An aerosol optical depth acquisition system and method based on multi-source data fusion

Abstract

The application discloses a kind of aerosol optical depth acquisition system and method based on multi-source data fusion in atmospheric remote sensing technical field, comprising: after obtaining multi-source heterogeneous aerosol optical depth data, pre-processing is carried out;Multi-modal feature extraction is carried out to the multi-source heterogeneous aerosol optical depth data after pre-processing;The multi-modal features after extraction are fused by introducing attention mechanism;The final fused aerosol optical depth data is obtained by inputting the feature vector after fusion into multilayer perception machine;The final fused aerosol optical depth data is filled in missing data.This application improves the reliability and spatial integrity of AOD data by deeply fusing multi-source satellite AOD data, reanalysis data and AERONET ground observation data to construct spatially fully covered high-precision AOD data.

Description

Technical Field

[0001] This invention relates to an aerosol optical thickness acquisition system and method based on multi-source data fusion, belonging to the field of atmospheric remote sensing technology. Background Technology

[0002] Aerosols, as one of the major sources of air pollution, can cause smog, acid rain, and other polluting weather, reduce ground visibility, and pose a threat to human health. Aerosol optical depth (AOD) is the most important optical property of aerosols, used to describe the aerosol content in the atmosphere, and is a core parameter for assessing the environmental and climate effects of aerosols.

[0003] In existing technologies, most solutions for obtaining aerosol optical thickness through satellite remote sensing focus on improving single satellite products or using relatively simple network structures for fusion, failing to fully consider and utilize the heterogeneous characteristics of different data sources in terms of spatial structure, time series, and data format. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide an aerosol optical thickness acquisition system and method based on multi-source data fusion. By deeply fusing multi-source satellite AOD data, reanalysis data, and AERONET ground observation data, high-precision AOD data with full spatial coverage is constructed, thereby improving the reliability and spatial integrity of AOD data.

[0005] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0006] In a first aspect, the present invention provides a method for obtaining aerosol optical thickness based on multi-source data fusion, comprising:

[0007] Preprocessing is performed after acquiring the optical thickness data of multi-source heterogeneous aerosols.

[0008] Multimodal feature extraction was performed on the preprocessed optical thickness data of multi-source heterogeneous aerosols.

[0009] An attention mechanism is introduced to fuse the extracted multimodal features;

[0010] The fused feature vectors are input into a multilayer perceptron to obtain the final fused aerosol optical thickness data.

[0011] Missing data were filled in for the final fused aerosol optical thickness data.

[0012] Furthermore, after acquiring the optical thickness data of the multi-source heterogeneous aerosols, preprocessing is performed, including:

[0013] Multi-source aerosol optical thickness data and auxiliary data were collected. All data underwent quality control, spectral matching, spatial matching, temporal synchronization, and standardization to construct a spatiotemporally unified multi-source dataset.

[0014] The quality control includes cleaning up the data and removing invalid values, wherein: the official quality label is used to filter invalid values ​​for FY4B data, and the QA mark is used to filter and retain high-quality data for MODIS data.

[0015] The spectral matching includes: calculating the aerosol optical thickness values ​​of different satellite AOD products in a specified band using an aerosol optical thickness value calculation formula, wherein the aerosol optical thickness value calculation formula is:

[0016]

[0017] in, Represents the optical thickness value of the aerosol. 1 represents the band, and α and β are auxiliary calculation coefficients;

[0018] The spatial matching includes: unifying all grid data to the same coordinate system, and resampling the grid data using bilinear interpolation. The calculation method for bilinear interpolation is as follows:

[0019]

[0020]

[0021]

[0022] In the formula: Z(x,y) represents the AOD value of point (x,y), Z(x,y1) represents the AOD value of point (x,y1), and Z(x,y2) represents the AOD value of point (x,y2). , , , Let x be the AOD value of four adjacent known points, x be the center longitude coordinate of the target grid, y be the center latitude coordinate of the target grid, x1 and x2 be the longitude coordinates of the adjacent known points, and y1 and y2 be the latitude coordinates of the adjacent known points.

[0023] For discrete AERONET ground observation data, inverse distance-weighted interpolation is used to assist spatial matching, as shown in the following formula:

[0024]

[0025] In the formula, For the observation value of the i-th AERONET site, Let k be the distance from the target grid to the i-th station, k be the distance attenuation coefficient, and n be the number of neighboring stations participating in the interpolation.

[0026] The time synchronization includes: selecting a specified time period before and after the hour as a time window, calculating the average value of AERONET AOD and MODIS AOD data within the time window, and unifying the time resolution of multi-source data to an hourly rate based on the spatially matched data;

[0027] The standardization process includes:

[0028]

[0029]

[0030]

[0031] In the formula: For the standardized results, Let be the original value of the i-th sample from the j-th data source. Let be the mean of the j-th data source. Let be the standard deviation of the j-th data source. The total number of valid samples after removing outliers from the j-th data source.

