Post-processing correction aerosol characteristic parameter estimation method based on multi-source data

Abstract

A post-processing correction aerosol characteristic parameter estimation method based on multi-source data, the estimation method comprises preparation of a data set and establishment of a post-processing correction aerosol characteristic parameter estimation model PPC-AeroNet based on the data set, the first part of the PPC-AeroNet model is a convolutional neural network CNN; the CNN is composed of three convolution blocks and uses Leaky Relu as an activation function. The data set uses MODIS, reanalysis data, surface elevation multi-source data and corresponding pixel matching, data containing missing values are filled by using a nearest neighbor method, and then data with coarse resolution is uniformly downscaled to a resolution of 1 kilometer by using a bilinear interpolation method; totally 14 input features are obtained. According to the method, three aerosol characteristic parameters with the daily resolution of 1 kilometer are seamlessly estimated, and finally the purpose of quantitative application is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing data processing, and particularly relates to a method for estimating aerosol characteristic parameter based on multi-source data after processing and correction. BACKGROUND

[0002] Atmospheric aerosol refers to solid, liquid and solid-liquid mixed particles with an aerodynamic diameter of 0.001-100 microns suspended in the atmosphere, which is generated by human and natural activities and mainly distributed in the troposphere and the stratosphere, and is an important part of the earth's atmosphere [1–3] . Aerosols can absorb and scatter solar shortwave and longwave radiation, directly affecting the balance of the earth's radiation budget, causing a decrease in ground temperature by reducing the amount of solar radiation reaching the ground, which is called the direct effect of aerosols [4] . A large amount of absorption of solar radiation by aerosols will cause the troposphere to warm up, affecting the relative humidity and stability, reducing cloud formation and life span, which is called the semi-direct effect of aerosols [5] . Aerosols can also act as cloud condensation nuclei to change the microphysical properties of clouds, such as absorption, scattering coefficient, liquid water content, and cloud droplet particle size distribution, thereby affecting the spatial and temporal distribution of solar radiation energy, which is called the indirect effect of aerosols [6] . Air pollution is a serious environmental problem faced by mankind in the 21st century [7] . Aerosol particles are the main pollutants affecting air environmental quality, and high concentrations of aerosol particles can seriously affect people's life and health and production and life [8, 9] . With the continuous development of human society, industrial pollution, urban traffic, biomass burning and other human factors have emitted a large amount of aerosols, while soil dust, forest fires, volcanic eruptions and other natural factors also produce aerosols. The sources of atmospheric aerosols are mainly divided into natural and human sources, including dust aerosols, black carbon aerosols, sulfate aerosols, sea salt aerosols and organic carbon aerosols, among which dust aerosols account for the largest proportion and have a significant impact on the environment and climate

[10] . Different particle sizes of aerosol particles have different degrees of harm, and coarse particles with a particle size of 10-100 microns, such as dust aerosols during sandstorms, not only affect atmospheric visibility, but also cause harm to the environment, transportation and human body. Humans living in a dust-polluted environment for a long time are prone to skin, respiratory and cardiovascular diseases, and in severe cases, even death; particulate matter (PM10) with a particle size of less than or equal to 10 microns mainly comes from various industrial pollution emissions, and once inhaled by the human body, it can cause various heart and lung diseases; particulate matter (PM2.5) with a particle size of less than or equal to 2.5 microns can be suspended in the air for a long time. The smaller the particle size of such fine particles, the deeper they can enter the human respiratory tract, affecting the cardiovascular, cerebral vascular and respiratory systems

[11] Aerosol affects the atmospheric environment, radiation balance, and land-atmosphere system of the entire ecosystem, and is closely related to human production and life. Its related research and monitoring have attracted widespread attention from many scholars at home and abroad.

[0003] The content of atmospheric aerosol monitoring mainly includes aerosol concentration and distribution, physical properties, optical properties, and evolution process

[12] For remote sensing monitoring of aerosols, according to the monitoring platform of the instrument, it can be divided into ground-based remote sensing and satellite remote sensing. Ground-based remote sensing is to measure direct solar radiation and scattered radiation by using a sun photometer, and to calculate the related characteristic parameters of aerosols such as AOD, AE, and spectral distribution according to the absorption and scattering of solar radiation by aerosols. Aerosol optical depth (AOD) represents the integral of the extinction coefficient of aerosols in the vertical direction, describes the attenuation of light by aerosols, represents the influence of atmospheric aerosols on solar radiation absorption and scattering, and is one of the most important optical parameters of aerosols, as well as an important parameter for estimating aerosol content and evaluating air quality

[49] The Angstrom Exponent (AE) represents the spectral dependence of AOD on incident light wavelength , which can provide information on aerosol particle size (the larger the exponent, the smaller the particle size) and aerosol phase function. It is an important parameter for describing aerosol particle size distribution and characterizing aerosol types

[50] Fine Mode Fraction (FMF) is a parameter that describes the proportion of fine particles in aerosols

[51] . Human-induced aerosols are mainly fine (mode) aerosols, while natural aerosols are mainly coarse (mode) aerosols. Studying FMF can help people better understand the impact of human-induced aerosols on the environment and climate

[52] .

[0004] Ground-based remote sensing can provide continuous real-time observation and high-precision inversion results, but the observation coverage is small, the site distribution is uneven, the number of sites is limited, the construction and maintenance cost is high, and a large amount of manpower and material resources are needed, making it difficult to achieve continuous monitoring of large areas. With the development and maturity of satellite remote sensing technology and the continuous iteration and update of high-precision sensors, satellite remote sensing can provide detailed information on the spatial variation of aerosols, has long-term and large-scale aerosol detection capabilities, and has important significance and broad research prospects for global aerosol research

[13] On the other hand, in recent years, the rise and development of artificial intelligence algorithms, data-driven machine learning methods have shown good performance in the inversion and prediction of aerosol optical thickness and particulate matter concentration. Machine learning methods have strong non-linear fitting capability for massive data and can be used as a statistical method to extract remote sensing information. Because the physical meaning of some inversion parameters is usually difficult to accurately describe, which limits the development of inversion methods based on physical models, while machine learning methods can better solve the problem of quantitative inversion of such parameters. Among them, the neural network model has more obvious advantages than other machine learning methods in the estimation accuracy of remote sensing parameters

[14] Using satellite remote sensing data and auxiliary data, using a neural network model to carry out high-precision estimation of aerosol optical thickness, Angstrom exponent, and fine mode fraction, can help promote regional and global atmospheric aerosol monitoring and research, and has important research value and practical significance.

