Satellite remote sensing aerosol optical depth data reconstruction method considering spatial complementation and downscaling

Abstract

The present application relates to the technical field of atmospheric environment monitoring, and in particular to a satellite remote sensing AOD data reconstruction method considering spatial data filling and downscaling, which comprises the following steps: collecting satellite remote sensing AOD data, auxiliary factors and ground verification data; performing data preprocessing to obtain an original spatial resolution data set and a target spatial resolution data set; combining the original spatial resolution data set and a random forest algorithm to construct an AOD spatial data filling and downscaling initial model, and obtaining an optimal auxiliary factor; optimizing the initial model based on the optimal auxiliary factor to obtain an AOD spatial data filling and downscaling optimal model; and inputting the optimal auxiliary factor and a time variable in the target spatial resolution data set into the optimal model to reconstruct AOD data and obtain satellite remote sensing AOD data with spatial data filling and downscaling. The method can realize spatial downscaling while filling data, and quickly and efficiently obtain target spatial resolution AOD data without data loss.

Description

Technical Field

[0001] This invention relates to the field of atmospheric environment monitoring technology, specifically to a satellite remote sensing AOD data reconstruction method that takes into account spatial gap filling and downscaling. Background Technology

[0002] Aerosols are of great significance to research on climate change and human health. In my country, aerosols are one of the major pollutants affecting air quality. Aerosol optical depth (AOD) is an important parameter describing atmospheric turbidity and total aerosol content. Currently, the main methods for acquiring AOD data include ground-based observation and satellite remote sensing. Ground-based observation has the advantages of high precision and strong temporal continuity, but its spatial coverage is limited due to the limited number of stations, making it unsuitable for large-scale monitoring. Satellite remote sensing, with its high spatial coverage and temporal resolution, can achieve global monitoring and has become an important means of monitoring AOD. However, satellite remote sensing is affected by weather factors, and the acquired AOD data still has significant gaps in both time and space.

[0003] Furthermore, many current studies, such as urban pollution source tracing and refined health risk assessment, urgently require high-resolution data to capture changes within cities and at local scales, especially AOD data with a resolution of hundreds of meters. However, current satellite remote sensing AOD products are generally at the kilometer scale, with low resolution, which is insufficient to meet research needs. Therefore, conducting high spatial resolution AOD research with complete data is particularly important.

[0004] Current research on obtaining spatially complete satellite remote sensing AOD (Air-Obtained Distance) coverage has made some progress. The main methods include spatial interpolation, multi-source AOD data fusion, and statistical modeling estimation. Spatial interpolation is simple and convenient, and can fill in missing AODs, but its accuracy is limited by the spatial coverage and accuracy of the original AOD data. While multi-source AOD data fusion can obtain spatially complete AOD data, significant differences often exist between data sources, potentially introducing errors and reducing accuracy. Statistical modeling estimation can achieve full coverage, but it is highly dependent on the selection of auxiliary elements and the amount of data.

[0005] There have been many attempts at AOD downscaling, mainly including interpolation downscaling and statistical downscaling. Interpolation downscaling is simple and fast, but its downscaling effect is poor, only smoothing the AOD spatially. Statistical downscaling methods can achieve better downscaling, but require more auxiliary elements and have a large computational burden. Currently, most techniques only consider spatial imputation or only spatial downscaling, without considering methods that can achieve both simultaneously.

[0006] The search revealed:

[0007] The publication number is CN114117899A, the publication date is March 1, 2022, and the invention title is "A Spatiotemporal Imputation Method, System, and Computer Equipment for Satellite Data." This application discloses a spatiotemporal imputation method, system, and computer equipment for satellite data. The method includes: acquiring an initial dataset; and imputing missing data in the initial dataset by combining a temporal interpolation model and a spatial interpolation model. This application addresses the technical problem of the lack of high temporal resolution satellite data imputation algorithms.

[0008] The invention, with publication number CN111859304B and publication date November 21, 2023, entitled "A Method and System for Predicting Satellite Aerosol Missing Data Based on Spatiotemporal Autocorrelation," discloses a method and system for predicting satellite aerosol missing data based on spatiotemporal autocorrelation. The method includes: creating a grid covering the target area and obtaining the projected coordinates of the center point of each grid; resampling the AOD (Aerosol Occurrence Discharge) of the target area and the variables mediating the spatiotemporal autocorrelation of satellite AOD into the grid, matching the grid with the data, and forming a separate dataset for each day's data; establishing a prediction dataset based on the spatiotemporal autocorrelation of satellite AOD; repeatedly sampling the prediction dataset multiple times, constructing a missing data prediction model based on a generalized additive model for each sampling; predicting AOD missing data based on the established multiple missing data prediction models, and using the average of the multiple predictions for each grid as the final AOD prediction value. This invention predicts satellite AOD missing data while considering the temporal and spatial autocorrelation of satellite AOD. The prediction accuracy index obtained by comparing the predicted values ​​with remote sensing monitoring values ​​reflects the high accuracy of the method in predicting missing data.