[0032] Furthermore, the multi-source AOD data includes: the official AOD product FY-4B AOD of Fengyun-4B satellite, the MODIS Level 2 AOD product MOD04 L2 AOD, the MERRA-2 AOD reanalysis product MERRA-2 AOD, and the AERONET Level 2.0 product AERONET AOD;

[0033] The auxiliary data include: SRTM 90m resolution digital elevation model (DEM); ERA5 meteorological data, including boundary layer height (BLH), total column water (TCW), surface pressure (SP), 2-meter temperature (T2M), 10-meter easterly wind component (10UW), 10-meter northerly wind component (10VW), and relative humidity (RH); quality assurance marks or pixel-level uncertainty estimates for each satellite product.

[0034] Furthermore, multimodal feature extraction was performed on the preprocessed multi-source heterogeneous aerosol optical thickness data, including: spatiotemporal feature extraction from satellite raster data and AERONET AOD feature extraction, wherein:

[0035] The extraction of spatiotemporal features from satellite raster data includes: extracting spatial and temporal features from FY4BAOD and MOD04 AOD based on a 3D convolutional neural network (CNN);

[0036] The AERONET AOD feature extraction includes: modeling sparse data relationships of ground-based sites based on graph neural networks (GNNs), specifically including:

[0037] A dynamic aerosol map is constructed, with all AERONET sites regarded as nodes of the graph. The node characteristics are its AOD measurement value, timestamp, and site location information.

[0038] Defining edge connections and weights: The edge weight between two sites is defined as follows:

[0039]

[0040] Where: w_ij is the edge weight between stations i and j, dist_ij is the distance between stations i and j, and wind_vector_ij is the average wind field vector between stations i and j;

[0041] Graph Convolution Operation: Dynamic aerosol map processing is performed using a GNN model. The GNN updates the feature representation of each node by aggregating information from neighboring nodes. After multi-layer GNN processing, the feature vector of each node not only contains its own measurement information but also incorporates information from other high-precision stations associated with it through the physical path. The graph neural network feature fusion formula is as follows:

[0042]

[0043] in, Is node v at the th The feature vector of the layer, Is node v in the 1st- The feature vector of the layer, Here, W is the activation function, W is the weight matrix, and u is the neighbor node of v. It is the set of neighbors of node v, d u is the degree of node u, and b is the bias.

[0044] Furthermore, an attention mechanism is introduced to fuse the extracted multimodal features, including attention weight calculation and weighted fusion, wherein:

[0045] The attention weight calculation includes: designing an attention network that uses the location information and time information of the target point p, as well as the preliminary features extracted from the CNN, as the "query," and using the feature vector from the FY-4B CNN branch. Feature vectors from the MODIS CNN branch "Influence" feature vectors from GNN modules As "keys" and "values", attention networks learn to calculate three weights. , , And the sum of the three is 1;

[0046] The weighted fusion includes obtaining the final fused feature vector through weighted summation. The expression is:

[0047]

[0048] In the formula: This is the final fused feature vector. , , There are three weights, and the sum of the three is 1. The feature vectors are from the FY-4B CNN branch. The feature vectors are from the MODIS CNN branch. This is the "influence" feature vector from the GNN module.

[0049] Furthermore, the fused feature vector is input into a multilayer perceptron to obtain the final fused aerosol optical thickness data, including: inputting the fused feature vector... A simple multilayer perceptron is input as the prediction head, and the final fused AOD value is output. The model is trained using a composite loss function.

[0050]

[0051] In the formula: This is the loss predicted by AOD. It represents the loss due to uncertain prediction, where λ is a hyperparameter.

[0052] Furthermore, missing data imputation is performed on the final fused aerosol optical thickness data, including: sample dataset construction, model building, and data imputation, wherein:

[0053] The construction of the sample dataset includes: merging the AOD data The MERRA-2 AOD data, ERA5 meteorological data, DEM topographic data, and topographic feature data are integrated to construct the model input dataset;

[0054] The model construction includes: taking the fused AOD as the target output, MERRA-2 AOD, meteorological parameters, DEM, and DEM topographic features. The LightGBM model is trained using the input data to construct the fused AOD data. The mapping relationship between MERRA-2 AOD data and meteorological parameters and other variables is as follows:

[0055]

[0056] In the formula: AOD is the final output AOD data. For the fused AOD data, DEM represents topographic data, BLH represents boundary layer height, TCW represents total column water, SP represents surface pressure, T2M represents 2-meter temperature, 10UW represents the 10-meter easterly component, 10VW represents the 10-meter northerly component, and RH represents relative humidity. This is terrain feature data;

[0057] The data imputation includes: imputing missing values ​​in the fused data based on the trained LightGBM model to obtain spatially full-coverage AOD data.