[0005] Traditional aerosol remote sensing inversion methods are mostly based on radiation transfer processes and involve many parameters such as surface reflectivity and observation geometry. Before inversion, all possible scenarios in the target area need to be considered to establish a lookup table (LUT). According to the simulated radiation brightness of LUT, the apparent reflectivity observed by satellite is compared to estimate the aerosol characteristic parameters. The difficulty of aerosol satellite remote sensing inversion lies in the uncertainty of cloud pixel interference, aerosol model assumption, and surface reflectivity estimation

[15] Among them, how to effectively decouple the surface and atmosphere to separate the surface contribution in the satellite received signal is the key. The accuracy of surface reflectivity estimation will greatly affect the final inversion results of aerosol satellite remote sensing. Studies have shown that when the surface reflectivity error is 0.01, the AOD error will reach 0.1, i.e. about 10 times the error

[16] Due to the non-uniformity, anisotropy, and different types of surface cover of the surface, the surface reflectivity varies in space. Due to the influence of sunlight conditions, vegetation growth, and other factors, the surface reflectivity also changes over time

[17] For single-view, non-polarized satellite sensors such as MODIS, the lack of multi-angle observation information and the influence of spectral band selection make it a challenge to accurately estimate the surface reflectivity. Building LUT is time-consuming, and the lack of multi-angle observation makes it difficult to pre-set reasonable scenarios and accurate parameters to build a more accurate LUT, thereby affecting the accuracy of aerosol characteristic parameter inversion results. Therefore, designing an aerosol characteristic parameter estimation model that balances efficiency and accuracy is a current research hotspot.

[0006] Due to the low signal-to-noise ratio of the sensor, the comprehensive constraints of the inversion algorithm, the scale reduction problem of different instruments, the satellite revisit orbit period, weather conditions and surface reflectivity, etc., the aerosol data set based on satellite inversion usually has a large number of non-random missing values, which brings additional uncertainty to the data itself, which not only affects the study of aerosol characteristics, but also affects the related research and application of fine particulate matter concentration estimation and climate variable analysis [18–22] Cloud is the main factor leading to missing satellite inversion data. According to statistics, the average coverage of clouds in the global atmosphere is about 40% to 70% [23–25] . At present, there are few seamless aerosol products with high resolution, and it is necessary to carry out full-coverage aerosol characteristic parameter estimation research

[26] .

[0007] In view of the problem of missing aerosol data products, researchers have developed various methods to fill in the data. These methods can be roughly divided into two categories: one method is to use geostatistical methods to fill in the blank area, such as Kriging and its branches [27,28] , Thiessen

[29] , optimal interpolation method, Bayesian model

[30] , etc.; the other method is to fuse aerosol data products from different satellite sensors, complementing the missing original data from multiple sources, and the methods used include geostatistical inverse modeling

[31] , spatial statistical data fusion

[32] , least squares estimation, maximum likelihood estimation, etc. Geostatistical methods usually have large computational load, slow running speed, and high requirements for data quality and distribution. If the data distribution around the blank area is uneven or even has large differences, or the data is missing seriously, it will affect the filling accuracy, and the results will be too smooth, which is difficult to reflect the spatial variation of the data [31, 33] . Multi-source data fusion will be affected by the differences between data, resulting in the accuracy of the fusion results being affected, and for passive remote sensing satellite sensors, it is difficult to fundamentally rely on multiple data merging to solve the problem of data missing caused by clouds and haze

[34] .

[0008] At present, existing research uses machine learning methods to learn the errors and uncertainties existing in the existing aerosol product inversion algorithm, thereby improving and correcting the data product without processing the original radiation data, so as to improve the accuracy [35–37] . CNN benefits from its strong feature extraction capability, parameter sharing mechanism, and inductive bias, and performs better and converges faster and more stably on small sample data sets than networks based on the Transformer architecture [38,39]Inspired by this, it is decided to use the aerosol data product, combined with other auxiliary data, to build a post-processing correction model using the CNN network. SUMMARY

[0009] The application provides a post-processing correction aerosol characteristic parameter estimation method based on multi-source data, aiming to seamlessly estimate three aerosol characteristic parameters with a resolution of 1 km per day, and ultimately achieve the purpose of quantitative application.

[0010] The technical scheme adopted by the application is as follows:

[0011] A post-processing correction aerosol characteristic parameter estimation method based on multi-source data, the estimation method comprising data set preparation and establishment of a post-processing correction aerosol characteristic parameter estimation model based on the data set: Post-processing Calibration Network for Aerosol Characteristic Parameters Estimation, PPC-AeroNet; the first part of the PPC-AeroNet model is a convolutional neural network CNN; the CNN is composed of three convolutional blocks: the first convolutional block comprises a 3x3x1 three-dimensional standard convolutional layer, a lightweight temporal attention module TAM, and a 2x2x1 three-dimensional maximum pooling layer; the second convolutional block comprises a 3x3x1 three-dimensional standard convolutional layer and a 2x2x1 three-dimensional maximum pooling layer; the third convolutional block comprises a 3x3x1 three-dimensional standard convolutional layer and a three-dimensional global maximum pooling layer; Leaky Relu is used as the activation function; in the three convolutional blocks, the number of convolution kernels of the standard convolutional layer is set to 256, 128, and 64 respectively; the dimension of the original input data is 14x12x12x5, wherein 14 is the number of channels, 12x12 is the image width and height, and 5 represents the time step, i.e. five days including the current day and the previous four days.

[0012] The second part of the PPC-AeroNet model is that two LSTM layers are added after the CNN, which are used to process the one-dimensional time series output by the CNN part, and the hidden layer nodes are both set to 64; after the LSTM, two Fully Connected, FC layers with 64 hidden nodes are connected to integrate global feature information, and finally a single target parameter is output.

[0013] The dataset uses MODIS, reanalysis data, multi-source surface elevation data, corresponding pixel matching, and nearest neighbor method to fill the data with missing values, and then uses the bilinear interpolation method to uniformly reduce the resolution of the coarse resolution data to 1km resolution; 14 input features are obtained, and a 12x12 pixel image dataset centered on the target pixel is made, each picture contains 14 input feature data of the day and the previous four days, a total of 70 channel data.

[0014] The light weight time attention module TAM is composed of: one branch adopts a 1x1x1 three-dimensional convolution to enhance the nonlinearity of the input feature map, and the other branch extracts time step information through a global maximum pooling layer and a one-dimensional convolution layer, and the feature map obtained by multiplying the two branches is used as the output of the TAM.

