Multi-source variable random forest VOD downscaling method considering vegetation optical thickness memory

Abstract

The invention provides a novel method for improving the spatial resolution of the vegetation optical thickness (VOD) by using a random forest method. According to the method, dynamic and static variables related to VOD, such as optical vegetation indexes, terrain parameters, soil humidity and rainfall, are comprehensively considered, and the memory effect (lag VOD) of vegetation optical thickness is introduced. By means of the method, researchers can downscale 0.1 degree VOD data based on the multi-frequency radiometer to the VOD data with the resolution of 0.01 degree. The research shows that the accuracy of the model is remarkably improved by adding the lag VOD variable. Besides, a normalized vegetation index (NDVI), an enhanced vegetation index (EVI) and a leaf area index (LAI) are used for evaluating the downscaled 0.01-degree VOD, a result shows that the downscaled 0.01-degree VOD is highly consistent with the spatial-temporal pattern of the optical vegetation indexes, and spatial-temporal dynamic characteristics of vegetation can be better captured.

Description

Technical Field

[0001] The present invention relates to the fields of remote sensing information technology and geographic information processing technology, in particular to a method and an implementation system thereof that combines the vegetation optical thickness memory (VOD memory) characteristic and utilizes multi-source variables and a random forest (RF) algorithm to perform spatial downscaling of VOD data. Background Art

[0002] Vegetation Optical Depth (VOD) describes the attenuation of microwave (passive or active) radiation in the vegetation layer and is related to vegetation water content (VWC) and canopy structure. Quantifying this attenuation effect is very important for accurately retrieving surface soil moisture from passive microwave satellite observations and analyzing the fields of global climate and environmental change. It has been widely used in global-scale carbon dynamics, vegetation phenology, global isotropic changes and drought detection, burn area trends and fire risks. As mentioned in the above study, VOD products mainly come from passive microwave sensors of different frequencies. These datasets include high-frequency (C / X / Ku band) VOD in AMSR-E (Advanced Microwave Scanning Radiometer) and its successor AMSR2 (Advanced Microwave Scanning Radiometer 2) and L-band low-frequency VOD (L-VOD) in Soil Moisture and Ocean Salinity (SMOS) and Soil Moisture Active Passive (SMAP).

[0003] However, most of the above VOD products are based on radiometers with low spatial resolution (0.25°-0.1°), which hinders the application of vegetation dynamic monitoring at regional scale (such as 1km). In addition, high-resolution VOD is also important for improving high-resolution soil moisture inversion as a necessary radiation transfer variable for microwave soil moisture inversion. Since the spatial resolution of these developed VOD products is relatively coarse, in order to meet the resolution requirements in practical applications, obtaining high-resolution VOD has become increasingly necessary in many research fields. Domestic and foreign researchers have proposed statistical and physical methods to obtain high-resolution VOD, such as the sharpened microwave brightness temperature inversion based on the smoothing filter modulation (SFIM) technology to obtain resolution-enhanced VOD, and the high-resolution VOD of a grassland in the Heihe River Basin obtained by inverting the backscatter data of Sentinel-1C band based on the water cloud model; these methods use the idea of ​​establishing statistical correlation or physically based models between coarse-scale VOD and fine-scale auxiliary variables, and require setting prior conditions, such as linear fitting assumptions, which makes it difficult to fully describe the complex relationship between VOD and auxiliary variables;

[0004] Inspired by the strong ability to model the nonlinear relationship between auxiliary variables and predictor variables, domestic and foreign scholars have introduced machine learning methods and applied them to the downscaling research of different data products, rather than just limiting them to a certain type of data; for example, the mass conservation correction is introduced into the RF model to downscale soil moisture (SM) to obtain high-resolution SM data; the 36km and 9km resolution SM data in SMAP are used as training for the neural network (NN) algorithm, and based on the training relationship between the two, 2.25km resolution SM data are obtained; in addition to being applied to the downscaling of SM data, the spatial correlation information between precipitation is considered and the precipitation product is downscaled based on the RF model to generate high-precision and high-resolution precipitation data; in the above-mentioned downscaling research based on machine learning methods, the nonlinear relationship between the independent variable and the dependent variable at high resolution is fully considered, and the downscaling results show more spatial heterogeneity while retaining the spatial pattern of the original large-scale data; however, there are few studies on the downscaling of VOD obtained from the passive microwave radiation transfer equation through microwave measurements using machine learning methods. Summary of the invention

[0005] In order to overcome the shortcomings of the prior art, the present invention aims to provide a method for improving the spatial resolution of vegetation optical depth based on the random forest method, and constructs a random forest downscaling model (RFDM-VOD) which, in addition to optical vegetation index, surface parameters and environmental factors, additionally considers lagged VOD variables, so as to provide more fine-scale information and vegetation memory information about vegetation optical depth for the vegetation optical depth downscaling method, and at the same time make up for the lack of space and time of available multispectral data in the study area, and fully explore the nonlinear relationship between vegetation optical depth and multi-feature data to obtain vegetation optical depth with high spatial resolution.