[0009] In summary, there is an urgent need to design a satellite remote sensing AOD data reconstruction method that takes into account spatial missing data and downscaling to solve the problems of missing satellite remote sensing AOD data and insufficient spatial resolution in existing technologies. Summary of the Invention

[0010] The purpose of this invention is to provide a satellite remote sensing AOD data reconstruction method that takes into account spatial missing data filling and downscaling. Through machine learning and spatial statistical techniques, it can learn the statistical regression relationship between satellite remote sensing AOD data and auxiliary factors from a large amount of data, thereby achieving accurate reconstruction of missing AOD information. It can simultaneously fill in data gaps and perform spatial downscaling, avoiding a large amount of repetitive statistical calculations. This method can quickly and efficiently achieve AOD data filling and spatial downscaling, completing the supplementation of missing satellite remote sensing AOD data and improving spatial resolution. The specific technical solution is as follows:

[0011] A method for reconstructing satellite remote sensing AOD data that takes into account spatial missing data and downscaling includes the following steps:

[0012] Step 1: Collect satellite remote sensing AOD data, ancillary factor data, and ground-based validation data for a specific area over a period of more than one month. The ancillary factor data includes numerical model AOD data, topographic data, land use type data, vegetation enhancement index data, relative humidity data, atmospheric boundary layer height data, total precipitation data, surface air pressure data, air temperature data, dew point temperature data, wind speed data, and wind direction data. The ground-based validation data is AERONET ground-based monitoring data.

[0013] Step 2: Preprocess the satellite remote sensing AOD data, auxiliary factor data and ground-based validation data collected in Step 1 to obtain the original spatial resolution dataset with the same spatial scale as the satellite remote sensing AOD data and the target spatial resolution dataset to be achieved by spatial downscaling.

[0014] A spatial imputation and downscaling initial model for AOD is constructed based on the random forest algorithm. The original spatial resolution dataset is input into the AOD spatial imputation and downscaling initial model to obtain the optimal auxiliary factor.

[0015] The optimal auxiliary factor is input into the initial model of AOD space missing and downscaling for iterative training to obtain the optimal model of AOD space missing and downscaling.

[0016] Step 3: Input the optimal auxiliary factors and time variables from the target spatial resolution dataset into the optimal AOD spatial missing and downscaling model to reconstruct AOD data at the target spatial resolution scale, and obtain spatially missing and downscaled satellite remote sensing AOD data.

[0017] Preferably, the preprocessing in step two includes:

[0018] The projection transformation process specifically involves: performing projection transformation on satellite remote sensing AOD data and auxiliary factor data using ArcGIS software to unify them into the WGS1984 coordinate system.

[0019] Spatial matching processing specifically involves spatial matching of satellite remote sensing AOD data and auxiliary factor data. Specifically, spatial matching from high spatial resolution to low spatial resolution uses a resampling method, while spatial matching from low spatial resolution to high spatial resolution uses an interpolation method.

[0020] The null data removal process specifically involves removing null data from satellite remote sensing AOD data and auxiliary factor data.

[0021] The satellite remote sensing AOD data and auxiliary factor data are used to create an original spatial resolution dataset with the same spatial scale as the satellite remote sensing AOD data and a target spatial resolution dataset to be achieved by spatial downscaling, based on the results of spatial matching processing and missing data removal processing.

[0022] Preferably, the resampling method uses a statistical method of averaging, as shown in the following formula:

[0023]

[0024] Where: R H It is high spatial resolution data; R L It is low spatial resolution data; (lon) p ,lat p (lon, lat) are the pixel coordinates in the high spatial resolution data; (lon, lat) are the corresponding pixel coordinates in the low spatial resolution data; n is the number of pixels in the high spatial resolution data covered by each pixel in the low spatial resolution data.

[0025] The interpolation method used is bilinear interpolation, and the formula is as follows:

[0026] f(lon,lat)=(1-w)(1-h)f(lon1,lat1)+w(1-h)f(lon2,lat1)+(1-w)hf(lon1,lat2)+whf(lon2,lat2);

[0027] w = (lon - lon1) / (lon2 - lon1);

[0028] h = (lat - lat1) / (lat2 - lat1);

[0029] Where: (lon1,lat1), (lon2,lat1), (lon1,lat2), and (lon2,lat2) are four known data points, with corresponding pixel values ​​f(lon1,lat1), f(lon2,lat1), f(lon1,lat2), and f(lon2,lat2), respectively; w and h represent the horizontal and vertical distance ratios of the target position relative to the known data points, and f(lon,lat) is the estimated pixel value of the point;

[0030] The specific process for removing missing data is as follows: the location and time of missing data in satellite remote sensing AOD data are marked, and then auxiliary factor data at the same spatial location and time are removed based on the marking.

