A high temporal resolution spatially seamless aerosol optical depth filling method

Abstract

The present invention discloses a high-temporal-resolution spatially seamless aerosol optical depth filling method, belonging to the technical field of aerosol optical depth filling. The present invention decomposes the geostationary satellite aerosol optical depth product into a linear component and a nonlinear component. The linear component is solved by solving the linear relationship between the reanalysis data and the satellite-inverted aerosol optical depth product for different surface cover types, and correcting the reanalysis data based on this linear relationship; the corrected linear component is subtracted from the satellite aerosol optical depth to obtain the nonlinear component of the effective observation pixel, and a machine learning model is used to map the correlation between the nonlinear component and the influencing factor. The relationship is applied to non-effective observation pixels to obtain the spatially seamless nonlinear component; finally, the linear component and the nonlinear component are superimposed to obtain the spatially seamless aerosol optical depth.

Description

Technical Field

[0001] The present invention belongs to the technical field of aerosol optical depth filling, and in particular relates to a high temporal resolution spatial seamless aerosol optical depth filling method. Background Art

[0002] Algorithms and scientific data products for retrieving aerosol optical depth (AOD) from satellite visible light observations have matured. However, the susceptibility of satellite visible light bands to cloud and fog, as well as outliers in satellite surface reflectance products, result in significant non-random missing values ​​in satellite-derived AOD products. To address this issue, studies have integrated satellite-derived AOD products with reanalysis data, along with surface, atmospheric, and socioeconomic parameters. Based on the assumption of continuity in the spatiotemporal distribution of AOD, these missing values ​​are then filled. However, the vast majority of these studies rely on polar-orbiting satellite products with higher spatial resolution. This is because, due to limitations in sensor hardware, the coarse spatial resolution of geostationary satellites made it difficult to accurately reflect the spatial distribution of AOD. Consequently, few studies of filling algorithms used geostationary satellites as input parameters. However, with the launch of the new-generation geostationary satellite, Himawari-8, the spatial resolution of geostationary satellites has significantly improved, and their AOD products now meet application requirements. A small number of studies have been conducted to transplant the aerosol optical depth filling algorithm developed on polar-orbiting satellites to geostationary satellites, but the following problems still exist:

[0003] (1) Polar-orbiting satellites have a coarse temporal resolution. Most polar-orbiting satellites only have one valid AOD inversion result per day, while geostationary satellites have a higher temporal resolution, with valid observations every hour or less. The AOD filling algorithm developed based on polar-orbiting satellites does not take into account the differences in AOD patterns at different local times, and therefore cannot be directly transferred to geostationary satellite AOD products that have inversion results at multiple local times.

[0004] (2) The main way to use surface cover type information and reanalysis data aerosol optical depth products in the existing aerosol optical depth filling algorithm is to use them as input parameters of the machine learning model. However, this method is not sufficient to characterize the differences in the distribution patterns of aerosol optical depth under different surface cover types, nor does it effectively utilize the linear relationship between the reanalysis data aerosol optical depth product and the satellite inversion aerosol optical depth product. Summary of the Invention

[0005] The present invention provides a high-temporal-resolution spatially seamless aerosol optical depth filling method to achieve high-temporal-resolution and spatially seamless aerosol optical depth acquisition.

[0006] The technical solution adopted in the present invention is:

[0007] A high temporal resolution spatially seamless aerosol optical depth filling method, the method comprising the following steps:

[0008] Step 1: Train the regression model of the aerosol optical depth influencing factors and aerosol optical depth at low spatial resolution:

[0009]

[0010] in, represents the low spatial resolution aerosol optical depth obtained by regression model fitting, kernel low represents the low spatial resolution aerosol optical depth impact factor, and f() represents the downscaling mapping relationship of each input variable at low spatial resolution. It should be noted that in this invention, high (low) temporal resolution space and low (high) spatial resolution are artificial divisions. If the temporal resolution of the current object is higher than (or equal to) the set temporal resolution threshold, it is a high temporal resolution space, otherwise it is a low temporal resolution space. If the spatial resolution of the current object is lower than (or equal to) the set spatial resolution threshold, it is a low spatial resolution space, otherwise it is a high spatial resolution space. The thresholds for the division (temporal resolution threshold and spatial resolution threshold) are set based on the scenario.