[0006] In order to achieve the above object, the present invention provides a multi-source data random forest downscaling method considering vegetation optical thickness memory, which specifically includes the following steps:

[0007] (1) Obtain vegetation optical thickness data, SMAP soil moisture data, MODIS data, and SRTM digital elevation data in the target area. MODIS data includes surface temperature data, vegetation index data, terrain data, and WorldClim precipitation data;

[0008] (2) Design the vegetation optical thickness memory data as one of the auxiliary factors for vegetation optical thickness downscaling, and select the vegetation optical thickness with a lag of seven days as the auxiliary data;

[0009] (3) Perform data quality control on the above auxiliary data products; perform projection conversion, splicing, resampling and clipping on MODIS data, soil moisture data and precipitation data, and perform projection conversion, resampling and clipping on SRTM digital elevation data, obtain MODIS data, soil moisture data, precipitation data and SRTM digital elevation data of the target area with the same spatial resolution as the vegetation optical depth data, and then perform spatiotemporal matching with the vegetation optical depth data to serve as the input data of the vegetation optical depth downscaling model;

[0010] (4) A vegetation optical thickness downscaling model was constructed based on random forests, and the vegetation optical thickness downscaling model was trained using lagged vegetation optical thickness data, soil moisture data, processed MODIS data, precipitation data, and SRTM digital elevation data;

[0011] (5) obtain the feature importance ranking of the above-mentioned VOD spatial downscaling model considering vegetation optical thickness memory (i.e., lagged VOD) (RFDM-VOD);

[0012] (6) The lagged vegetation optical depth data, soil moisture data, MODIS data and SRTM digital elevation data were resampled to 0.25° / 0.1° spatial resolution and input into the vegetation optical depth downscaling model for hierarchical downscaling, and finally the vegetation optical depth data with a spatial resolution of 0.01° in the target area were obtained.

[0013] (7) Application of the VOD spatial downscaling model considering vegetation optical depth memory (RFDM-VOD) and results evaluation.

[0014] Furthermore, in step (1), the vegetation index product includes normalized difference vegetation index product, enhanced vegetation index product and leaf area index data; in addition to vegetation index data, MODIS data also includes surface temperature data and land cover category of the region; SRTM digital elevation data includes elevation, aspect and slope data;

[0015] Furthermore, step (3) involves relevant data preprocessing: all MODIS data are resampled using ArcGIS, ENVI and Matlab software, and bilinear interpolation is selected as the resampling method to reduce the impact of geographic reference errors. The resampled MODIS data are geo-corrected to obtain the same geographic alignment as the vegetation optical thickness data, and then all data are cropped to ensure that the images have exactly the same coverage area. All data are preprocessed to obtain their pixel values. The preprocessed auxiliary data and the vegetation optical thickness data have the same projection, spatial resolution and consistent coverage;

[0016] Furthermore, in step (4), the general expression of the vegetation optical thickness downscaling model is:

[0017]

[0018] In the above formula, ξ is the residual, are dynamic variables, corresponding to NDVI, LST mean 、PRE、EVI、SM、7-day delayed VOD(VOD lag-7 ); are non-dynamic variables, corresponding to DEM, slope (SLOPE), aspect (ASPECT), longitude (LON), and latitude (LAT); f(·) is the modeling function relationship between the dependent variable and the explanatory variable.