[0031] Preferably, it also includes band conversion processing, specifically: interpolating the ground-based verification data of adjacent bands of the satellite remote sensing AOD data using the Angstrom index to obtain the AOD value of the satellite remote sensing AOD data band. The calculation expression is as follows:

[0032]

[0033]

[0034] Where: α represents the Angstrom index, λ1 and λ2 represent the wavelengths of two adjacent bands of satellite remote sensing AOD data, and λ3 represents the wavelength of satellite remote sensing AOD data; and λ1 and λ2 represent the AOD values ​​corresponding to them; τ represents the ground-based verification AOD of the same band as the satellite remote sensing AOD data.

[0035] Preferably, the initial model for AOD spatial imputation and downscaling in step two is as follows:

[0036] AOD SAT ~RF(AOD) R , Mete, DEM, LUC, EVI, Time);

[0037]

[0038] Among them: AOD SAT Represents satellite remote sensing AOD data; AOD R Represents AOD reanalysis data; Mete represents a matrix composed of meteorological variables; DEM represents topographic data; LUC represents land use type data; EVI represents vegetation enhancement index data; Time represents the corresponding observation time; RF represents the random forest model; x represents the AOD numerical model, Mete, DEM, EVI, and Time; T i D represents the i-th decision tree; i represents the training dataset for the i-th decision tree; N represents the number of decision trees; m represents the minimum number of samples per leaf node; Sur represents the surrogate splitting function; PS represents the independent variable selection criteria; OOBP represents the out-of-bag prediction function; OOBPI represents the out-of-bag independent variable importance estimation function; Method represents the method type; on represents the function being enabled; curvature represents the use of curvature as the selection criterion; regression represents the method type as a regression model.

[0039] Preferred methods for obtaining the optimal auxiliary factor include:

[0040] The importance estimates of the out-of-bag independent variables are calculated using the following formula:

[0041]

[0042] Where: Baseline OOB Error is the baseline out-of-bag prediction error obtained by splitting other independent variables; VI(k) is the importance estimate of independent variable k; OOB Error(k) is the out-of-bag prediction error when splitting independent variable k, and the calculation expression for the out-of-bag prediction error OOB Error is: n represents the total number of samples, y j Let f represent the true label of the j-th sample. -j (x j ) represents the prediction result of the random forest model trained using other samples on the j-th sample, and L is the loss function;

[0043] The optimal auxiliary factor is obtained based on the importance estimate.

[0044] Preferably, the process of obtaining the optimal AOD space-patch and downscaling model through iterative training includes:

[0045] Step ①: Input the optimal auxiliary factor into the AOD space imputation and downscaling initial model to obtain the following expression:

[0046] AOD SAT ~RF(OIF1,OIF2,……,OIF N ,Time)[N=50,m=30,Sur=on,PS=curvature,OOBP=on,OOBPI=on,Method=regression];

[0047] Where: OIF1-OIF N This represents the first to Nth terms in the optimal auxiliary factor;

[0048] Set the coefficient of determination R. 2 The root mean square error (RMSE) and the mean bias (Bias) are used as model evaluation metrics, as follows:

[0049]

[0050] Where: Y and X are the spatially imputed downscaled AOD and the satellite remote sensing AOD, respectively; S is the number of matched samples; ∑* is the averaging operator; ∑* is the summation operator; The square root operator;

[0051] Step 2: The dataset consisting of satellite remote sensing AOD, optimal auxiliary factors, and time variables is fixed and randomly shuffled. 60% of the dataset is used as the training set and 40% is used as the test set. The expected results of model training using the training set and modeling using the test set are verified to obtain the model accuracy.

[0052] Step ③: Judge based on the model accuracy. Specifically: if the root mean square error (RMSE) of the test set is less than 0.1 and the average bias (Bias) is less than 0.001, the accuracy index meets the requirements, stop model iteration, and obtain the optimal model for AOD space filling and downscaling; otherwise, optimize and adjust the model parameters and return to step ②.

[0053] The parameters for model optimization and tuning include the number of decision trees and the depth of leaf nodes. The initial number of decision trees is 50, and the minimum number of samples per leaf node is 30. The formula for parameter tuning is as follows:

[0054] N r+1 =N r +10;

[0055] m r+1 =m r +10;

[0056] Where: N i m i These represent the number of policy trees and the minimum number of leaf nodes in the r-th iteration, respectively. The number of policy trees and the minimum number of leaf nodes are increased by 10 in each iteration.

[0057] The parameters of the optimal model for AOD spatial imputation and downscaling are as follows:

[0058]

[0059] Where N_opt and m_opt represent the optimal parameters for N and m.