[0011] Define the residual ΔAOD between the original aerosol optical depth at low resolution and the aerosol optical depth obtained by regression model fitting low for: Among them, AOD low represents the low-resolution raw aerosol optical depth;

[0012] The regression model trained at low spatial resolution is applied to high spatial resolution to obtain the predicted high-resolution aerosol optical depth results. Among them, kernel high Indicates the aerosol optical depth factor with high spatial resolution;

[0013] ΔAOD low The high spatial resolution residual ΔAOD is obtained by bilinear sampling to high spatial resolution. high ;

[0014] According to the formula Get the AOD of a single downscaling model high ;

[0015] Step 2, linear component (LC) estimation:

[0016] The linear relationship between the reanalysis data aerosol optical depth and the satellite aerosol optical depth is fitted based on the least squares method: AOD LC,i =A i AOD merra2,i +B i ;

[0017] Among them, AOD LC,i Represents the linear component of a certain land cover type, A i 、B i are the slope and intercept of the linear relationship fitted on a certain land cover type, AOD merra2,i represents the aerosol optical depth of the reanalysis data on a certain land cover type; i represents different land cover types;

[0018] Step 3, nonlinear component (NC) estimation:

[0019] The relationship between the nonlinear components and their influencing factors is mapped using a machine learning algorithm to obtain the mapping relationship at the non-missing value pixel: AOD NC-clr,j =RF j (Φ)+ε j , where Φ represents the influence factor of the nonlinear component, j represents different places, and ε j Represents the residual at local time j, RF j () represents the mapping relationship between nonlinear components and their influencing factors;

[0020] According to the formula AOD NC-gap-filled,j =RF j (Φ)+ε j Calculate the predicted value of the nonlinear component of aerosol optical depth (AOD) at the missing pixel at local time j NC-gap-filled,j ;

[0021] Step 4: Superimpose the estimated linear and nonlinear components to obtain the spatially seamless aerosol optical depth estimate:

[0022] AOD NC-gap-filled,j =A i AOD merra2,i +B i +RF j (Φ)+ε j

[0023] Among them, AOD NC-gap-filled,j It represents the aerosol optical depth after filling, thereby obtaining the high temporal resolution and spatially seamless aerosol optical depth after filling.

[0024] Furthermore, the low spatial resolution aerosol optical depth factor kernel low and high spatial resolution aerosol optical depth factor kernel high The same factors affecting AOD are included: vegetation index, land cover type, elevation, slope, wind speed, temperature, humidity and air pressure.

[0025] Furthermore, before training the regression model at low spatial resolution, each aerosol optical depth influencing factor was upsampled to the same spatial resolution as the aerosol optical depth of the reanalysis data.

[0026] Furthermore, upsampling is performed using area weighting: Among them, S sum Represents the low spatial resolution pixel area; S i represents the area of ​​the i-th high spatial resolution pixel within the low spatial resolution pixel; x i It represents the value of the aerosol optical depth impact factor corresponding to the i-th high spatial resolution pixel within the low spatial resolution pixel.

[0027] Furthermore, the influencing factors Φ of the nonlinear component include: temperature T 2m , water vapor TQV, horizontal direction (U direction) wind speed U 2m , vertical direction (V direction) wind speed V 2m , planetary boundary layer height PBLH, air pressure PS, elevation DEM, land cover type LC, normalized vegetation index NDVI, sensor azimuth SAA, sensor zenith angle SZA, annual cumulative day DOY, latitude Lat, longitude Lon, geographical distance D from the pixel to the center of the study area center and population density pop.

[0028] Furthermore, the geographical distance D center The calculation formula is:

[0029]

[0030] Among them, γ and φ represent the longitude and latitude of the pixel, respectively. The subscripts of γ and φ are used to distinguish the two pixels or geographical locations to be calculated, and r represents the radius of the earth.

[0031] The technical solution provided by the present invention brings at least the following beneficial effects:

[0032] The present invention can use the re-classified data to fill the missing values ​​of aerosol optical thickness of geostationary satellites with high time resolution;

[0033] The present invention decomposes the surface temperature into a normal component and a sub-normal component from the perspective of information decomposition of reanalysis data aerosol optical thickness and geostationary satellite. The high temporal resolution characteristic can be used to constrain the reconstructed value.

[0034] The present invention takes into account the variation of aerosol optical depth at different locations and times and for different types of land cover. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0036] Figure 1 Schematic diagram of the processing process of the method of the present invention.

[0037] Figure 2 This is a histogram of the test set accuracy of the model established by the method of the present invention at different locations and different surface cover types. Figure 2 In the figure, (a) is MBE; (b) is RMSE; (c) is R; and (d) is EE.