[0019] Furthermore, in step (5), the features in the feature importance ranking of the downscaling (RFDM-VOD) model include NDVI, LST mean 、PRE、EVI、SM、7-day delayed VOD(VOD lag-7 ), DEM, slope (SLOPE), aspect (ASPECT), longitude (LON), latitude (LAT);

[0020] Furthermore, in step (6), obtaining vegetation optical thickness data with a spatial resolution of 0.01° in the target area requires hierarchical downscaling, including downscaling of vegetation optical thickness data from 0.25° to 0.1° and from 0.1° to 0.01°;

[0021] Furthermore, in step (7), the evaluation includes the spatiotemporal correlation analysis between the downscaled vegetation optical thickness and the vegetation index. In addition, the indicators for model accuracy verification include the root mean square error RMSE, the positive Pearson correlation coefficient R and the mean absolute error MAE; their formulas are as follows:

[0022]

[0023]

[0024] In the above formula, y si is the predicted value of the i-th VOD; y oi is the original value of the i-th VOD; is the arithmetic mean of the i-th VOD prediction value; is the arithmetic mean of the original value of the i-th VOD.

[0025] The present invention comprehensively considers the optical vegetation index, terrain parameters, soil moisture, precipitation and other parameters related to VOD and the nonlinear relationship between them, constructs a random forest-based multi-source variable VOD spatial downscaling method (RFDM-VOD) that considers vegetation optical thickness memory, and uses this model to apply to the low spatial resolution VOD data of AMSR-E / 2 to obtain the downscaled high spatial resolution VOD data. The high spatial resolution VOD after RFDM-VOD downscaling shows more spatial heterogeneity that the original large-scale VOD data does not have, and shows a more refined spatial pattern; the temporal dynamics of the high spatial resolution VOD after downscaling are basically consistent with the vegetation index NDVI, LAI and EVI, and can more clearly capture the dynamic changes of vegetation on the annual, seasonal and monthly scales. Overall, the downscaling method of the present invention performs well in the downscaling results of the VOD products of AMSR-E / 2, and the development of high-resolution VOD information is of great significance for monitoring vegetation dynamics and evaluating the health of ecosystems. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is the geographical location and elevation of the study area in the example of the present invention;

[0027] Figure 2 This is a flow chart of vegetation optical thickness downscaling in the present invention;

[0028] Figure 3 It is a scatter plot of the estimation and target data of the training and test data sets under the RFDM-VOD downscaling model of the present invention;

[0029] Figure 4 Comparison diagram of vegetation optical thickness before and after downscaling in the study area provided by the example of the present invention;

[0030] Figure 5 A bar chart showing the relative importance of the prediction factors in the RFDM-VOD downscaling model of the present invention;

[0031] Figure 6 It is a time series diagram of high spatial resolution VOD and vegetation index after downscaling of different vegetation cover types in the present invention;

[0032] Figure 7 This is the spatial distribution diagram of the determination coefficient of high spatial resolution VOD and vegetation index after downscaling in the present invention. DETAILED DESCRIPTION

[0033] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not intended to limit the scope of use of the present invention.

[0034] The present invention proposes a multi-source data random forest downscaling method that considers vegetation optical thickness memory. Southwest China is taken as the study area. The geographical location is between 20°54′N—34°19′N, 97°21′E—112°04′E. The area has a subtropical monsoon climate. The average annual temperature in the study area is between 0 and 23°C, and the annual precipitation is between 600 and 2400 mm, forming a distinct dry and wet season change. The wet season precipitation accounts for more than 85% of the annual precipitation, and the dry season precipitation is less and the evaporation is large. The overall terrain in the study area is high in the west and low in the east, with an elevation between -1 and 7156m. The topography is complex and it is a typical karst landform area. High-resolution VOD products have important application value for ecological and environmental management in the southwest region. The spatial distribution of elevation in the study area can be seen in Figure 1 .

[0035] The multi-source data random forest downscaling method considering vegetation optical thickness memory provided by the present invention specifically comprises the following steps:

[0036] (1) Obtain vegetation optical thickness data, soil moisture data, MODIS data, and SRTM digital elevation data in the target area. MODIS data includes surface temperature data, vegetation index data, terrain data, and WorldClim precipitation data;

[0037] Among them, the vegetation optical thickness data and soil moisture data are based on the microwave measurement data and cross-calibrated brightness temperature data of the multi-frequency radiometers AMSR-E and AMSR2, and the physical relationship between multiple channels is comprehensively considered. The vegetation optical thickness products and soil moisture products are obtained using the multi-channel collaborative inversion algorithm (MCCA);

[0038] The NDVI and EVI data in MODIS data come from MOD13A3 products, with a spatial resolution of 1 km and a temporal resolution of 16 days. The LAI data comes from MOD15A2H products, with a spatial resolution of 500 meters and a temporal resolution of 8 days. The mean of daytime and nighttime surface temperatures is used as auxiliary data in the downscaling model. The data set comes from MOD11A2, with a spatial resolution of 1 km and a temporal resolution of 8 days. The land cover level of each pixel is determined using the annual MCD12Q1 product of MODIS version 6.1. In the example of the present invention, the land cover classification scheme of the International Geosphere-Biosphere Program (IGBP) is selected to determine different land cover types.