[0060] Preferably, the AOD data reconstruction in step three includes:

[0061] Target spatial resolution:

[0062]

[0063] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The embodiments described below will further illustrate the invention in detail. Detailed Implementation

[0064] The present invention will be described in detail below with reference to the embodiments, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0065] Example:

[0066] A method for reconstructing satellite remote sensing AOD data that takes into account spatial missing data and downscaling includes the following steps:

[0067] Step 1: Collect satellite remote sensing AOD data, auxiliary factor data, and ground-based validation data for a certain area for more than one month. Among them, the auxiliary factor data includes numerical model AOD data, topographic data, land use type data, vegetation enhancement index data, relative humidity data, atmospheric boundary layer height data, total precipitation data, surface air pressure data, air temperature data, dew point temperature data, wind speed data, and wind direction data; the ground-based validation data is AERONET ground-based monitoring data.

[0068] This embodiment collects MAIAC AOD data, auxiliary factor data, and ground-based validation data for the Beijing-Tianjin-Hebei region in 2020. Auxiliary factor data includes: MERRA-2 AOD numerical model data, STRM DEM data, CLCD land use type data, MODIS vegetation enhancement index data, and ERA5 meteorological data (including relative humidity, air temperature, 2m dew point temperature, air pressure, boundary layer height, total precipitation, 10m U-shaped wind volume, and 10m V-shaped wind volume).

[0069] Step 2: Preprocess the satellite remote sensing AOD data, auxiliary factor data, and ground-based validation data collected in Step 1 to create an original spatial resolution dataset with the same spatial scale as the satellite remote sensing AOD data and a target spatial resolution dataset to be achieved by spatial downscaling, i.e., a 1km resolution dataset (original spatial resolution dataset) and a 500m resolution dataset (target spatial resolution dataset).

[0070] Since the collected satellite remote sensing AOD data and auxiliary factor data are in TIFF format, the data is incomplete and the spatial resolution is inconsistent, which is not convenient for subsequent analysis and calculation. Therefore, it is necessary to preprocess them. In this embodiment, the preprocessing includes the following steps:

[0071] The projection transformation process specifically involves: performing projection transformation on satellite remote sensing AOD data and auxiliary factor data using the GDAL library or ArcGIS software, and unifying them to the WGS1984 coordinate system.

[0072] Spatial matching processing specifically involves spatially matching satellite remote sensing AOD data and auxiliary factor data. Specifically, spatial matching from high spatial resolution to low spatial resolution uses a resampling method, while spatial matching from low spatial resolution to high spatial resolution uses an interpolation method. Further optimization includes:

[0073] The resampling method uses a statistical method of averaging, as shown in the following formula:

[0074]

[0075] Where: R H It is high spatial resolution data; R L It is low spatial resolution data; (lon) p ,lat p (lon, lat) are the pixel coordinates in the high spatial resolution data; (lon, lat) are the corresponding pixel coordinates in the low spatial resolution data; n is the number of pixels in the high spatial resolution data covered by each pixel in the low spatial resolution data.

[0076] The interpolation method used is bilinear interpolation, and the formula is as follows:

[0077] f(lon,lat)=(1-w)(1-h)f(lon1,lat1)+w(1-h)f(lon2,lat1)+(1-w)hf(lon1,lat2)+whf(lon2,lat2);

[0078] w=(lon-lon1) / (lon2-lon1); h=(lat-lat1) / (lat2-lat1);

[0079] Where: (lon1,lat1), (lon2,lat1), (lon1,lat2), and (lon2,lat2) are four known data points, with corresponding pixel values ​​f(lon1,lat1), f(lon2,lat1), f(lon1,lat2), and f(lon2,lat2), respectively; w and h represent the horizontal and vertical distance ratios of the target position relative to the known data points, and f(lon,lat) is the estimated pixel value of the point;

[0080] The null data removal process involves removing null data from satellite remote sensing AOD data and auxiliary factor data. This mainly involves marking the locations and times where missing data exists in the satellite remote sensing AOD data, and then removing auxiliary factor data at the same spatial location and time based on the markings.

[0081] Satellite remote sensing AOD data and auxiliary factor data were used to create two datasets with 1km and 500m resolutions based on spatial matching and the removal of missing data.

[0082] Because different satellite remote sensing AOD data are acquired in different bands, band conversion of the ground-based verification data is necessary to achieve band-consistent AOD data ground-based verification. In this embodiment, the ground-based AERONET data includes AOD monitoring values ​​in multiple bands (340, 380, 440, 500, 675, 870, and 1020 nm), but lacks 550 nm band monitoring values. Therefore, the ground-based verification data needs to be converted to the 550 nm band for subsequent AOD data ground-based verification. This embodiment preferably uses a 550 nm band conversion process, specifically: interpolating adjacent bands using the Angstrom index to obtain the AOD value of the target band. The specific calculation expression is as follows:

[0083]

[0084]

[0085] Where: α represents the Angstrom index, λ1 and λ2 represent the wavelengths of two adjacent bands of satellite remote sensing AOD data, and λ3 represents the wavelength of satellite remote sensing AOD data; and λ1 and λ2 represent the AOD values ​​corresponding to λ1 and λ2; τ represents the ground-based verification AOD of the same band as the satellite remote sensing AOD data.