[0058] Secondly, the present invention provides an aerosol optical thickness acquisition system based on multi-source data fusion, comprising:

[0059] Data processing module: Preprocesses the optical thickness data of multi-source heterogeneous aerosols after acquisition;

[0060] Feature extraction module: Performs multimodal feature extraction on the preprocessed multi-source heterogeneous aerosol optical thickness data;

[0061] Feature fusion module: This module fuses the extracted multimodal features by introducing an attention mechanism.

[0062] Multilayer sensing module: Input the fused feature vector into the multilayer perceptron to obtain the final fused aerosol optical thickness data;

[0063] Data imputation module: Imputes missing data in the final fused aerosol optical thickness data.

[0064] Thirdly, the present invention provides an aerosol optical thickness acquisition device based on multi-source data fusion, including a processor and a storage medium;

[0065] The storage medium is used to store instructions;

[0066] The processor is configured to operate according to the instructions to perform the steps of the method according to any of the foregoing.

[0067] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0068] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0069] I. This invention constructs a spatiotemporal attention map fusion network to deeply fuse multi-source satellite AOD data, MERRA-2 AOD data, and AERONET AOD data to reduce overall error. Combining the spatiotemporal continuity advantage of MERRA-2 AOD, spatial completion is performed on the fused AOD product to generate a high-precision spatiotemporally continuous AOD product. This solution innovatively designs a multimodal attention mechanism to achieve dynamic weight allocation of heterogeneous data, and solves the data missing problem through a spatiotemporal interpolation network, significantly improving the spatial coverage of AOD data, especially suitable for complex terrain areas and data-sparse regions.

[0070] Second, this solution effectively corrects the systematic bias of satellite products by deeply fusing high-precision AERONET data and using GNN to model its physical propagation, significantly improving the accuracy of fused AOD products nationwide. Compared with single satellite products or traditional fusion methods, the RMSE of this invention is lower, the R² value with AERONET is higher, and it has higher accuracy and reliability.

[0071] Third, this solution combines the high temporal resolution of FY-4B, the high-quality data of MODIS, and the spatiotemporal continuity of MERRA-2 AOD to generate spatiotemporally continuous AOD products, effectively filling data gaps. Intelligent uncertainty perception: The uncertainty of the data source is used as model input, and the weights of different data sources are dynamically adjusted through an attention mechanism, making the fusion process more intelligent and robust. The model can also output the uncertainty of each predicted value, providing crucial quality assessment information for subsequent applications.

[0072] Fourth, by introducing physical constraints based on meteorological information such as wind fields into the GNN, the model not only fits the data but also learns the physical laws of aerosol transport. Therefore, it has stronger generalization ability and prediction accuracy in sparse areas of AERONET sites. In addition, the STAG-FuseNet framework has good scalability and can be easily connected to more satellite data sources such as VIIRS and GEMS in the future. It is only necessary to add corresponding parallel branches to the CNN feature extraction module. Attached Figure Description

[0073] The accompanying drawings, which form part of this specification, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0074] Figure 1 This is a flowchart illustrating a method for obtaining aerosol optical thickness based on multi-source data fusion, as provided in Embodiment 1 of the present invention. Detailed Implementation

[0075] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0076] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this invention is for describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0077] To address the need for constructing spatially continuous, high-quality aerosol optical thickness (AOD) data, this solution proposes a multi-source aerosol AOD data fusion method based on a Spatio-Temporal Attention Graph Fusion Network (STAG-FuseNet). This method utilizes an innovative deep learning model architecture comprising three parallel feature extraction modules, a multimodal attention fusion module, and a prediction output module to achieve multi-source heterogeneous data fusion, thereby improving the accuracy of AOD products. Subsequently, based on MERRA-2 AOD data, combined with ERA5 meteorological parameters, DEM, and other auxiliary data, spatial completion is performed on the fused AOD data to obtain a high-precision aerosol AOD product with full spatial coverage. The overall technical approach is as follows:

[0078] (1) Preparation of multi-source heterogeneous data. First, multi-source heterogeneous data are preprocessed. Multi-source satellite AOD data (geostationary satellite AOD data, polar-orbiting satellite AOD data), ground-based observation AOD data (AERONET AOD), reanalysis AOD data (MERRA-2 AOD), and auxiliary data (ERA5 meteorological data, DEM data) are collected. All data are subjected to quality control, spectral matching, time synchronization, spatial matching, and standardization to construct a spatiotemporally unified multi-source dataset for subsequent fusion processing.

[0079] (2) Multimodal feature extraction based on parallel networks. Then, multimodal feature extraction is performed for different AOD data sources. Spatiotemporal features of satellite raster data are extracted based on convolutional neural network 3D-CNN, sparse data relationship modeling of ground-based stations is performed based on graph neural network GNN, and the dimensionality unification of heterogeneous features is achieved through feature alignment module.