[0039] (2) Perform data quality control on the above auxiliary data products; perform projection conversion, splicing, resampling and clipping on MODIS data, soil moisture data and precipitation data, and perform projection conversion, resampling and clipping on SRTM digital elevation data, obtain MODIS data, soil moisture data, precipitation data and SRTM digital elevation data of the target area with the same spatial resolution as the vegetation optical depth data, and then perform spatiotemporal matching with the vegetation optical depth data to serve as the input data of the vegetation optical depth downscaling model;

[0040] The relevant data preprocessing is involved: all MODIS data are resampled using ArcGIS, ENVI and Matlab software, and bilinear interpolation is selected as the resampling method to reduce the impact of geographic reference errors. The resampled MODIS data are geo-corrected to obtain the same geographic registration as the vegetation optical thickness data, and then all data are cropped to ensure that the images have exactly the same coverage area. All data are preprocessed to obtain their pixel values. The preprocessed auxiliary data and the vegetation optical thickness data have the same projection, spatial resolution and consistent coverage.

[0041] (3) A vegetation optical thickness downscaling model was constructed based on random forests, and the vegetation optical thickness downscaling model was trained using lagged vegetation optical thickness data, soil moisture data, processed MODIS data, precipitation data, and SRTM digital elevation data. The training and test sets of the downscaling model are as follows: Figure 3 As shown;

[0042] The general expression of the vegetation optical thickness downscaling model is:

[0043]

[0044] In the above formula, ξ is the residual, are dynamic variables, corresponding to NDVI, LST mean 、PRE、EVI、SM、7-day delayed VOD(VOD lag-7 ); are non-dynamic variables, corresponding to DEM, slope (SLOPE), aspect (ASPECT), longitude (LON), and latitude (LAT); f(·) is the modeling function relationship between the dependent variable and the explanatory variable.

[0045] The example downscaling process is as follows Figure 2 As shown, the specific steps are:

[0046] (3.1) First, the data is preprocessed. The downscaling model construction process needs to maintain consistency in temporal and spatial resolution. In this study, the temporal resolution is set to the monthly scale, where SM, VOD and data are synthesized from daily means to monthly data, NDVI and EVI are synthesized from 16-day means to monthly data, and LAI and data are synthesized from 8-day means to monthly data. Since the original VOD data has a spatial resolution of 0.25° and 0.1°, the dynamic and non-dynamic variable data at different high resolutions are upscaled to 0.25°, 0.1° and 0.01° spatial resolutions in the same coordinate system using the pixel averaging method.

[0047] (3.2) The dynamic and non-dynamic variable data and VOD data with a spatial resolution of 0.25° at the monthly scale obtained in the first step are substituted into the expression, where the dynamic and non-dynamic variable data are input as independent variables and VOD is output as the dependent variable to construct a 0.25° downscaling model. The residual between the downscaled 0.25° resolution VOD data and the original 0.25° resolution VOD data in the study area is obtained, and the residual is interpolated to 0.1° resolution through bilinear interpolation;

[0048] (3.3) Based on the downscaling model constructed at 0.25°, the pre-processed 0.1° resolution dynamic and non-dynamic variable data and VOD data are input into the model to obtain the downscaled 0.1° VOD data, and the residual between it and the original 0.1° VOD is calculated. The residual of the previous 0.25° model is superimposed, and the superimposed residual is interpolated to 0.01° resolution through bilinear interpolation. Finally, based on the constructed 0.1° downscaling model, the 0.01° NDVI and LST are mean 、PRE、EVI、SM、VOD lag-7 , DEM, SLOPE, ASPECT, LON, and LAT are used as input variables to obtain the downscaled 0.01° VOD data, and the residuals of the first two are further superimposed to obtain the final 0.01° VOD downscaling results.

[0049] at the same time Figure 3 Scatter plots of the 0.1° resolution RFDM-VOD estimates versus the target 0.1° VOD for the training and test datasets are shown. Both datasets have similar R 2 , R, MAE, and RSME, which shows that the RFDM-VOD model can generalize beyond the training data. 2 The values ​​are all around 0.90, showing a high correlation between the target and the estimated values, and indicating that the experimentally selected parameter settings are able to learn the relationship between coarse-scale and fine-scale vegetation optical depth.