[0086] This embodiment employs steps such as projection transformation, spatial matching, and missing data removal. These steps aim to provide a consistent, high-quality data foundation for subsequent analysis and calculations, ensuring consistency and usability among the data. Band conversion is performed on the ground-based validation data to achieve data consistency between the MAIAC AOD data and the ground-based validation data, enabling more accurate validation of the MAIAC AOD data and facilitating subsequent accuracy verification.

[0087] An initial model for spatial imputation and downscaling of AOD is constructed based on the random forest algorithm. The original spatial resolution dataset is input into the initial model for spatial imputation and downscaling of AOD to obtain the optimal auxiliary factor. The preferred scheme in this embodiment is as follows:

[0088] A spatial missing and downscaling initial model for AOD is constructed based on the random forest algorithm. A dataset with a 1km spatial resolution is input into the AOD spatial missing and downscaling initial model, with the MAIAC AOD data as the dependent variable and the auxiliary factor data as the independent variables. The expression is as follows:

[0089] AOD MAIAC ~RF(AOD) MERRA-2 , Mete, DEM, LUC, EVI, Time);

[0090]

[0091] Among them: AOD MAIAC This refers to satellite remote sensing MAIACAOD data; AOD MERRA-2 This represents MERRA-2 AOD reanalysis data; Mete represents a matrix of meteorological variables, including relative humidity, air temperature, 2m dew point temperature, air pressure, boundary layer height, total precipitation, 10m U-shaped wind volume, and 10m V-shaped wind volume; DEM represents STRM DEM data; LUC represents CLCD land use type data; EVI represents MODIS vegetation enhancement index data; Time represents the corresponding observation time; RF represents the random forest model; x represents the AOD numerical model, Mete, DEM, EVI, and Time; T i D represents the i-th decision tree; i represents the training dataset for the i-th decision tree; N represents the number of decision trees; m represents the minimum number of samples per leaf node; Sur represents the surrogate splitting function; PS represents the independent variable selection criteria; OOBP represents the out-of-bag prediction function; OOBPI represents the out-of-bag independent variable importance estimation function; Method represents the method type; on represents the function being enabled; curvature represents the use of curvature as the selection criterion; regression represents the method type as a regression model.

[0092] For the initial AOD spatial imputation and downscaling model, the relative importance scores of each auxiliary factor are obtained by enabling the out-of-bag independent variable importance estimation function. These scores can measure the influence of each auxiliary factor on the AOD spatial imputation and downscaling model; specifically, the higher the importance score, the greater the contribution of the auxiliary factor to AOD reconstruction. For a given set of auxiliary factors, they are sorted according to their importance scores, and the top N most important auxiliary factors are selected as the independent variables for subsequent training of the optimal AOD spatial imputation and downscaling model. This reduces the dimensionality of the model input and improves the model's computational efficiency and prediction accuracy.

[0093] Methods for obtaining the optimal auxiliary factor include:

[0094] The importance estimates of the out-of-bag independent variables are calculated using the following formula:

[0095]

[0096] Where: Baseline OOB Error is the baseline out-of-bag prediction error obtained by splitting other independent variables; VI(k) is the importance estimate of independent variable k; OOB Error(k) is the out-of-bag prediction error when splitting independent variable k, and the calculation expression for the out-of-bag prediction error OOB Error is: n represents the total number of samples, y j Let f represent the true label of the j-th sample. -j (x j ) represents the prediction result of the random forest model trained using other samples on the j-th sample, and L is the loss function;

[0097] The optimal auxiliary factor is obtained based on the importance estimate.

[0098] In this embodiment, the optimal auxiliary factors include MERRA-2 AOD, boundary layer height, 10mV-type air volume, 10mU-type air volume, STRM DEM, air temperature, dew point temperature, air pressure, and CLCD land use type.

[0099] The optimal auxiliary factor is input into the initial model for AOD space missing-downscaling and iterative training to obtain the optimal model for AOD space missing-downscaling. In this embodiment, the process of iterative training to obtain the optimal model for AOD space missing-downscaling includes the following:

[0100] Step ①: Input the optimal auxiliary factor into the AOD space imputation and downscaling initial model to obtain the following expression:

[0101] AOD MAIAC ~RF(AOD) MERRA-2 Mete BLH,V10,U10,TP,TDM,SP ,DEM,LUC,Time)[N=50,m=30,Sur=on,PS=curvature,OOBP=on,OOBPI=on,Method=regression];

[0102] Where: BLH in Mete represents the boundary layer height; V10 represents a 10m V-shaped airflow; U10 represents a 10m U-shaped airflow; TP represents the air temperature; TDM represents the dew point temperature; SP represents the air pressure;

[0103] Set the coefficient of determination R. 2 The root mean square error (RMSE) and the mean bias (Bias) are used as model evaluation metrics, as follows:

[0104]

[0105] Where: Y and X are the AOD of spatially imputed downscaled and the original MAIAC AOD at 1km resolution, respectively; S is the number of matched samples; ∑* is the averaging operator; ∑* is the summation operator; The square root operator;

[0106] Step 2: Set a fixed random number generation seed, and shuffle the dataset consisting of MAIAC AOD, the optimal auxiliary factor, and the time variable. Use 60% of the dataset as the training set and 40% as the test set. Use the training set to train the model and the test set to test the expected performance of the model to obtain the model accuracy.