[0080] (3) Feature fusion based on multimodal attention. Secondly, an attention mechanism is introduced to effectively fuse the surface features extracted by CNN and the point features extracted by GNN, thereby fusing features from different data sources (such as geostationary satellite AOD, polar-orbiting satellite AOD, and ground-based station AOD) to generate a unified, high-precision AOD feature vector for subsequent AOD prediction and uncertainty quantification.

[0081] (4) Fusing AOD output. Next, a simple multilayer perceptron (MLP) is constructed, with the fused feature vector as input and the final fused AOD value as output.

[0082] (5) AOD data imputation. Finally, the fused AOD data is further imputed based on the LightGBM model. For the data missing areas that still exist after fusion, the LightGBM model is used in combination with auxiliary data based on the spatially continuous reanalysis data MERRA-2 AOD to imput the data, and a high-precision AOD data product with complete spatial coverage is output.

[0083] Example 1:

[0084] The specific embodiments of the present invention will be further described in detail below with reference to the examples and accompanying drawings, such as... Figure 1 As shown in the figure, this embodiment designs an aerosol optical thickness inversion optimization method based on multi-source data fusion. The specific steps are as follows:

[0085] S1: Multi-source data preprocessing. The collected multi-source AOD data includes: the official AOD product of Fengyun-4B satellite (FY-4BAOD), the MODIS Level 2 AOD product (MOD04 L2 AOD), the MERRA-2 AOD reanalysis product (MERRA-2 AOD), and the AERONET Level 2.0 product (AERONET AOD).

[0086] In addition, the collected auxiliary data include: SRTM 90m resolution digital elevation model (DEM); ERA5 meteorological data, including boundary layer height (BLH), total column water (TCW), surface pressure (SP), 2-meter temperature (T2M), 10-meter easterly component (10UW), 10-meter northerly component (10VW), relative humidity (RH), and other meteorological parameters; quality assurance marks or pixel-level uncertainty estimates for each satellite product.

[0087] S1-1: Quality Control. The raw data is cleaned by removing invalid values ​​to ensure all parameters are within the valid range. Quality control is performed on satellite AOD data. For FY4B data, official quality labels (including cloud and snow labels) are used to filter invalid values. For MODIS data, the QA label is used for filtering, retaining high-quality data (QA=3).

[0088] S1-2: Spectral Matching. Due to band differences between AOD products from different satellites, spectral matching is required before fusion processing. FY4B and MODIS data both provide AOD at 550 nm, while the AOD product from the AERONET site does not include this band. Therefore, band conversion is needed using the Angstrom index for the 500-675 nm band. First, substitute the AOD values ​​for the 500 nm and 675 nm bands provided by the AERONET product into the following formula to calculate β and α; then, substitute the wavelength 550 nm into the following formula again to obtain the AOD value for the 550 nm band.

[0089]

[0090] in, Represents the optical thickness value of the aerosol. 1 represents the band, and α and β are auxiliary calculation coefficients.

[0091] S1-3: Spatial Matching. Using FY4B AOD data as a reference, all grid data were unified to the same coordinate system (WGS84), and bilinear interpolation was used to resample the grid data to a spatial resolution of 10 km for subsequent matching. The calculation method for bilinear interpolation is as follows:

[0092] Set the center coordinates of the target grid Where x is longitude and y is latitude. Find the bounding box in the original data. The coordinates of the four adjacent known points are as follows: , (Same latitude, adjacent longitude) , (Same latitude, adjacent longitude), the corresponding raw data value is , , , .

[0093] First, linear interpolation is performed along the longitude direction to obtain... and The median value at:

[0094]

[0095]

[0096] In the formula: Z(x,y1) represents the AOD value of point (x,y1), and Z(x,y2) represents the AOD value of point (x,y2). , , , Let x be the AOD value of four adjacent known points, x be the center longitude coordinate of the target grid, y be the center latitude coordinate of the target grid, x1 and x2 be the longitude coordinates of the adjacent known points, and y1 and y2 be the latitude coordinates of the adjacent known points.

[0097] Then, linear interpolation is performed along the latitudinal direction to obtain the target point. Value:

[0098]

[0099] In the formula: Z(x,y) represents the AOD value of point (x,y).

[0100] The weighting coefficient is calculated inversely proportional to the distance between the target point and its neighboring points to ensure spatial smoothness.

[0101] For discrete AERONET ground observation data, inverse distance-weighted (IDW) interpolation is used to assist spatial matching, as shown in the following formula:

[0102]

[0103] In the formula, For the observation value of the i-th AERONET site, is the distance from the target grid to the i-th station, k is the distance attenuation coefficient (2 in this embodiment), and n is the number of neighboring stations participating in the interpolation (the 8 nearest ones).

