity from satellites and true aerosol parameter data from ground-based observation stations, and perform spatiotemporal matching between the observational data and the true aerosol parameter data; Based on the observation data and multi-source auxiliary data, dynamic observation features and static background features are constructed respectively. The multi-source auxiliary data includes atmospheric reanalysis dataset, aerosol climate dataset, topographic elevation and human activity proxy variables. Using the true values ​​of the aerosol parameters as training labels, the dynamic observation features are input into the first-stage learning inversion model for strong nonlinear fitting, and the initial prediction results of multiple aerosol characteristic parameters are output. The statistics of the output of the first-stage learning inversion model are calculated to generate predictive statistical features. The static background features and the predicted statistical features are fused and input into the second-stage learning inversion model to learn the residual structure and long-range dependence, so as to correct the initial prediction results and output the inversion results of the multiple aerosol characteristic parameters.

2. The multi-parameter, multi-band joint remote sensing inversion method for aerosol optical properties according to claim 1, characterized in that, The method further includes: Based on the inversion results of the multiple aerosol characteristic parameters, derived indices for source analysis or event identification are calculated. These derived indices include one or more of the following: visible light aerosol optical thickness, fine mode fraction, combustion aerosol index, dust index, and urban aerosol index.

3. The multi-parameter, multi-band joint remote sensing inversion method for aerosol optical properties according to claim 1, characterized in that, The acquisition of satellite multi-band atmospheric top-layer reflectivity observation data includes: The original satellite images of the relatively sensitive aerosol bands in the multi-band atmospheric top reflectivity observations were used as the dynamic input data source. The original satellite imagery is geocorrected based on the satellite's geographic coordinates to ensure that the pixel positions of the original satellite imagery are registered with the real scene. Invalid pixels in the original satellite imagery are removed using cloud masking and / or snow masking. Pixels in the original satellite image whose solar zenith angle and observed zenith angle are both greater than a preset standard angle are removed, and the scattering angle is calculated; The original satellite imagery is processed for temporal consistency while retaining its original data type to obtain observational data on the reflectivity of the upper atmosphere in multiple bands.

4. The multi-parameter, multi-band joint remote sensing inversion method for aerosol optical properties according to claim 3, characterized in that, The dynamic observation features include at least the multi-band atmospheric top reflectivity and observation geometry in the observation data, the dynamic background field of aerosol optical thickness provided by the atmospheric reanalysis dataset, and absorbing gases. The observation geometry includes the solar zenith angle, the observation zenith angle, and the scattering angle. The aerosol optical thickness dynamic background field is used as a priori background input when learning the inversion model in the first stage to perform the aerosol optical thickness inversion task. The static background features include at least spatial and temporal variables, the topographic elevation, the human activity proxy variable, and the single scattering albedo and / or scattering phase function asymmetric factor climate background field provided by the aerosol climate dataset.

5. The multi-parameter, multi-band joint remote sensing inversion method for aerosol optical properties according to claim 1, characterized in that, The first-stage learning inversion model employs a gradient boosting decision tree model and operates in multi-task regression mode. It is used to jointly invert the multiple aerosol optical property parameters and outputs the corresponding initial prediction results for each of the multiple aerosol optical property parameters. The multiple aerosol characteristic parameters include at least: multi-band aerosol optical thickness, fine-mode aerosol optical thickness and / or coarse-mode aerosol optical thickness, multi-band single-scattering albedo and / or scattering phase function asymmetry factor.

6. The multi-parameter, multi-band joint remote sensing inversion method for aerosol optical properties according to claim 1, characterized in that, The second-stage learning inversion model is a neural network model based on the self-attention mechanism. It uses the mean squared error loss function and a graphics processor for model training, and includes residual connection layers and normalization layers to stabilize the training.

7. A multi-parameter, multi-band joint remote sensing inversion device for aerosol optical properties, characterized in that, The device includes: The training data acquisition module is used to acquire observation data of satellite multi-band atmospheric top reflectivity and ground-based observation station aerosol parameter true data, and to perform spatiotemporal matching of the observation data and the aerosol parameter true data. The feature construction module is used to construct dynamic observation features and static background features based on the observation data and multi-source auxiliary data, respectively. The multi-source auxiliary data includes atmospheric reanalysis dataset, aerosol climate dataset, topographic elevation and human activity proxy variables. The first-stage learning inversion module is used to use the true value data of the aerosol parameters as training labels, input the dynamic observation features into the first-stage learning inversion model for strong nonlinear fitting, output the initial prediction results of multiple aerosol characteristic parameters, and calculate the statistics of the output of the first-stage learning inversion model to generate predictive statistical features. The second-stage learning inversion module is used to use the true aerosol parameter data as training labels, fuse the static background features with the predicted statistical features, and input them into the second-stage learning inversion model to learn the residual structure and long-range dependence, so as to correct the initial prediction results and output the inversion results of the multiple aerosol characteristic parameters.

8. An electronic device, characterized in that, include: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method according to any one of claims 1 to 6.

10. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 6.

Images