[0107] Step ③: Judge based on the model accuracy. Specifically: if the root mean square error (RMSE) of the test set is less than 0.1 and the average bias (Bias) is less than 0.001, then the accuracy index meets the requirements. In this case, it can be considered that the scale residual of fitting AOD at the 1km scale can be ignored, and the model iteration is stopped to obtain the optimal model for AOD spatial missing and downscaling. Otherwise, optimize and adjust the model parameters and return to step ②.

[0108] In this embodiment, the parameters for model optimization and adjustment include the number of decision trees and the depth of leaf nodes. The initial number of decision trees is 50, the minimum number of leaf nodes is 30, the surrogate splitting function is enabled, the out-of-bag prediction function is enabled, the out-of-bag independent variable importance estimation function is enabled, the curvature is specified as the criterion for independent variable selection, and the model is specified as a regression model.

[0109] The formula for parameter adjustment is as follows:

[0110] N r+1 =N r +10; m r+1 =m r +10;

[0111] Where: N i m i These represent the number of policy trees and the minimum number of leaf nodes in the r-th iteration, respectively. The number of policy trees and the minimum number of leaf nodes are increased by 10 in each iteration.

[0112] The optimal model parameters for AOD spatial imputation and downscaling are as follows:

[0113]

[0114] In this embodiment, the annual model training set and test set R 2 All were above 0.94, RMSE were between 0.03 and 0.09, and Bias were less than 0.0002.

[0115] Step 3: Input the optimal auxiliary factors and time variables from the 500m spatial resolution dataset into the optimal AOD spatial missing and downscaling model to reconstruct the AOD data at the 500m spatial resolution scale, obtaining spatially missing and downscaled satellite remote sensing AOD data. This embodiment preferably uses:

[0116] Input the 500m resolution dataset from step two into the optimal AOD spatial missing and downscaling model obtained in step two to predict AOD data at 500m resolution, and obtain spatial missing and downscaling AOD data at 500m spatial resolution, that is, obtain spatial missing and downscaling satellite remote sensing AOD data.

[0117] The refactoring process includes:

[0118] 500m resolution (i.e., target spatial resolution):

[0119]

[0120] In addition, this embodiment uses the preprocessed AERONET ground monitoring data obtained in step two to verify the accuracy of the spatially missing and downscaled AOD and MAIACAOD at 500m, and compares the accuracy of the two to verify the data missing and downscaling effect of the spatially missing and downscaled AOD, as follows:

[0121] To ensure comparability of verification accuracy between AOD data at different spatial resolutions during ground-based verification comparison, this embodiment employs a consistent spatiotemporal window for data verification. Specifically, a ground-based AERONET site is selected as the center, and the original satellite remote sensing AOD spatial resolution is used as the range. The average value of the original satellite AOD and the spatially downscaled satellite remote sensing AOD within this range are calculated and matched with the ground-based AERONET site. For ground-based AERONETAOD, the average value 35 minutes before and after satellite transit is selected and matched with both the original satellite remote sensing AOD and the spatially downscaled satellite remote sensing AOD. The verification index selected is the coefficient of determination (R²). 2 Ground-based verification confirmed that the 2020 500m resolution spatially amended and downscaled AOD data validated R... 2 The value is 0.79, and the MAIAC AOD data validates the R value. 2 The R-value was 0.72, and the AOD data was validated with spatial imputation and downscaling at 500m resolution. 2 Greater than MAIAC AOD R 2 Spatial gap filling and downscaling AOD data for this region has been successfully obtained.

[0122] The technical solution applied in this embodiment is as follows: First, satellite remote sensing AOD data, auxiliary factors, and ground-based validation data are collected within a certain area. Then, the satellite remote sensing AOD data, auxiliary factors, and ground-based validation data are preprocessed to ensure data consistency, resulting in two datasets with different spatial resolutions: 1km and 500m. Next, an initial AOD spatial missing-filling and downscaling model is constructed based on the random forest algorithm. The 1km spatial resolution dataset is input into the initial AOD spatial missing-filling and downscaling model to obtain the optimal auxiliary factor. The optimal auxiliary factor is then input into the initial AOD spatial missing-filling and downscaling model for iterative training to obtain the optimal AOD spatial missing-filling and downscaling model. Finally, the 500m spatial resolution dataset is input into the obtained optimal AOD spatial missing-filling and downscaling model to reconstruct the AOD data at the 500m spatial resolution, obtaining spatially missing and downscaled AOD data at the 500m spatial resolution. Finally, an accuracy comparison is conducted based on the ground-based validation data.