[0104] S1-4: Time Synchronization. MERRA-2 and ERA5 data are observed hourly, while FY4B AOD has a time resolution of 15 minutes. Multi-source data at the same hourly interval are matched. AERONET AOD and MODIS AOD data have variable update times; therefore, a 15-minute window before and after each hourly interval is selected as the time window. The average value of AERONET AOD and MODIS AOD data within this time window is calculated. Using the spatially matched data, the time resolution of the multi-source data is unified to one hourly, thus achieving time synchronization.

[0105] S1-5: Standardization Processing. In multi-source data fusion, the physical meanings and numerical ranges of different data sources vary significantly (e.g., AOD values ​​are typically between 0 and 5, ERA5 relative humidity is 0-100%, and DEM elevation can range from -400 to 8000 meters). Directly inputting these values ​​into the model can lead to variables with large numerical ranges dominating feature learning, masking other crucial information. By performing Z-score standardization on different data sources, the data is converted into a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating the dimensional differences between different data points.

[0106]

[0107] in, This represents the original value of the i-th sample from the j-th data source. The mean of the j-th data source reflects the central tendency of the data; Let be the standard deviation of the j-th data source, reflecting the degree of dispersion of the data; The result is standardized, dimensionless, with a mean of 0 and a standard deviation of 1.

[0108]

[0109] in, The total number of valid samples after removing outliers from the j-th data source.

[0110]

[0111] Let be the standard deviation of the j-th data source, and use the sample standard deviation, taking the denominator as . To avoid underestimating the degree of dispersion.

[0112] S2: Multimodal Feature Extraction. Dedicated feature extraction networks are designed to address the characteristics of different data sources. Spatiotemporal features of FY4B and MODIS satellite raster data are extracted using Convolutional Neural Networks (CNNs). Relationships of sparse data from AERONRT ground-based stations are modeled using Graph Neural Networks (GNNs). A feature alignment module unifies the dimensions of heterogeneous features, providing a foundational feature representation for subsequent fusion.

[0113] S2-1: Spatiotemporal Feature Extraction from Satellite Raster Data. Based on a 3D-CNN convolutional neural network, features in both the spatial and temporal dimensions of FY4BAOD and MOD04 AOD are extracted. 3D-CNN can extract features simultaneously from both temporal and spatial dimensions, effectively capturing dynamic processes such as aerosol transport and diffusion.

[0114] The module stacks AOD data from FY-4B and MODIS, along with the corresponding observation geometry (satellite zenith angle), quality indicators / uncertainty values, and other rasterized auxiliary data into a multi-channel input. For FY-4B, the input tensor has dimensions (T, H, W, C), where T is the time step (four consecutive 15-minute observations), H and W are the spatial dimensions, and C is the number of feature channels (AOD, satellite zenith angle, data quality indicator). The output of this module is a depth feature vector representing the spatial and temporal neighborhood of each grid point.

[0115] S2-2: AERONET AOD Feature Extraction. Modeling sparse data relationships at ground-based sites using graph neural networks (GNNs).

[0116] A dynamic aerosol graph is constructed, treating all AERONET sites as nodes in the graph. Node characteristics include their AOD measurement value, timestamp, and site location information.

[0117] Defining edge connectivity and weights: The graph edges are not only based on geographical distance, but also innovatively incorporate meteorological information. The edge weight between two stations is defined as follows:

[0118]

[0119] Where w_ij is the edge weight between stations i and j, dist_ij is the distance between stations i and j, and wind_vector_ij is the average wind vector between stations i and j. This means that if station j is located upwind of station i, the connection weight between them will be higher. This allows the graph to simulate the physical transport path of aerosols.

[0120] Graph convolution operation: The graph is processed using a GNN model. The GNN updates the feature representation of each node by aggregating information from its neighbors. After multiple layers of GNN processing, each node's feature vector not only contains its own measurement information but also incorporates information from other high-precision sites associated with it via physical paths. This allows ground truth information to propagate to sparse data regions in a physically consistent manner.

[0121] The feature fusion formula for graph neural networks is as follows:

[0122]

[0123] in, Is node v at the th The eigenvectors of the layer, where W is the weight matrix. It is an activation function, where u is a neighboring node of v. It is the set of neighbors of node v, d uis the degree of node u (i.e., the number of neighbors), and b is the bias.

[0124] S3: Feature fusion based on multimodal attention. An attention mechanism is introduced to solve the problem of effectively fusing surface features extracted by CNN and point features extracted by GNN.

[0125] Suppose we have a query vector Q, a set of keys K, and corresponding values ​​V. In attention mechanisms, attention weights are typically calculated based on a similarity function and normalization operations. The basic formula is as follows:

[0126]

[0127]

[0128] in, It is the attention weight of the i-th query to the j-th key, reflecting the degree of contribution of the key to the query; Let be the similarity function, representing the similarity score between the i-th query and the j-th key; This is the result of attention, where N represents the number of vectors. Represents the i-th query vector. Represents the j-th key. This represents the j-th value.