[0015] The present application uses MCD19A2 and MERRA-2 data as input, combines other meteorological auxiliary data as constraint, and uses a post-processing correction model built by combining CNN and LSTM to realize seamless estimation of three aerosol characteristic parameters with 1km resolution per day, and finally achieves the purpose of quantitative application. The model not only uses the spatial information around the target pixel, but also considers using the time information of the day and the previous four days to capture aerosol related information from the dimensions of channel, space and time. In the CNN part, a light weight time attention module (TAM) is designed to extract local time information. After CNN, LSTM is used to further capture the long-term dependence of time series. Through three cross-validation and independent site verification, the advancement and effectiveness of the present application are proved. In comparison with existing data products, it can be known that the model can effectively correct the original data, improve the data precision, and improve the original coarse resolution to high resolution, and enrich the spatial distribution details of the target parameters. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The overall structure diagram of the present application PPC-AeroNet is shown in the figure;

[0017] Figure 2 The specific structure diagram of the present application TAM is shown in the figure;

[0018] Figure 3 The independent site verification results of different models for estimating AOD in the northwest region are shown in the figure;

[0019] Figure 4 The independent site verification results of different models for estimating AE in the northwest region are shown in the figure;

[0020] Figure 5 The independent site verification results of different models for estimating FMF in the northwest region are shown in the figure;

[0021] Figure 6 Spatial distribution comparison for AOD on August 1, 2022;

[0022] Figure 7 Spatial distribution comparison for AE on August 1, 2022;

[0023] Figure 8 Spatial distribution comparison for FMF on August 1, 2022;

[0024] Figure 9 Spatial distribution comparison for AOD on September 1, 2022;

[0025] Figure 10 Spatial distribution comparison for AE on September 1, 2022;

[0026] Figure 11 Spatial distribution comparison for FMF on September 1, 2022. DETAILED DESCRIPTION

[0027] The present application is further illustrated in theory and specific experiments in combination with the accompanying drawings.

[0028] Reference Figures 1-2 A post-processing correction aerosol characteristic parameter estimation method based on multi-source data, the estimation method includes data set preparation and establishment of a post-processing correction aerosol characteristic parameter estimation model based on the data set: Post-processing Calibration Network for Aerosol Characteristic Parameters Estimation, PPC-AeroNet.

[0029] 1. PPC-AeroNet Model: The first part of the PPC-AeroNet model is a Convolutional Neural Network (CNN). This CNN consists of three convolutional blocks: the first convolutional block contains a 3×3×1 standard 3D convolutional layer, a Temporal Attention Module (TAM), and a 2×2×1 3D max-pooling layer; the second convolutional block contains a 3×3×1 standard 3D convolutional layer and a 2×2×1 3D max-pooling layer; the third convolutional block contains a 3×3×1 standard 3D convolutional layer and a 3D global max-pooling layer. Leaky ReLU is used as the activation function. The standard 3D convolutional layer was chosen because, during experiments, its estimation performance was found to be better than that of alternatives such as depthwise convolution and depthwise separable convolution. Convolutional Long Short-Term Memory (ConvLSTM) networks are better; in the three convolutional blocks, the number of convolutional kernels in the standard convolutional layers are set to 256, 128, and 64, respectively; the original input data has a dimension of 14×12×12×5, where 14 is the number of channels, 12×12 is the image width and height, and 5 represents the time step, i.e., five days (the current day and the previous four days). After the original data is input into the model, it undergoes the first convolutional processing to map the low-dimensional data to a high-dimensional space, and then the dimensionality is reduced layer by layer, making it easier to extract rich feature information and improve the model's fitting ability. In order to learn dynamically changing features from the time step, Tan et al.

[40] Inspired by the proposed TAU module, a lightweight temporal attention module (TAM) was designed. The TAM consists of two branches: one branch uses a 1×1×1 3D convolution to enhance the non-linearity of the input feature map, and the other branch extracts temporal step information through a global max pooling layer and a 1D convolutional layer. The feature map obtained by multiplying the two branches is used as the output of the TAM. The specific structure is as follows: Figure 2 As shown. Unlike the TAU module, the designed TAM does not perform tensor reshaping, meaning it does not need to merge time steps and channels. When extracting local receptive field information from the input feature map, it removes depthwise convolutions and depthwise dilated convolutions, using only 1×1×1 three-dimensional convolutions to improve computational efficiency and enhance non-linear expressive power. Furthermore, when extracting dynamic attention between time series of the input feature map, it uses max pooling to obtain a one-dimensional time series, replaces fully connected layers with a one-dimensional convolution with a 3-kernel, and adds a sigmoid activation function to obtain time step weights, reducing computational cost and learning local temporal relationships.

[0030] The second part of the PPC-AeroNet model is two LSTM layers added after the CNN, which are used to process the one-dimensional time series output by the CNN part, and the hidden layer nodes are all set to 64. After LSTM, two Fully Connected, FC, layers with 64 hidden nodes are connected to integrate global feature information, and finally output the target parameters; the data set uses MODIS, reanalysis data, land elevation and other multi-source data for pixel matching, and uses the nearest neighbor method to fill in the data containing missing values, and then uses the bilinear interpolation method to uniformly downscale the coarse resolution data to 1 kilometer resolution. Through this series of processing, a total of 14 input features are obtained, and a 12x12 pixel image data set centered on the target pixel is made, each picture containing 14 input feature data of the day and the previous four days, a total of 70 channel data.

[0031] 2. Data set

[0032] 2.1 Data source

[0033] 2.1.1 AERONET ground observation data

[0034] The global aerosol automatic observation network AERONET uses the CIMEL CE318 automatic sun photometer as the basis observation instrument, provides 8 channels including AOD of various aerosol related parameters including AOD, for satellite inversion verification and cooperation with other databases. A total of 52 observation data of ground stations are selected in the present application, see Table 1. Among them, the time range used by 50 stations is three years from 2020 to 2022, and the data time range used by two independent stations is five years from 2009 to 2013. The AERONET observation parameters required by the present application are AOD at 550 nm, AE in the range of 440-870 nm, and FMF, which are used as true values for training the model and for testing and verification.

[0035] Since AERONET does not directly provide AOD data at 550 nm wavelength, it is necessary to obtain the AOD value at 550 nm by using the exponential band interpolation, and the specific calculation process is as follows:

[0036] (1)

[0037] (2)

[0038] (3)

[0039] the wavelength is AOD value at 440 nm and 670 nm; denotes the Angstrom exponent; is the turbidity coefficient.

[0040] In the present invention, AOD at 440 nm and 670 nm are used to calculate the AOD value at 550 nm. The data used are downloaded from the official website (https: / / aeronet.gsfc.nasa.gov).

[0041] 2.1.2 MODIS data

[0042] MODIS (Moderate Resolution Imaging Spectroradiometer) is a spectrometer with 36 bands from visible to thermal infrared (0.405-14.385 microns) on board Terra and Aqua satellites, which is suitable for remote sensing of atmospheric aerosols. Terra crosses the equator from north to south at about 10:30 local solar time, and Aqua crosses the equator from south to north at about 1:30 local solar time. Each observation covers a width of about 2330 kilometers, and the entire Earth's surface can be observed every 1 to 2 days, and the orbit is repeated every 16 days.