[0123] Compared with existing technologies, the technical advantages of this embodiment are as follows: In the case of insufficient spatial resolution and a large amount of missing data in satellite remote sensing AOD data, this invention develops a method for constructing high spatial resolution AOD data with complete spatial coverage by coupling machine learning and geographic information theory and technology. Through machine learning and spatial statistical techniques, it is possible to learn the statistical regression relationship between satellite remote sensing AOD data and auxiliary factors from a large amount of data, thereby achieving accurate reconstruction of missing AOD information. It can achieve spatial downscaling while filling in data gaps, avoiding a large amount of repetitive statistical calculations, and can quickly and efficiently achieve AOD data filling and spatial downscaling.

[0124] This embodiment also discloses a computer storage medium storing computer program instructions, which, when executed by a processor, implement the above-mentioned satellite remote sensing AOD data reconstruction method that takes into account spatial missing data and downscaling.

[0125] This embodiment discloses an apparatus, including: at least one processor, at least one memory, and computer program instructions stored in the memory, wherein the computer program instructions are executed by the processor as described above in the satellite remote sensing AOD data reconstruction method that takes into account spatial missing data and downscaling.

[0126] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for reconstructing satellite remote sensing AOD data that takes into account spatial missing data and downscaling, characterized in that, Includes the following steps: Step 1: Collect satellite remote sensing AOD data, ancillary factor data, and ground-based validation data. The ancillary factor data includes numerical model AOD data, topographic data, land use type data, vegetation enhancement index data, relative humidity data, atmospheric boundary layer height data, total precipitation data, surface air pressure data, air temperature data, dew point temperature data, wind speed data, and wind direction data. The ground-based validation data is AERONET ground-based monitoring data. Step 2: Preprocess the satellite remote sensing AOD data, auxiliary factor data and ground-based verification data collected in Step 1 to obtain the original spatial resolution dataset and the target spatial resolution dataset; A spatial imputation and downscaling initial model for AOD is constructed based on the random forest algorithm. The original spatial resolution dataset is input into the AOD spatial imputation and downscaling initial model to obtain the optimal auxiliary factor. The optimal auxiliary factor is input into the initial model of AOD space missing and downscaling for iterative training to obtain the optimal model of AOD space missing and downscaling. Step 3: Input the optimal auxiliary factors and time variables from the target spatial resolution dataset into the optimal AOD spatial missing and downscaling model to reconstruct AOD data at the target spatial resolution scale, and obtain spatially missing and downscaled satellite remote sensing AOD data.

2. The satellite remote sensing AOD data reconstruction method considering spatial gap filling and downscaling according to claim 1, characterized in that, The preprocessing in step two includes: The projection transformation process specifically involves: performing projection transformation on satellite remote sensing AOD data and auxiliary factor data using ArcGIS software to unify them into the WGS1984 coordinate system. Spatial matching processing specifically involves spatial matching of satellite remote sensing AOD data and auxiliary factor data. Specifically, spatial matching from high spatial resolution to low spatial resolution uses a resampling method, while spatial matching from low spatial resolution to high spatial resolution uses an interpolation method. The null data removal process specifically involves removing null data from satellite remote sensing AOD data and auxiliary factor data. Satellite remote sensing AOD data and auxiliary factor data are processed through spatial matching and missing data removal to create an original spatial resolution dataset with the same spatial scale as the satellite remote sensing AOD data and a target spatial resolution dataset to be achieved through spatial downscaling.

3. The satellite remote sensing AOD data reconstruction method considering spatial gap filling and downscaling according to claim 2, characterized in that, The resampling method uses a statistical method of averaging, as shown in the following formula: Where: R H It is high spatial resolution data; R L It is low spatial resolution data; (lon) p ,lat p (lon, lat) are the pixel coordinates in the high spatial resolution data; (lon, lat) are the corresponding pixel coordinates in the low spatial resolution data; n is the number of pixels in the high spatial resolution data covered by each pixel in the low spatial resolution data. The interpolation method used is bilinear interpolation, and the formula is as follows: f(lon,lat)=(1-w)(1-h)f(lon1,lat1)+w(1-h)f(lon2,lat1)+(1-w)hf(lon1,lat2)+whf(lon2,lat2); w = (lon - lon1) / (lon2 - lon1); h = (lat - lat1) / (lat2 - lat1); Where: (lon1,lat1), (lon2,lat1), (lon1,lat2), and (lon2,lat2) are four known data points, with corresponding pixel values ​​f(lon1,lat1), f(lon2,lat1), f(lon1,lat2), and f(lon2,lat2), respectively; w and h represent the horizontal and vertical distance ratios of the target position relative to the known data points, and f(lon,lat) is the estimated pixel value of the point; The specific process for removing missing data is as follows: the location and time of missing data in satellite remote sensing AOD data are marked, and then auxiliary factor data at the same spatial location and time are removed based on the marking.