[0129] The formula for multimodal feature fusion is as follows:

[0130]

[0131] in, It is attention weight. It is a value vector. This formula describes how to perform a weighted summation of the value vector using attention weights to generate the final output.

[0132] For each target grid point p that needs to have its AOD predicted, its fusion feature consists of three contributions:

[0133] a. Feature vectors from the FY-4B CNN branch .

[0134] b. Feature vectors from the MODIS CNN branch .

[0135] c. "Influence" feature vector from the GNN module This vector is obtained by weighted summation of the output features h_i of all AERONET nodes n_i, with the weights related to the distance between the target point p and node n_i.

[0136] S3-1: Attention Weight Calculation: Design an attention network that uses the location and time information of the target point p, as well as the preliminary features extracted from the CNN, as the "query". , , As "keys" and "values," attention networks learn to calculate three weights. , , And the sum of the three is 1.

[0137] These weights are dynamic; for example, if MODIS has high data quality at that point (good QA indicators, low uncertainty). It will be higher; if the target point is adjacent to an AERONET site, This will significantly improve performance; at night or when there is no MODIS crossing. It will approach 0, and the model will mainly rely on the features of FY-4B for time inference.

[0138] S3-2: Weighted fusion. The final fused feature vector. We obtain the result through weighted summation:

[0139]

[0140] S4: Fuse AOD output. The fused feature vector... The model takes a simple multilayer perceptron (MLP) as input as its prediction head and outputs the final fused AOD value. The model is trained using a composite loss function.

[0141]

[0142] This is the loss predicted by AOD, using the root mean square error (RMSE). The loss is for uncertain predictions, using negative log-likelihood loss, which encourages the model to produce low uncertainty when predictions are accurate and high uncertainty when predictions are inaccurate. λ is a hyperparameter used to balance the two losses.

[0143] S5: Missing Data Imputation. Due to factors such as cloud cover or sensor malfunctions, satellite data has low spatial coverage, resulting in spatial gaps in the AOD data fused from satellite data. However, MERRA-2 AOD, acquired through assimilation methods, possesses spatiotemporal continuity and can be used to fill these gaps in satellite AOD data. In this step, the LightGBM model is used to imput the missing data after fusion, generating a spatially complete AOD product.

[0144] S5-1: Sample Dataset Construction. This involves fusing the AOD data (… The model input dataset is constructed by integrating MERRA-2 AOD data, ERA5 meteorological data, DEM topographic data, and topographic feature data.

[0145] Meteorological data include boundary layer height (BLH), total column water (TCW), surface pressure (SP), 2-meter temperature (T2M), 10-meter easterly component (10UW), 10-meter northerly component (10VW), and relative humidity (RH), used to provide meteorological constraints, DEM topographic data, and topographic feature data. This is used to provide terrain constraints to ensure that the interpolation results conform to physical laws.

[0146] S5-2: Model Construction. The target output is the fused AOD, including MERRA-2 AOD, meteorological parameters (BLH, TCW, SP, T2M, 10UW, 10VW, RH), DEM, and DEM topographic features (…). Using this as input data, the LightGBM model is trained to construct the fused AOD data. The mapping relationship between ) and MERRA-2 AOD data and meteorological parameters and other variables:

[0147]

[0148] In the formula: AOD is the final output AOD result. For the fused AOD data, DEM represents topographic data, BLH represents boundary layer height, TCW represents total column water, SP represents surface pressure, T2M represents 2-meter temperature, 10UW represents the 10-meter easterly component, 10VW represents the 10-meter northerly component, and RH represents relative humidity. This is terrain feature data;

[0149] S5-3: Data Imputation. Missing values ​​are imputed in the fused data based on the trained LightGBM model to obtain spatially complete AOD data.

[0150] Example 2:

[0151] An aerosol optical thickness acquisition system based on multi-source data fusion, which can implement the aerosol optical thickness acquisition method based on multi-source data fusion described in Example 1, includes:

[0152] Data processing module: Preprocesses the optical thickness data of multi-source heterogeneous aerosols after acquisition;

[0153] Feature extraction module: Performs multimodal feature extraction on the preprocessed multi-source heterogeneous aerosol optical thickness data;

[0154] Feature fusion module: This module fuses the extracted multimodal features by introducing an attention mechanism.

[0155] Multilayer sensing module: Input the fused feature vector into the multilayer perceptron to obtain the final fused aerosol optical thickness data;

[0156] Data imputation module: Imputes missing data in the final fused aerosol optical thickness data.

[0157] Example 3:

[0158] This invention also provides an aerosol optical thickness acquisition device based on multi-source data fusion, which can realize the aerosol optical thickness acquisition method based on multi-source data fusion described in Embodiment 1, including a processor and a storage medium;

[0159] The storage medium is used to store instructions;

[0160] The processor is configured to operate according to the instructions to perform the steps of the following method:

[0161] Preprocessing is performed after acquiring the optical thickness data of multi-source heterogeneous aerosols.