[0043] The MODIS data used in the present invention are MCD19A2, MCD12Q1 and MOD13A3. MCD19A2 is a Level 2 product based on the MAIAC algorithm, using MODIS data on Terra and Aqua as input, the full name of the product is "Multi-Angle Implementation of Atmospheric Correction algorithm-based Level-2 gridded aerosol optical thickness over land surfaces product". MCD19A2 provides multiple scientific data sets (SDS) at 1 km resolution per day, such as water vapor column content, AOD at 470 nm and 550 nm, etc.

[41] . MCD12Q1 product describes the land surface cover type according to one year of observation data of Terra and Aqua, providing 5 cover types with a spatial resolution of 500 meters

[42] . MOD13A3 product provides monthly normalized vegetation index at 1 km resolution. The data used are downloaded from the official website (https: / / modis.gsfc.nasa.gov).

[0044] 2.1.3 ERA5-Land data

[0045] ERA5-Land is an enhanced global reanalysis dataset generated by the European Centre for Medium-Range Weather Forecasts (ECMWF) by replaying the land component of ERA5 climate reanalysis. ERA5-Land provides hourly 0.1 ◦ ×0.1 ◦ spatial resolution (about 9 km) of land surface variables such as surface temperature, surface pressure, etc. The data used is downloaded from the official website (https: / / cds.climate.copernicus.eu).

[0046] 2.1.4 ETOPO2v2c data

[0047] ETOPO2v2c is a global, full-coverage, gridded land topography and seafloor elevation dataset. The dataset is published by the National Geophysical Data Center (NGDC) of the National Oceanic and Atmospheric Administration (NOAA) of the United States, with a spatial resolution of 2 arc minutes (about 3.7 km). The data can be downloaded from the NGDC official website (https: / / www.ngdc.noaa.gov).

[0048] Table 1 AERONET ground sites used in the present invention

[0049]

[0050] 2.1.5 MERRA-2 data

[0051] MERRA-2 (Modern-Era Retrospective Analysis for Research and Applications, Version 2) is an atmospheric reanalysis dataset developed by the Goddard Space Flight Center of the National Aeronautics and Space Administration (NASA) of the United States, providing high-resolution atmospheric data from 1980 to the present, and widely used in climate research, environmental monitoring, and weather forecasting

[43] . MERRA-2 aerosol reanalysis provides long-term aerosol parameter simulation from 1980 to the present. The present invention will use MERRA-2 tavg1_2d_aer_Nx products, which provide global hourly spatial resolution of 0.5 ◦ ×0.625 ◦The column mass density, surface mass concentration and other parameters of aerosol components (black carbon, dust, sea salt, sulfate and organic carbon). The products used are downloaded from the GES DISC website (https: / / disc.gsfc.nasa.gov). The AOD of 550 nm and the AE of 470-870 nm in the product are selected. The AE of 470-870 nm is very close to that of 440-870 nm, and the difference is negligible, so the two can be directly compared. MERRA-2 does not provide FMF parameters, which can be converted by formula 4

[44] :

[0052] (4)

[0053] In the above formula, is the total aerosol optical thickness at a wavelength of 550 nm, 、 、

[0054] 、 respectively represent the optical thickness of black carbon aerosol, organic carbon aerosol, sea salt aerosol and sulfate aerosol.

[0055] 2.2 Data set preparation

[0056] Due to the cloud interference, there are a lot of missing data in MCD19A2 product. MCD19A2 contains data from two satellites in multiple time periods, in order to improve the coverage of data, the data in these time periods are averaged as daily data. For the AOD data of 550 nm, a sliding window of 3x3 pixels (each pixel represents 1 km resolution data) is used to calculate the mean of valid data in each window to further improve the coverage and availability of data. The missing parts of the water vapor column content data are filled using the nearest neighbor method. In this way, 3x3 pixel mean AOD data of 550 nm and fully covered water vapor column content data after filling are obtained from MCD19A2 product. From MCD12Q1 product, select Land Cover Type 1. This classification is the global vegetation classification scheme defined by the International Geosphere-Biosphere Program (IGBP), which includes 11 natural vegetation categories, 3 developed and vegetation-covered land categories, and 3 non-vegetation land categories, a total of 17 land cover categories. The resolution of MCD12Q1 is 500 meters, and the corresponding data of MCD19A2 is extracted by nearest pixel matching. From the MOD13A3 product, extract the normalized vegetation index (NDVI). The ERA5-Land product only provides data on land, and for the ocean, river, and lake areas, the data is missing and needs to be filled using the nearest neighbor method, and then the data is interpolated to 1 km resolution using the bilinear interpolation method. Need to extract from the ERA5-Land product, forecast albedo, leaf area index, skin temperature, surface latent heat flux, surface pressure, 10m u-component of wind, 10m v-component of wind. MERRA-2 tavg1_2d_aer_Nx product, ETOPO2v2c product are fully covered, but the resolution is relatively rough, also need to be bilinearly interpolated to reduce the data to 1 km resolution. From the MERRA-2 product, extract AOD, AE and FMF data, and from the ETOPO2v2c product, extract elevation data. Fill the valid data in the 3x3 pixel mean AOD data of 550 nm to the corresponding position of the MERRA-2 AOD data. Through the above series of processing, a total of 14 input features are obtained, as shown in Table 2.Pictures of 12x12 pixels centered on the target pixel are made, each containing 14 input feature data of the day and the previous four days, a total of 70 channel data. Using the time range of 2021-2022, the observation data of 50 AERONET ground stations are used as the target true value, and after spatiotemporal matching of multi-source data, a total of 14702 valid data are obtained. 14702 images containing 5 time steps are prepared as a dataset for three cross-validation experiments.