[0038] Figure 3 Comparison results of the Himawari-8AHI aerosol optical density (AOD) and the padded AOD obtained using the present method, according to a specific embodiment of the present invention (the left column shows images of the Himawari-8AHI AOD at 03:00 (UTC) on the 60th, 206th, 270th, and 344th days of 2016; the right column shows images of the padded AOD obtained using the present method at 03:00 (UTC) on the 60th, 206th, 270th, and 344th days of 2016). DETAILED DESCRIPTION

[0039] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0040] An embodiment of the present invention provides a high-temporal-resolution spatially seamless aerosol optical depth filling method. By utilizing the linear relationship between the aerosol optical depth product of reanalysis data and the aerosol optical depth product retrieved from satellites, combined with the spatiotemporal distribution characteristics of the aerosol optical depth product retrieved from satellites, the aerosol optical depth retrieved from satellites is decomposed into linear and nonlinear components. The non-random missing elements in the aerosol optical depth inversion product of geostationary satellites are filled, thereby achieving high-temporal-resolution and spatially seamless aerosol optical depth acquisition.

[0041] Most existing satellite aerosol optical depth filling methods are targeted at polar-orbiting satellites, without considering the high temporal resolution characteristics of geostationary satellites. The embodiment of the present invention considers the significant variation of geostationary satellite aerosol optical depth products with surface cover type and local time, as well as their correlation with reanalysis data, and decomposes the geostationary satellite aerosol optical depth product into a linear component (LC) and a nonlinear component (NC). The linear component is solved by solving the linear relationship between the reanalysis data and the satellite-inverted aerosol optical depth product for different surface cover types, and correcting the reanalysis data based on the linear relationship. The corrected linear component is subtracted from the satellite aerosol optical depth to obtain the nonlinear component of the effective observation pixel. A machine learning model is used to map the correlation between the nonlinear component and the influencing factor, and the relationship is applied to the ineffective observation pixel to obtain the spatially seamless nonlinear component. Finally, the linear component and the nonlinear component are superimposed to obtain the spatially seamless aerosol optical depth.

[0042] like Figure 1 As shown, the high temporal resolution spatial seamless aerosol optical depth filling method provided by the embodiment of the present invention includes:

[0043] (1) Downscaling of the aerosol optical depth of the reanalysis data. The downscaling process is to first train the regression model f of the aerosol optical depth influencing factors and aerosol optical depth at low spatial resolution:

[0044]

[0045] Where, the subscript low represents low spatial resolution, kernel low Representative factors affecting aerosol optical depth include vegetation index, land cover type, elevation, slope, wind, temperature, humidity, and pressure. It refers to the low spatial resolution aerosol optical depth obtained by fitting the trained model. That is, f() represents the downscaling mapping relationship of each input variable at low spatial resolution. The model established between the influencing factors and aerosol optical depth cannot fully explain the scale variation of aerosol optical depth, and the residual can represent factors not taken into account by the model. Therefore, the model needs to consider the residual ΔAOD low :

[0046]

[0047] Where AOD low is the low-resolution original aerosol optical depth; ΔAOD lowis the residual between the original aerosol optical thickness at low resolution and the aerosol optical thickness obtained by model fitting. Based on the "relationship scale invariance" hypothesis, the model trained at low resolution is applied to high resolution to obtain the predicted high-resolution aerosol optical thickness result.

[0048]

[0049] Where, the subscript high represents high spatial resolution; kernel high Represents high spatial resolution vegetation index, land cover type, elevation, slope, wind, temperature, humidity, pressure and other factors affecting aerosol optical depth. Since the spatial resolution of aerosol optical depth in reanalysis data is much coarser than that of kernel high Therefore, before using the above method to train the regression model at low spatial resolution, each influencing factor needs to be upsampled to the same spatial resolution as the aerosol optical depth of the reanalysis data. In this invention, the area weighted method is used for upsampling:

[0050]

[0051] Where S sum Represents the low-resolution pixel area; S i represents the area of ​​the i-th high-resolution pixel within the low-resolution pixel; x i is the value of the impact factor corresponding to the i-th high-resolution pixel in the low-resolution pixel. low Sampling to high spatial resolution ΔAOD by bilinear method high . Add the residual ΔAOD high You can get the single downscaling model result AOD high :

[0052]

[0053] (2) Estimation of linear component LC.

[0054] In this paper, the least squares method is used to fit the linear relationship between the aerosol optical depth of the reanalysis data and the satellite aerosol optical depth for different surface cover types, and the aerosol optical depth of the reanalysis data corrected by this linear relationship is used as the linear classification LC.