4. The satellite remote sensing AOD data reconstruction method considering spatial gap filling and downscaling according to claim 2, characterized in that, It also includes band conversion processing, specifically: interpolating the ground-based verification data of adjacent bands of the satellite remote sensing AOD data using the Angstrom index to obtain the AOD value of the satellite remote sensing AOD data band. The calculation expression is as follows: Where: α represents the Angstrom index, λ1 and λ2 represent the wavelengths of two adjacent bands of satellite remote sensing AOD data, and λ3 represents the wavelength of satellite remote sensing AOD data; and λ1 and λ2 represent the AOD values ​​corresponding to λ1 and λ2; τ represents the ground-based verification AOD of the same band as the satellite remote sensing AOD data.

5. The satellite remote sensing AOD data reconstruction method considering spatial gap filling and downscaling according to any one of claims 1-4, characterized in that, The initial model for AOD spatial imputation and downscaling in step two is as follows: AOD SAT ~RF(AOD R ,Mete,DEM,LUC,EVI,Time); Among them: AOD SAT Represents satellite remote sensing AOD data; AOD R Represents AOD reanalysis data; Mete represents a matrix composed of meteorological variables; DEM represents topographic data; LUC represents land use type data; EVI represents vegetation enhancement index data; Time represents the corresponding observation time; RF represents the random forest model; x represents the AOD numerical model, Mete, DEM, EVI, and Time; T i D represents the i-th decision tree; i represents the training dataset for the i-th decision tree; N represents the number of decision trees; m represents the minimum number of samples per leaf node; Sur represents the surrogate splitting function; PS represents the independent variable selection criteria; OOBP represents the out-of-bag prediction function; OOBPI represents the out-of-bag independent variable importance estimation function; Method represents the method type; on represents the function being enabled; curvature represents the use of curvature as the selection criterion; regression represents the method type as a regression model.

6. The satellite remote sensing AOD data reconstruction method considering spatial gap filling and downscaling according to claim 5, characterized in that, Methods for obtaining the optimal auxiliary factor include: The importance estimates of the out-of-bag independent variables are calculated using the following formula: Where: Baseline OOB Error is the baseline out-of-bag prediction error obtained by splitting other independent variables; VI(k) is the importance estimate of independent variable k; OOB Error(k) is the out-of-bag prediction error when splitting independent variable k, and the calculation expression for the out-of-bag prediction error OOB Error is: n represents the total number of samples, y j Let f represent the true label of the j-th sample. -j (x j ) represents the prediction result of the random forest model trained using other samples on the j-th sample, and L is the loss function; The optimal auxiliary factor is obtained based on the importance estimate.

7. The satellite remote sensing AOD data reconstruction method considering spatial gap filling and downscaling according to claim 6, characterized in that, The process of obtaining the optimal AOD space-patch and downscaling model through iterative training includes: Step ①: Input the optimal auxiliary factor into the AOD space imputation and downscaling initial model to obtain the following expression: AOD SAT ~RF(OIF1,OIF2,……,OIF N ,Time)[N=50,m=30,Sur=on,PS=curvature,OOBP=on,OOBPI=on,Method=regression]; Where: OIF1-OIF N This represents the first to Nth terms in the optimal auxiliary factor; Set the coefficient of determination R. 2 The root mean square error (RMSE) and the mean bias (Bias) are used as model evaluation metrics, as follows: Where: Y and X are the spatially imputed downscaled AOD and the satellite remote sensing AOD, respectively; S is the number of matched samples; ∑* is the averaging operator; ∑* is the summation operator; It is the square root operator; Step 2: The dataset consisting of satellite remote sensing AOD, optimal auxiliary factors, and time variables is fixed and randomly shuffled. 60% of the dataset is used as the training set and 40% is used as the test set. The expected results of model training using the training set and modeling using the test set are verified to obtain the model accuracy. Step ③: Judge based on the model accuracy. Specifically: if the root mean square error (RMSE) of the test set is less than 0.1 and the average bias (Bias) is less than 0.001, the accuracy index meets the requirements, stop model iteration, and obtain the optimal model for AOD space filling and downscaling; otherwise, optimize and adjust the model parameters and return to step ②. The parameters for model optimization and tuning include the number of decision trees and the depth of leaf nodes. The initial number of decision trees is 50, and the minimum number of samples per leaf node is 30. The formula for parameter tuning is as follows: N r+1 =N r +10; m r+1 =m r +10; Where: N i m i These represent the number of policy trees and the minimum number of leaf nodes in the r-th iteration, respectively. The number of policy trees and the minimum number of leaf nodes are increased by 10 in each iteration. The parameters of the optimal AOD spatial imputation and downscaling model are as follows: Where N_opt and m_opt represent the optimal parameters for N and m.

8. The satellite remote sensing AOD data reconstruction method considering spatial gap filling and downscaling according to claim 7, characterized in that, Step three, AOD data reconstruction, includes: Target spatial resolution:

Images