[0057]

[0058] Experimental verification of the application:

[0059] 1. Ground stations and verification area

[0060] China's northwest region is located in the hinterland of the Eurasian continent, in the mid-latitude arid and semi-arid zone, with a dry climate and a shortage of water resources. In recent years, affected by global warming, the precipitation in some areas has increased, and the climate has gradually shown a trend of changing from "warm and dry" to "warm and wet" [45–48] . The northwest region has vast deserts and gobi, and dust weather is frequent. With the development of the economy in the northwest region, the increase of human activities such as industrial emissions, motor vehicle exhaust emissions, and energy consumption has also produced a large amount of aerosols. The present application selects the region from 76.7 ◦ to 106.1 ◦ east longitude and 34.1 ◦ to 41.3 ◦ north latitude in China's northwest as the verification area. The area involved includes the southern Xinjiang Uygur Autonomous Region, the northern Tibet Autonomous Region, the northern Qinghai Province, and the central Gansu Province, among others, covering parts of the Taklimakan Desert, the Tengger Desert, the Badain Jaran Desert, and other deserts, with high altitudes and complex terrain. Due to the scarcity of stations in the northwest region, the publicly available data has a relatively short time range and insufficient data samples, so a total of 50 AERONET ground stations in the eastern, central, and southern regions of Asia are selected to obtain long-term ground observation data for model training and evaluation. In addition, the Qinghai Lake-Waliwan Mountain (Mt_WLG) in Qinghai Province and the Lanzhou University Semi-Arid Climate and Environment Observation Station (SACOL) in Gansu Province are selected to obtain their observation data for independent station verification of the model.

[0061] 2. Experimental setup

[0062] The experiment uses a dataset prepared from multi-source data such as observation data of 50 AERONET sites in 2021-2022, MODIS satellite product data, ERA5-Land data, MERRA-2 data, etc. after preprocessing and spatiotemporal matching. The model is evaluated based on 10-fold cross-validation of sites, time, and samples, and independent site validation is performed using sites in the validation area that are not involved in training to further test the generalization and robustness of the model. For 10-fold cross-validation based on sites, it needs to be randomly divided into 10 groups, and each group contains data of 45 sites in the training set and data of the remaining 5 sites in the validation set. The specific grouping is shown in Table 3:

[0063] Table 3: Validation set division for 10-fold cross-validation based on sites

[0064]

[0065] The above cross-validation is a statistical method for evaluating the performance and generalization ability of machine learning models, and is also a widely used evaluation technique in the field of remote sensing

[57] . This method divides the dataset into multiple subsets and then trains and validates on these subsets multiple times to obtain more reliable performance evaluation. There are various division criteria, and the present invention performs cross-validation based on three division methods of ground stations, time, and samples to evaluate the model performance. Specifically, using ground station observation data as the target true value, according to the location of the ground station used, randomly dividing it into 10 groups of training set and validation set for 10-fold site cross-validation of the model; arranging the dataset in chronological order and dividing it into 10 groups of training set and validation set in chronological order for 10-fold time cross-validation of the model; finally, randomly dividing the dataset into 10 different groups of training set and validation set in the ratio of 8:2 for 10-fold sample cross-validation of the model. In addition, by training the model with all the dataset as training samples and testing with observation data of independent sites not used for training, the generalization and robustness of the model are further verified. The present invention uses several statistical equations such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Mean Bias Error (MBE), Pearson Correlation Coefficient (R), Index of Agreement (IOA), and Expected Error (EE) to evaluate the accuracy, bias, and correlation of the estimated results of each model.

[0066] In the experiment, the initial learning rate of the neural network model is set to 0.00005, and the Adam optimizer is used to adjust the learning rate. The number of iterations for training on the dataset is set to 300, the batch size is set to 1000, and MAE is used as the loss function. The input of the TCN model is a one-dimensional time series data containing time steps of the center pixel; the input of the ResNetReg model is only 14-dimensional feature data of the center pixel estimated on the same day. The input of the CNN model is image data of the time series. For the machine learning model, the input is also 14-dimensional data of the center pixel, and the hyperparameters are determined according to Bayesian optimization.

[0067] 3. Experimental results

[0068] 3.1 Ablation experiment

[0069] To evaluate the performance gain of the LSTM network and the TAM module of the present application on the complete model, based on the 10-fold cross-validation of the site, ablation experiments are performed on each module.

[0070]

[0071]

[0072]

[0073] Tables 4, 5 and 6 respectively show the gain of each module on the estimation of AOD, AE and FMF. It can be seen that the effect of using only CNN to estimate aerosol characteristic parameters is not good, especially in estimating AE and FMF parameters, the performance is poor in MAE, MBE and RMSE. After adding LSTM for further processing behind the CNN network, or only adding the TAM module in the CNN, the estimation results have been significantly improved in all indicators, and the complete PPC-AeroNet model can achieve the best effect, except for the MBE indicator, only the RMSE in estimating AOD ranks second, and the IOA in estimating AE ranks second, and all the remaining indicators are the best.

[0074] 3.2 Model comparison

[0075] In order to verify the advancement of the model proposed in the present application in estimating aerosol characteristic parameters, under the same experimental environment, seven mainstream algorithm models are compared with each other in estimating AOD, AE and FMF parameters through three 10-fold cross-validation and independent station verification.

[0076] 3.2.1 Cross-validation results

[0077] In the same experimental equipment environment, four machine learning models and three neural network models are compared to evaluate the advancement of the PPC-AeroNet model proposed in the application. Tables 7, 10, and 13 respectively show the station 10-fold cross-validation results of AOD, AE, and FMF, Tables 8, 11, and 14 respectively show the time 10-fold cross-validation results of the three parameters, and Tables 9, 12, and 15 respectively show the sample 10-fold cross-validation results of the three parameters.

[0078]

[0079]

[0080]

[0081]

[0082]

[0083]

[0084]

[0085]

[0086] According to the results in Tables 7 to 15, in the site 10-fold cross-validation, compared with the estimation results of three characteristic parameters, PPC-AeroNet performs best in estimating AOD, with four indicators (MAE, RMSE, IOA, and EE) being the best among all models, and the R index is only second to the LGBM model; the RF model performs best in estimating AE, with MAE, MBE, RMSE, and EE being the best, and PPC-AeroNet is the best in the IOA index, and the performance of the remaining indicators is basically in the top three among all models except the MBE index; the XGB model performs best in estimating FMF, with RMSE, R, and EE being the best, and PPC-AeroNet is the best in MAE and IOA. Overall, PPC-AeroNet performs well and stably in estimating the three parameters, and the performance of the four machine learning models in the MBE index is usually better than that of the neural network model. In the 10-fold cross-validation based on time and samples, due to the small difference between the divided training and verification data, PPC-AeroNet begins to exert its powerful nonlinear fitting capability and can use time information to improve the estimation accuracy. In the estimation results of AOD, AE, and FMF, almost all indicators are the best except that the MBE performance is slightly poor. Only in the time 10-fold cross-validation for estimating FMF, the R index is only second to the LGBM model, fully demonstrating its excellent performance. Except that EE is only 57.74% in the site 10-fold cross-validation for estimating AOD, EE of all other cross-validations reaches more than 68%.