[0055] AOD LC,i =A i AOD merra2,i +B i (i=0,1,2...16) (6)

[0056] Where AOD LC,i is the linear component of a certain land cover type; A i 、B i are the slope and intercept of the linear relationship fitted on a certain land cover type; AOD merra2,i is the aerosol optical depth of the reanalysis data on a certain surface cover type; i represents a different surface cover type, and 0-16 correspond to the different surface cover type numbers in the IGBP (International Geosphere-biosphere Programme).

[0057] (3) Estimation of nonlinear component NC. Use machine learning algorithms (including but not limited to random forest, lightGBM, neural network, convolutional neural network and other algorithms) to map the relationship between NC and its influencing factors. For the convenience of description, we use random forest as an example here. The NC corresponding to the non-missing value pixel of the geostationary satellite is obtained by subtracting the aerosol optical thickness of the geostationary satellite from the linear component obtained in the second stage. Then, the NC is classified according to local time. A random forest model is established for the NC of each local time to map the relationship between NC and related factors, and this relationship is used to generate NC at the missing value pixel. The mapping relationship at the non-missing value pixel can be simply expressed as follows:

[0058]

[0059] Where AOD NC-clr,j represents the nonlinear component of the aerosol optical depth at the non-missing pixel at local time j, T 2m 、TQV、U 2m 、V 2m ,PBLH,PS,DEM,LC,NDVI,SAA,SZA,DOY,Lat,Lon,D center and pop are respectively air temperature, water vapor, wind speed in the U direction, wind speed in the V direction, planetary boundary layer height, air pressure, elevation, land cover type, normalized vegetation index, sensor azimuth, sensor zenith angle, annual cumulative days, latitude, longitude, geographical distance from the pixel to the center of the study area, and population density; j is the time at different places; ε j is the residual when the place is j. Formula (7) is used to characterize the model trained using data from non-missing regions.

[0060] The formula for calculating geographic distance is as follows:

[0061]

[0062] Wherein, formula (8) is used to calculate the geographic distance between two pixels (or geographic locations), DIS is the geographic distance, asin() represents the inverse sine function, γ and φ represent the longitude and latitude of the pixel, respectively. The subscripts of γ and φ are used to distinguish the two pixels (or geographic locations) to be calculated, and r represents the radius of the Earth (approximately 6371 km). When the mapping factor used is sufficient for the NC response, the residual can be ignored to a certain extent. Then, the aerosol optical depth after filling in any missing value pixel M is:

[0063]

[0064] AOD NC-gap-filled,j represents the random forest model prediction value of the nonlinear component of aerosol optical depth at the missing pixel at local time j, that is, formula (9) expresses: the model trained using formula (7) is used to predict the missing part.

[0065] (4) Estimation of spatially seamless aerosol optical depth. The LC and NC obtained in stages 2 and 3 are superimposed to obtain the filled high-temporal-resolution and spatially seamless aerosol optical depth.

[0066]

[0067] Where AOD gap-filled,i,j is the filled aerosol optical depth of surface cover type i at local time j; AOD LC,i is the LC component on land cover type i; AOD NC,j is the NC component at local time j. Substituting equations (6) and (9) into equation (10) yields the final aerosol optical depth (AOD) after filling: NC-gap-filled,j , whose expression is:

[0068]

[0069] The high-temporal-resolution, spatially seamless aerosol optical depth filling method provided in an embodiment of the present invention is based on the integration of geostationary satellite aerosol optical depth (AOD) and reanalysis data. To further validate its filling performance, the embodiment of the present invention was applied to the Himawari-8 L3-level hourly aerosol optical depth product and the MERRA-2 reanalysis data hourly aerosol optical depth product, filling a large number of non-random missing values ​​in the Himawari-8 L3-level aerosol optical depth product. The method implemented in this embodiment utilizes the strong linear correlation between reanalysis data and satellite-derived products to decompose the satellite-derived aerosol optical depth product into linear and nonlinear components. The linear component is corrected and constrained for different land cover types, and the nonlinear component is fitted using a machine learning algorithm. The linear and nonlinear components are superimposed to obtain the filled, high-temporal-resolution, spatially seamless aerosol optical depth data.