[0162] Multimodal feature extraction was performed on the preprocessed optical thickness data of multi-source heterogeneous aerosols.

[0163] An attention mechanism is introduced to fuse the extracted multimodal features;

[0164] The fused feature vectors are input into a multilayer perceptron to obtain the final fused aerosol optical thickness data.

[0165] Missing data were filled in for the final fused aerosol optical thickness data.

[0166] Example 4:

[0167] This invention also provides a computer-readable storage medium that implements the aerosol optical thickness acquisition method based on multi-source data fusion described in Embodiment 1. The medium stores a computer program that, when executed by a processor, performs the steps of the following method:

[0168] Preprocessing is performed after acquiring the optical thickness data of multi-source heterogeneous aerosols.

[0169] Multimodal feature extraction was performed on the preprocessed optical thickness data of multi-source heterogeneous aerosols.

[0170] An attention mechanism is introduced to fuse the extracted multimodal features;

[0171] The fused feature vectors are input into a multilayer perceptron to obtain the final fused aerosol optical thickness data.

[0172] Missing data were filled in for the final fused aerosol optical thickness data.

[0173] As is known from common technical knowledge, this invention can be implemented through other embodiments that do not depart from its spirit or essential characteristics. Therefore, the disclosed embodiments described above are merely illustrative in all respects and are not the only ones. All modifications within the scope of this invention or its equivalents are included in this invention.

[0174] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0175] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0176] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0177] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0178] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for obtaining aerosol optical thickness based on multi-source data fusion, characterized in that, include: Preprocessing is performed after acquiring the optical thickness data of multi-source heterogeneous aerosols. Multimodal feature extraction was performed on the preprocessed optical thickness data of multi-source heterogeneous aerosols. An attention mechanism is introduced to fuse the extracted multimodal features; The fused feature vectors are input into a multilayer perceptron to obtain the final fused aerosol optical thickness data. Missing data were filled in for the final fused aerosol optical thickness data; Multimodal feature extraction was performed on the preprocessed multi-source heterogeneous aerosol optical thickness data, including: spatiotemporal feature extraction from satellite raster data and AERONET AOD feature extraction, wherein: The extraction of spatiotemporal features from satellite raster data includes: extracting spatial and temporal features from FY4B AOD and MOD04 AOD based on convolutional neural network 3D-CNN; The AERONET AOD feature extraction includes: modeling sparse data relationships of ground-based sites based on graph neural networks (GNNs), specifically including: A dynamic aerosol map is constructed, with all AERONET sites regarded as nodes of the graph. The node characteristics are its AOD measurement value, timestamp, and site location information. Defining edge connections and weights: The edge weight between two sites is defined as follows: ； Where: w_ij is the edge weight between stations i and j, dist_ij is the distance between stations i and j, and wind_vector_ij is the average wind field vector between stations i and j; Graph Convolution Operation: Dynamic aerosol map processing is performed using a GNN model. The GNN updates the feature representation of each node by aggregating information from neighboring nodes. After multi-layer GNN processing, the feature vector of each node not only contains its own measurement information but also incorporates information from other high-precision stations associated with it through the physical path. The graph neural network feature fusion formula is as follows: ； in, Is node v at the th The feature vector of the layer, Is node v in the 1st- The feature vector of the layer, Here, W is the activation function, W is the weight matrix, and u is the neighbor node of v. It is the set of neighbors of node v, d u is the degree of node u, and b is the bias. An attention mechanism is introduced to fuse the extracted multimodal features, including attention weight calculation and weighted fusion, where: The attention weight calculation includes: designing an attention network that uses the location information and time information of the target point p, as well as the preliminary features extracted from the CNN, as the "query," and using the feature vector from the FY-4B CNN branch. Feature vectors from the MODIS CNN branch "Influence" feature vectors from GNN modules As "keys" and "values", attention networks learn to calculate three weights. , , And the sum of the three is 1; The weighted fusion includes obtaining the final fused feature vector through weighted summation. The expression is: ； In the formula: This is the final fused feature vector. , , There are three weights, and the sum of the three is 1. The feature vectors are from the FY-4B CNN branch. The feature vectors are from the MODIS CNN branch. This is the "influence" feature vector from the GNN module.