[0087]

[0088] Table 16 shows the results of MCD19A2, MERRA-2 data products, verified according to the partitioning of sample-based 10-fold cross-validation. The 10 groups of average effective data for MCD19A2 AOD data used for verification are 2156, and the effective data used for verification of MERRA-2 are consistent with the number of effective data tested by the above eight models, which are 2941. Compared with the estimation results of PPC-AeroNet in the three aerosol characteristic parameters of 10-fold cross-validation samples, PPC-AeroNet has achieved improvement in almost all indicators for the original input data, except for the MBE and RMSE of MCD19A2 AOD data and the MBE of MERRA-2 FMF data, which have slightly decreased indicators. The EE indicators of the two AOD data are all less than 64%, and after correction, they reach 73.33%. The improvement of the other two data of MERRA-2 is also significant. The R, IOA, EE of AE data are improved from 0.6301, 0.7911, 79.82% to 0.7617, 0.8612, 90.54%, respectively, and the R, IOA, EE of FMF data are improved from 0.6878, 0.8303, 81.00% to 0.8016, 0.8882, 88.00%, respectively. This shows the effectiveness of the post-processing correction model after training. It can be seen that when the training data quality is good (that is, there is no large difference between the training data and the test data), the post-processing correction model trained with a small amount of training data (based on 10-fold cross-validation of samples, each training set is 11761 effective data) can effectively correct the original data and reduce errors.

[0089] 3.2.2 Independent site verification results

[0090] Using the observation data from 2009 to 2013 for preprocessing and spatiotemporal matching, a total of 272 effective data were obtained. Figure 3 、 4 , 5 shows the estimation results of three aerosol characteristic parameters.

[0091] From Figures 3 to 5It can be seen that in the independent station verification of AOD estimation, the four machine learning models perform well in the EE index, all above 60%, among which the LGBM model is the highest, reaching 63.24%, and the neural network model performs worse, with the highest PPC-AeroNet of only 47.43%. The PPC-AeroNet model performs well overall, with all indicators basically in the top, MAE second only to CatBoost, IOA second only to RF, MBE after CNN_LSTM and RF, and RMSE and R are the best; in the verification results of AE estimation, the MAE of PPC-AeroNet model is second only to RF, and the RMSE, R and IOA are the best; in the results of FMF estimation, the EE of ResNetReg is the highest, reaching 54.78%, while the EE of PPC-AeroNet is only 45.22%, but the MAE, RMSE, R and IOA are the best. Although the time interval between the training data and the independent verification data is as long as 7 years, and the geographical location of the verification station and the training station is also far apart, PPC-AeroNet still performs well among all models. But also because of the big difference between the training data and the verification data, the approximation error of the model fitting the original data from the training data did not show good results in the verification data, and the EE of all models in estimating the three parameters did not exceed 68%. It can be seen that the quality of the training data will greatly affect the training and application effect of deep learning and machine learning models.

[0092] Table 17 also only shows the complexity of different neural network models. From the table, it can be seen that the model complexity of each model is generally low, and PPC-AeroNet uses slightly more computing overhead to achieve better performance. PPC-AeroNet can achieve full coverage estimation and is not affected by cloud, ice and snow pixels. PPC-AeroNet is a post-processing correction of existing data products, and the sample quality of the training data is higher, and the better the sample quality, the higher the estimation accuracy of the model.

[0093]

[0094] 3.3 Spatial distribution display

[0095] In order to evaluate the spatial distribution effect of PPC-AeroNet in estimating various aerosol characteristic parameters in the verification area, the AOD of MCD19A2, the AOD, AE and FMF data of MERRA-2 are used for comparison, Figures 6 to 11 The estimation results on August 1, 2022 and September 1, 2022.

[0096] According to the estimated result map of the verification area, it can be seen that the PPC-AeroNet model effectively improves the spatial resolution of the estimated results of the three aerosol property parameters, and more delicately displays the characteristics and details of the aerosol property parameter distribution, on the basis of retaining the spatial distribution characteristics of the MERRA2 data, combining the MCD19A2 data, and supplementing information with other auxiliary data and constraints. Of course, due to the influence of using nearest neighbor to fill in missing values for coarse resolution original data and using bilinear interpolation to reduce the scale, it can still be found from the result map of the model estimation that there is a small part of the area with the outline of the coarse resolution data, which affects the overall smoothness of the data.

[0097] By consulting the literature, Huang et al. [44,53,54,55,56] The index results of the verification of MOD04_3K, MOD04_L2, MCD19A2, MERRA-2 and other products are summarized in Table 18. The MCD19A2 in the table shows the verification index of the AOD data, where A and T represent Aqua and Terra satellites respectively, and C6 and C6.1 represent Collection 6 and Collection 6.1 respectively. By comparing these indexes with the sample 10-fold cross-validation results of the model shown above, it can be found that when there is sufficient training data and the sample quality is good, the estimation performance of the PPC-AeroNet proposed in the present application can reach the level of the existing public products, and even be better, especially for AE and FMF parameters, which has the potential for quantitative estimation.

[0098] Table 18 Performance evaluation of data products

[0099]

[0100] References

[0101] [1] Lee K H, Li Z, Kim Y J, et al. Atmospheric aerosol monitoring from satellite observations: A history of three decades[J]. Atmospheric andBiological Environmental Monitoring, 2009. 13–38.

[0102] [2] Kang F G, Li Y H. Review of dust aerosol research in northwest China in the past decade[J]. Arid Meteorol, 2011, 29(2): 144-150.

[0103] [3] Tang YM, Deng RX, Xu MT, et al. Diurnal variation of aerosol optical properties in Guangzhou in autumn[J]. Journal of Sun Yat-Sen University (Natural Science Edition), 2019, 58(2):58–67.

[0104] [4] Rap A, Scott C E, Spracklen D V, et al. Natural aerosol direct and indirect radiative effects[J]. Geophysical Research Letters, 2013, 40(12):3297–3301.

[0105] [5] Procopio A, Artaxo P. Direct and semi-direct aerosol effects: a modeled study for biomass burning aerosol radiative forcing in the amazon region[C]. Proceedings of AIP Conference Proceedings, volume 1100, 649–652. American Institute of Physics, 2009.

[0106] [6] Liao L, Lou S J, Fu Y, et al. Radiative forcing and surface temperature effects of aerosols in the eastern China on the synoptic scale[J]. Chinese Journal of Atmospheric Sciences, 2015, 39(1):68–82.

[0107] [7] Jion M M M F, Islam A R M T, Shahrier M, et al. A critical review of no2 and aod in major asian cities: challenges, mitigation approaches and way forwards[J]. Air Quality, Atmosphere & Health, 2024.1–17.

[0108] [8] She L. Aerosol optical depth retrieval and dust monitoring based on himawari-8 / ahi data[D]. Ph.D. Dissertation]. Beijing: University of Chinese Academy of Sciences (Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences), 2018.