[0070] The data selected in this embodiment mainly include the L3 AOD product of the Himawari-8 Advanced Himawari Imager (AHI), the latest version of the global atmospheric reanalysis dataset (MERRA-2) produced by the NASA Global Modeling and Assimilation Office using the Goddard Earth Observing System model version 5.12.4, and the main parameter input parameter set can be divided into the following types: (1) aerosol optical depth product of MERRA-2 reanalysis data; (2) atmospheric parameters, including cloud cover, 2m air temperature, surface pressure, 2m U / V wind speed and planetary boundary layer height provided by ERA5 reanalysis data; (3) surface parameters, including normalized vegetation index, surface cover type and elevation; (4) observation geometry parameters, including sensor zenith angle and sensor azimuth; (4) spatiotemporal parameters, including pixel longitude and latitude and the geographical distance from the pixel to the center of the study area, local time, and annual cumulative day; (5) socioeconomic parameters, namely population density. The above data is preprocessed to unify the spatiotemporal resolution and perform spatiotemporal matching. The implementation can be divided into the following four steps.

[0071] (1) Downscaling of aerosol optical depth products from reanalysis data

[0072] First, the vegetation index, land cover type, elevation, slope, wind, temperature, humidity, pressure and other AOD influencing factors used to construct the coarse-resolution downscaling regression model in Equation (2) are upsampled to the same spatial resolution as the MERRA-2 AOD product using Equation (3). Then, a machine learning algorithm is used to fit the regression relationship in Equation (2), and this regression relationship is substituted into Equation (4). The high-resolution AOD influencing factors are used as input parameters to obtain the downscaled MERRA-2 AOD product, which is used as the input for the second step.

[0073] (2) Linear component LC estimation

[0074] The effective observation pixel samples of Himawari-8AHI aerosol optical depth and the corresponding MERRA-2 aerosol optical depth samples are obtained, and the samples are divided according to the surface cover type. The coefficient A in equation (6) is calculated using the least squares method under different surface cover types. i 、B i The corrected MERRA-2 aerosol optical depth is obtained by fitting as the linear component LC of the padded aerosol optical depth product.

[0075] (3) Nonlinear component NC estimation

[0076] The nonlinear component NC at the effective observation pixel of Himawari-8AHI aerosol optical depth is subtracted from the linear component obtained in the second step. Then, according to formula (7), a machine learning algorithm (random forest is used in this example) is used to map the relationship between the effective observation pixel NC and the corresponding influencing factors (the main factors used in this example are air temperature, water vapor, wind speed in the U direction, wind speed in the V direction, planetary boundary layer height, air pressure, elevation, surface cover type, normalized vegetation index, sensor azimuth, sensor zenith angle, annual cumulative days, latitude, longitude, geographical distance from the pixel to the center of the study area, and population density). Finally, based on the trained machine learning model, the nonlinear component NC of the ineffective observation pixel after filling is obtained (formula (9)).

[0077] (4) High temporal resolution and spatially seamless aerosol optical depth filling

[0078] The linear component LC obtained in the second step and the padded nonlinear component NC obtained in the third step are superimposed to obtain the padded high-time-resolution spatial seamless aerosol optical depth. Figure 2The model accuracy of the proposed method on different surface cover types is shown in the figure. It can be seen from the figure that the proposed method has a good model training effect on most surface cover types. To further verify the accuracy of the proposed method, the surface station data and the high temporal resolution spatial seamless aerosol optical depth product obtained by the proposed method are used for verification. Figure 3 As shown in the figure, and compared with the existing MERRA-2 aerosol optical depth product, the results show that the verification accuracy of the MERRA-2 aerosol optical depth product at the selected 32 stations is low and there is a systematic underestimation. However, the high temporal resolution spatially seamless aerosol optical depth product generated by the method of the present invention is more accurate than the MERRA-2 aerosol optical depth product and has no significant systematic deviation.

[0079] In response to the shortcomings of existing methods, the present invention decomposes the geostationary satellite aerosol optical depth product into linear components and nonlinear components, and solves them separately. In the process of solving the linear component, the linear correlation between the reanalysis data aerosol optical depth product and the geostationary satellite aerosol optical depth product is used, and the reanalysis data data is corrected separately using satellite data at different surface cover types. Due to the spatial seamlessness of the reanalysis data, there is no spatial missing in the correction result. As for the nonlinear component, the powerful nonlinear fitting ability of machine learning is used to map the correlation between the nonlinear component and the surface state, atmospheric state, observation geometry and other influencing factors at the time of observation, and the mapping is used to fill in the missing values ​​of the nonlinear component. This method also deeply considers the impact of the spatiotemporal differences of the geostationary satellite aerosol optical depth product, and can be used to obtain high-temporal resolution and full-coverage aerosol optical depth by integrating geostationary satellite aerosol optical depth and reanalysis data.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

[0081] The above are only some embodiments of the present invention. For those skilled in the art, several modifications and improvements can be made without departing from the inventive concept of the present invention, which all fall within the scope of protection of the present invention.