2. The method for obtaining aerosol optical thickness based on multi-source data fusion according to claim 1, characterized in that, After acquiring the optical thickness data of multi-source heterogeneous aerosols, preprocessing is performed, including: Multi-source aerosol optical thickness data and auxiliary data were collected. All data underwent quality control, spectral matching, spatial matching, temporal synchronization, and standardization to construct a spatiotemporally unified multi-source dataset, including: The quality control includes cleaning up the data and removing invalid values, wherein: the official quality label is used to filter invalid values ​​for FY4B data, and the QA mark is used to filter and retain high-quality data for MODIS data. The spectral matching includes: calculating the aerosol optical thickness values ​​of different satellite AOD products in a specified band using an aerosol optical thickness value calculation formula, wherein the aerosol optical thickness value calculation formula is: ； in, Represents the optical thickness value of the aerosol. 1 represents the band, and α and β are auxiliary calculation coefficients; The spatial matching includes: unifying all grid data to the same coordinate system, and resampling the grid data using bilinear interpolation. The calculation method for bilinear interpolation is as follows: ； ； ； In the formula: Z(x,y) represents the AOD value of point (x,y), Z(x,y1) represents the AOD value of point (x,y1), and Z(x,y2) represents the AOD value of point (x,y2). , , , Let x be the AOD value of four adjacent known points, x be the center longitude coordinate of the target grid, y be the center latitude coordinate of the target grid, x1 and x2 be the longitude coordinates of the adjacent known points, and y1 and y2 be the latitude coordinates of the adjacent known points. For discrete AERONET ground observation data, inverse distance-weighted interpolation is used to assist spatial matching, as shown in the following formula: ； In the formula, For the observation value of the i-th AERONET site, Let k be the distance from the target grid to the i-th station, k be the distance attenuation coefficient, and n be the number of neighboring stations participating in the interpolation. The time synchronization includes: selecting a specified time period before and after the hour as a time window, calculating the average value of AERONET AOD and MODIS AOD data within the time window, and unifying the time resolution of multi-source data to an hourly rate based on the spatially matched data; The standardization process includes: ； ； ； In the formula: For the standardized results, Let be the original value of the i-th sample from the j-th data source. Let be the mean of the j-th data source. Let be the standard deviation of the j-th data source. The total number of valid samples after removing outliers from the j-th data source.

3. The method for obtaining aerosol optical thickness based on multi-source data fusion according to claim 2, characterized in that, The multi-source AOD data includes: FY-4B AOD, the official AOD product of Fengyun-4B satellite; MODIS Level 2 AOD product MOD04 L2AOD; MERRA-2 AOD reanalysis product MERRA-2 AOD; and AERONET Level 2.0 product AERONET AOD. The auxiliary data include: SRTM 90m resolution digital elevation model (DEM); ERA5 meteorological data, including boundary layer height (BLH), total column water (TCW), surface pressure (SP), 2-meter temperature (T2M), 10-meter easterly wind component (10UW), 10-meter northerly wind component (10VW), and relative humidity (RH); quality assurance marks or pixel-level uncertainty estimates for each satellite product.

4. The method for obtaining aerosol optical thickness based on multi-source data fusion according to claim 1, characterized in that, The fused feature vectors are input into a multilayer perceptron to obtain the final fused aerosol optical thickness data, including: the fused feature vectors... A simple multilayer perceptron is input as the prediction head, and the final fused AOD value is output. The model is trained using a composite loss function. ‘； In the formula: This is the loss predicted by AOD. It represents the loss due to uncertain prediction, where λ is a hyperparameter.

5. The method for obtaining aerosol optical thickness based on multi-source data fusion according to claim 4, characterized in that, Missing data imputation was performed on the final fused aerosol optical thickness data, including: sample dataset construction, model building, and data imputation, where: The construction of the sample dataset includes: merging the AOD data The MERRA-2 AOD data, ERA5 meteorological data, DEM topographic data, and topographic feature data are integrated to construct the model input dataset; The model construction includes: taking the fused AOD as the target output, MERRA-2 AOD, meteorological parameters, DEM, and DEM topographic features. The LightGBM model is trained using this as input data to construct the fused AOD data. The mapping relationship between MERRA-2AOD data and meteorological parameters and other variables is as follows: ； In the formula: AOD is the final output AOD data. For the fused AOD data, DEM represents topographic data, BLH represents boundary layer height, TCW represents total column water, SP represents surface pressure, T2M represents 2-meter temperature, 10UW represents the 10-meter easterly component, 10VW represents the 10-meter northerly component, and RH represents relative humidity. This is terrain feature data; The data imputation includes: imputing missing values ​​in the fused data based on the trained LightGBM model to obtain spatially full-coverage AOD data.

6. An aerosol optical thickness acquisition system based on multi-source data fusion, characterized in that, The method for obtaining aerosol optical thickness based on multi-source data fusion as described in any one of claims 1 to 5 includes: Data processing module: Preprocesses the optical thickness data of multi-source heterogeneous aerosols after acquisition; Feature extraction module: Performs multimodal feature extraction on the preprocessed multi-source heterogeneous aerosol optical thickness data; Feature fusion module: This module fuses the extracted multimodal features by introducing an attention mechanism. Multilayer sensing module: Input the fused feature vector into the multilayer perceptron to obtain the final fused aerosol optical thickness data; Data imputation module: Imputes missing data in the final fused aerosol optical thickness data.

7. An aerosol optical thickness acquisition device based on multi-source data fusion, characterized in that, Including processor and storage media; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 5.

Images