[0109] [9] Wang Y, Ali MA, Bilal M, et al. Identification of no2 and so2pollution hotspots and sources in jiangsu province of china[J]. RemoteSensing, 2021, 13(18):3742.

[0110]

[10]

[10] Zhao H, Gui K, Wang Y, et al. Long-term distribution and evolution trends of absorption aerosol optical depth with different chemical components in global and typical regions[J]. Atmospheric Research, 2025, 314:107819.

[0111]

[11] Su Danfeng, Zeng Sanwu. Study on the hazards of fine particulate matter PM2.5 to various systems of the human body [J]. Medical Information, 2019, 32(18):32–34.

[0112]

[12] Zhang Chunyang. Study on the inversion of aerosol characteristic parameters in Kunming area using CE318 observation data [D]. Kunming: Yunnan University, 2017.

[0113]

[13] Tang Yuming, Deng Ruru, Liu Yongming, et al. A review of research on remote sensing inversion of atmospheric aerosols [J]. Remote Sensing Technology and Application, 2018, 33(1):25–34.

[0114]

[14] Fan Y, Sun L. Satellite aerosol optical depth retrieval based on fully connected neural network (fcnn) and a combine algorithm of simplified aerosol retrieval algorithm and simplified and robust surface reflectance estimation (sremara) [J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2023, 16:4947-4962.

[0115]

[15] Li Z, Zhao X, Kahn R, et al. Uncertainties in satellite remote sensing of aerosols and impact on monitoring its long-term trend: a review and perspective [C]. Proceedings of Annales Geophysicae, volume 27, 2755-2770. Copernicus GmbH, 2009.

[0116]

[16] Wang Y, Xue Y, Guang J, et al. Simultaneously retrieval of aerosol optical depth and surface albedo with fy-2 geostationary data [C]. Proceedings of 2011 IEEE International Geoscience and Remote Sensing Symposium, 2912-2914. IEEE, 2011.

[0117]

[17] He T, Liang S, Wang D, et al. Estimation of surface albedo and directional reflectance from moderate resolution imaging spectroradiometer (modis) observations [J]. Remote Sensing of Environment, 2012, 119:286-300.

[0118]

[18] Singh M K, Venkatachalam P, Gautam R, et al. Geostatistical methods for filling gaps in level-3 monthly-mean aerosol optical depth data from multi-angle imaging spectroradiometer [J]. Aerosol and Air Quality Research, 2017, 17(8):1963-1974.

[0119]

[19] Han M, Jia S, Zhang C. Estimation of high-resolution pm2.5 concentrations based on gap filling aerosol optical depth using gradient boosting model [J]. Air Quality, Atmosphere & Health, 2022, 15(4):619-631.

[0120]

[20] Zhang R, Di B, Luo Y, et al. A nonparametric approach to filling gaps in satellite-retrieved aerosol optical depth for estimating ambient pm2.5 levels [J]. Environmental Pollution, 2018, 243:998-1007.

[0121]

[21] Zhao C, Liu Z, Wang Q, et al. High-resolution daily aodestimated to full coverage using the random forest model approach in thebeijing-tianjin-hebei region[J]. Atmospheric Environment, 2019, 203:70–78.

[0122]

[22] Wang W, He Q, Zhang M, et al. Full-coverage 1-km estimates and spatiotemporal trends of aerosol optical depth over taiwan from 2003 to 2019[J]. Atmospheric Pollution Research, 2022, 13(11):101579.

[0123]

[23] Yi B, Li R, Yang Z. Global distributions of multi-layer and multi-phase clouds and their cloud radiative effects[J]. Journal ofQuantitative Spectroscopy and Radiative Transfer, 2025, 333:109334.

[0124]

[24] Zhu Z, Qiu S, He B, et al. Cloud and cloud shadow detection forlandsat images: The fundamental basis for analyzing landsat time series[J].Remote Sensing Time Series Image Processing, 2018. 3–23.

[0125]

[25] He Junliang. Research on the influence of cloud proximity effect on remote sensing inversion of aerosol optical thickness and its reduction method [J]. Acta Geodaetica et Cartographica Sinica, 2020, 49(1):132.

[0126]

[26] Guo B, Wang Z, Pei L, et al. Reconstructing modis aerosol optical depth and exploring dynamic and influential factors of aod via random forest at the global scale[J]. Atmospheric Environment, 2023, 315:120159.

[0127]

[27] Yang J, Hu M. Filling the missing data gaps of daily modis aod using spatiotemporal interpolation[J]. Science of the Total Environment, 2018, 633:677-683.

[0128]

[28] Carrera-Hernández J, Gaskin S. Spatio temporal analysis of daily precipitation and temperature in the basin of mexico[J]. Journal of Hydrology, 2007, 336(3-4):231-249.

[0129]

[29] Buytaert W, Celleri R, Willems P, et al. Spatial and temporal rainfall variability in mountainous areas: A case study from the south ecuadorian andes[J]. Journal of Hydrology, 2006, 329(3-4):413-421.

[0130]

[30] Kaplan A, Kushnir Y, Cane M A, et al. Reduced space optimal analysis for historical data sets: 136 years of atlantic sea surface temperatures[J]. Journal of Geophysical Research: Oceans, 1997, 102(C13):27835-27860.

[0131]

[31] Wang J, Brown D G, Hammerling D. Geostatistical inverse modeling for super-resolution mapping of continuous spatial processes[J]. Remote Sensing of Environment, 2013, 139:205-215.

[0132]

[32] Nguyen H, Cressie N, Braverman A. Spatial statistical data fusion for remote sensing applications[J]. Journal of the American Statistical Association, 2012, 107(499):1004-1018.

[0133]

[33] Liu G, Wang Y, Zhang H, et al. Comparative study on several interpolation methods in spatial analysis[J]. Geographic Information World, 2011, 9(3):41-45.

[0134]

[34] Xia X, Zhao B, Zhang T, et al. Satellite-derived aerosol optical depth fusion combining active and passive remote sensing based on bayesian maximum entropy[J]. IEEE Transactions on Geoscience and Remote Sensing, 2021, 60:1-13.

[0135]

[35] Lipponen A, Kolehmainen V, Kolmonen P, et al. Model enforced post-process correction of satellite aerosol retrievals[J]. Atmospheric Measurement Techniques Discussions, 2020, 2020:1-17.

[0136]

[36] Lanzaco B L, Olcese L E, Palancar G G, et al. An improved aerosol optical depth map based on machine-learning and modis data: Development and application in south america[J]. Aerosol and Air Quality Research, 2017, 17(6):1623-1636.

[0137]

[37] Lary D J, Remer L, MacNeill D, et al. Machine learning and bias correction of modis aerosol optical depth[J]. IEEE Geoscience and Remote Sensing Letters, 2009, 6(4):694-698.