Claims

1. A high temporal resolution spatially seamless aerosol optical depth filling method, characterized in that: The following steps are involved: Step 1: Train the regression model of the aerosol optical depth influencing factors and aerosol optical depth at low spatial resolution: in, represents the low spatial resolution aerosol optical depth obtained by regression model fitting, kernel low represents the impact factor of aerosol optical depth at low spatial resolution, and f() represents the downscaling mapping relationship of each input variable at low spatial resolution; Define the residual ΔAOD between the original aerosol optical depth at low resolution and the aerosol optical depth obtained by regression model fitting low for: Among them, AOD low represents the low-resolution raw aerosol optical depth; The regression model trained at low spatial resolution is applied to high spatial resolution to obtain the predicted high-resolution aerosol optical depth results. Among them, kernel high Indicates the aerosol optical depth factor with high spatial resolution; ΔAOD low The high spatial resolution residual ΔAOD is obtained by bilinear sampling to high spatial resolution. high ; According to the formula Get the AOD of a single downscaling model high ; Step 2, linear component estimation: The linear relationship between the reanalysis data aerosol optical depth and the satellite aerosol optical depth is fitted based on the least squares method: AOD LC,i =A i AOD merra2,i +B i ; Among them, AOD LC,i Represents the linear component of a certain land cover type, A i 、B i are the slope and intercept of the linear relationship fitted on a certain land cover type, AOD merra2,i represents the aerosol optical depth of the reanalysis data on a certain land cover type; i represents different land cover types; Step 3, nonlinear component estimation: The relationship between the nonlinear components and their influencing factors is mapped using a machine learning algorithm to obtain the mapping relationship at the non-missing value pixel: AOD NC-clr,j =RF j (Φ)+ε j , where Φ represents the influence factor of the nonlinear component, j represents different places, and ε j Represents the residual at local time j, RF j () represents the mapping relationship between nonlinear components and their influencing factors; According to the formula AOD NC-gap-filled,j =RF j (Φ)+ε j Calculate the predicted value of the nonlinear component of aerosol optical depth (AOD) at the missing pixel at local time j NC-gap-filled,j ; Step 4: Superimpose the estimated linear and nonlinear components to obtain the spatially seamless aerosol optical depth estimate: AOD NC-gap-filled,j =A i ·AOD merra2,i +B i +RF j (F)+e j Among them, AOD NC-gap-filled,j It represents the aerosol optical depth after filling, thereby obtaining the high temporal resolution and spatially seamless aerosol optical depth after filling.

2. The method according to claim 1, wherein Low spatial resolution aerosol optical depth factor kernel low and high spatial resolution aerosol optical depth factor kernel high The same factors affecting AOD are included: vegetation index, land cover type, elevation, slope, wind speed, temperature, humidity and air pressure.

3. The method according to claim 1 or 2, wherein: Before training the regression model at low spatial resolution, each aerosol optical depth influencing factor was upsampled to the same spatial resolution as the aerosol optical depth of the reanalysis data.

4. The method according to claim 3, wherein The aerosol optical depth factor is upsampled using an area-weighted approach: Among them, S sum Represents the low spatial resolution pixel area, S i represents the area of ​​the i-th high spatial resolution pixel within the low spatial resolution pixel, x i It represents the value of the aerosol optical depth impact factor corresponding to the i-th high spatial resolution pixel within the low spatial resolution pixel. Represents x i The value of the aerosol optical depth factor after upsampling, where n represents the number of high spatial resolution pixels.

5. The method according to claim 1 or 2, wherein: The influencing factors of the nonlinear component Φ include: temperature T 2m , water vapor TQV, horizontal wind speed U 2m , vertical wind speed V 2m , planetary boundary layer height PBLH, air pressure PS, elevation DEM, land cover type LC, normalized vegetation index NDVI, sensor azimuth SAA, sensor zenith angle SZA, annual cumulative day DOY, latitude Lat, longitude Lon, geographical distance D from the pixel to the center of the study area center and population density pop.

6. The method according to claim 1 or 2, wherein: Geographic distance D center The calculation formula is: Among them, γ and φ represent the longitude and latitude of the pixel, respectively. The subscripts of γ and φ are used to distinguish the two pixels or geographical locations to be calculated, and r represents the radius of the earth.

Images