[0138]

[38] Li S, Chen X, He D, et al. Can vision transformers perform convolution?[J]. ArXiv Preprint ArXiv:2111.01353, 2021.

[0139]

[39] d’Ascoli S, Touvron H, Leavitt M L, et al. Convit: Improving vision transformers with soft convolutional inductive biases[C]. Proceedings of International Conference on Machine Learning, 2286-2296. PMLR, 2021.

[0140]

[40] Tan C, Gao Z, Wu L, et al. Temporal attention unit: Towards efficient spatiotemporal predictive learning[C]. Proceedings of Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition, 18770-18782, 2023.

[0141]

[41] Martins V S, Lyapustin A, Wang Y, et al. Global validation of columnar water vapor derived from eos modis-maiac algorithm against the ground-based aeronet observations[J]. Atmospheric Research, 2019, 225:181-192.

[0142]

[42] Zhao J, Dong Y, Zhang M, et al. Comparison of identifying land cover tempo-spatial changes using globcover and mcd12q1 global land cover products[J]. Arabian Journal of Geosciences, 2020, 13:1-12.

[0143]

[43] Randles C, Da Silva A, Buchard V, et al. The merra-2 aerosol reanalysis, 1980 onward. part i: System description and data assimilation evaluation[J]. Journal of Climate, 2017, 30(17):6823-6850.

[0144]

[44] Su X, Huang Y, Wang L, et al. Validation and diurnal variation evaluation of merra-2 multiple aerosol properties on a global scale[J]. Atmospheric Environment, 2023, 311:120019.

[0145]

[45] Yihui D, Yanju L, Ying X, et al. Regional responses to global climate change: progress and prospects for trend, causes, and projection of climatic warming-wetting in northwest china[J]. Advances in Earth Science, 2023, 38(6):551.

[0146]

[46] Yao X, Zhang M, Zhang Y, et al. New insights into climate transition in northwest china[J]. J. Arid. Land, 2022. 1-15.

[0147]

[47] Wang C, Zhang S, Li K, et al. Change characteristics of precipitation in northwest china from 1961 to 2018[J]. Chin. J. Atmos. Sci, 2021, 45(4):713-724.

[0148]

[48] Yang J, Jiang Z, Liu X, et al. Influence research on spring vegetation of eurasia to summer drought-wetness over the northwest china [J]. Arid Land Geography, 2012, 35(1): 10-22.

[0149]

[49] Li F G, Lin C Y, Li L G. Review of aerosol optical depth research in northwest China [J]. China Meteorological Society 2007 Annual Conference Atmospheric Composition Observation, Research and Forecast Branch Session Papers, 2007.

[0150]

[50] Jung C H, Lee J Y, Kim Y P. Changes in the ångstrom exponent during aerosol coagulation and condensation [J]. Asian Journal of Atmospheric Environment, 2012, 6(4): 304-313.

[0151]

[51] Kleidman R, O'Neill N, Remer L, et al. Comparison of moderate resolution imaging spectroradiometer (modis) and aerosol robotic network (aeronet) remote-sensing retrievals of aerosol fine mode fraction over ocean [J]. Journal of Geophysical Research: Atmospheres, 2005, 110(D22).

[0152]

[52] Zheng F, Hou W, Sun X, et al. Optimal estimation retrieval of aerosol fine-mode fraction from ground-based sky light measurements [J]. Atmosphere, 2019, 10(4): 196.

[0153]

[53] Wei J, Li Z, Sun L, et al. Modis collection 6.1 3 km resolution aerosol optical depth product: Global evaluation and uncertainty analysis[J]. Atmospheric Environment, 2020, 240:117768.

[0154]

[54] Wei J, Li Z, Peng Y, et al. Modis collection 6.1 aerosol optical depth products over land and ocean: validation and comparison[J]. Atmospheric Environment, 2019, 201:428-440.

[0155]

[55] Huang G, Su X, Wang L, et al. Evaluation and analysis of long-term modis maiac aerosol products in china[J]. Science of The Total Environment, 2024, 948:174983.

[0156]

[56] Ansari K, Ramachandran S. Optical and physical characteristics of aerosols over asia: Aeronet, merra-2 and cams[J]. Atmospheric Environment, 2024, 326:120470.

[0157]

[57] Cao M, Zhang M, Su X, et al. A two-stage machine learning algorithm for retrieving multiple aerosol properties over land: Development and validation[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023.

Claims

1. A method for estimating aerosol characteristic parameters based on post-processing correction of multi-source data, characterized in that, The estimation method comprises data set preparation and establishment of a post-processing calibration network for aerosol characteristic parameters estimation (PPC-AeroNet) based on the data set; the first part of the PPC-AeroNet model is a convolutional neural network (CNN); the CNN comprises three convolutional blocks: the first convolutional block comprises a 3×3×1 three-dimensional standard convolutional layer, a lightweight temporal attention module (TAM), and a 2×2×1 three-dimensional maximum pooling layer; the second convolutional block comprises a 3×3×1 three-dimensional standard convolutional layer and a 2×2×1 three-dimensional maximum pooling layer; and the third convolutional block comprises a 3×3×1 three-dimensional standard convolutional layer and a three-dimensional global maximum pooling layer; a Leaky Relu is used as an activation function; in the three convolutional blocks, the number of convolution kernels of the standard convolutional layers is set to 256, 128, and 64, respectively; and the dimension of the original input data is 14×12×12×5, wherein 14 is the number of channels, 12×12 is the image width and height, and 5 represents the time step, i.e., five days including the current day and the previous four days; The second part of the PPC-AeroNet model is that two LSTM layers are added after the CNN, which are used to process the one-dimensional time sequence output by the CNN part, and the number of hidden layer nodes is set to 64; after the LSTM, two fully connected (FC) layers with 64 hidden nodes are connected, which integrate global feature information, and finally output a single target parameter; The data set uses MODIS, reanalysis data, and multi-source surface elevation data, matches corresponding pixels, uses the nearest neighbor method to fill in missing values, and uses the bilinear interpolation method to uniformly reduce the resolution of the coarse resolution data to 1 km; a total of 14 input features are obtained, a 12×12 pixel image data set centered on the target pixel is made, and each picture contains 14 input feature data of the current day and the previous four days, a total of 70 channel data.

2. The method according to claim 1, wherein, The lightweight temporal attention module (TAM) comprises: one branch adopts a 1×1×1 three-dimensional convolution to enhance the nonlinearity of the input feature map, and the other branch extracts time step information through a global maximum pooling layer and a one-dimensional convolutional layer; and the feature map obtained by multiplication of the two branches is used as the output of the TAM.

Images