[0050] (4) obtain the feature importance ranking of the above-mentioned VOD spatial downscaling model considering vegetation optical thickness memory (i.e., lagged VOD) (RFDM-VOD);

[0051] Among them, the features in the feature importance ranking of the downscaling (RFDM-VOD) model include NDVI, LST mean 、PRE、EVI、SM、7-day delayed VOD(VOD lag-7 ), DEM, slope (SLOPE), aspect (ASPECT), longitude (LON), latitude (LAT); Figure 5 The relative importance of prediction factors in the RFDM-VOD downscaling model is shown in the bar chart. The relative importance of prediction factors in the 0.1° and 0.01° vegetation optical thickness downscaling is shown in the figure. It can be seen that the relative importance of the lagged VOD variable is the highest, and it has a greater impact on the downscaling model.

[0052] (5) The lagged vegetation optical depth data, soil moisture data, MODIS data and SRTM digital elevation data were resampled to 0.25° / 0.1° spatial resolution and input into the vegetation optical depth downscaling model for hierarchical downscaling, and finally the vegetation optical depth data with a spatial resolution of 0.01° in the target area were obtained.

[0053] Among them, obtaining the vegetation optical thickness data with a spatial resolution of 0.01° in the target area requires graded downscaling, including downscaling the vegetation optical thickness data from 0.25° to 0.1° and from 0.1° to 0.01°; the downscaled vegetation optical thickness of 0.1° and 0.01° is similar to the original low spatial resolution vegetation optical thickness data. Figure 4 As shown. Figure 4 (c) and (d) show that the spatial distribution of the downscaled VOD is similar to that of the original 0.25° and 0.1° resolution VOD ( Figure 4 (a), (b)), the positions of the low-value center and the high-value center of the downscaled VOD match well with the VOD at the original resolution. The downscaled VOD shows more spatial details and is smoother than the VOD at the original resolution. Overall, the spatial pattern is consistent with expectations, showing a finer spatial pattern with increasing resolution, with a more accurate expression of the spatial heterogeneity of the VOD. These results show that the RFDM-VOD downscaling algorithm preserves the large-scale spatial pattern of the VOD.

[0054] (6) Application of the VOD spatial downscaling model considering vegetation optical depth memory (RFDM-VOD) and results evaluation.

[0055] The evaluation includes the temporal and spatial correlation analysis between the downscaled vegetation optical thickness and the vegetation index. In addition, the indicators for model accuracy verification include the root mean square error RMSE, the positive Pearson correlation coefficient R and the mean absolute error MAE; the formulas are as follows:

[0056]

[0057] In the above formula, y si is the predicted value of the i-th VOD; y oi is the original value of the i-th VOD; is the arithmetic mean of the i-th VOD prediction value; is the arithmetic mean of the original value of the i-th VOD.

[0058] In order to further understand the performance of high spatial resolution VOD after RFDM-VOD downscaling, five vegetation communities with more pixels in the study area, including Mixed Forest (MXF), Woody Savanna (WS), Savanna (SAV), Croplands (CRP) and Croplands / Natural Vegetation Mosaic, were selected based on the IGBP land cover classification. The pixel values ​​of NDVI, EVI and LAI under the five vegetation cover classifications were obtained. At the same time, the high spatial resolution VOD after RFDM-VOD downscaling was analyzed for time series consistency with the vegetation index data under the five vegetation cover classifications. Figure 6 As shown in the figure, the downscaled VOD under vegetation cover has obvious seasonal changes similar to NDVI, EVI and LAI. With the aging of vegetation, VOD gradually increases in spring, reaches a peak in summer, and then decreases in autumn. At the same time, it shows a more stable change trend compared with NDVI, EVI and LAI.

[0059] The determination coefficient values ​​were calculated based on the high spatial resolution VOD data obtained by the RFDM-VOD downscaling model and the ENVI, EVI and LAI values ​​obtained pixel by pixel in the study area. Based on this, the data quality of high spatial resolution VOD under RFDM-VOD and its ability to monitor the dynamic changes of vegetation over a long period of time were evaluated. The spatial pattern of the determination coefficient values ​​is shown in Figure 2. Figure 7 As shown, in most areas of the study area, the overall correlation between VOD and vegetation index after downscaling is good, and the downscaling results perform better.

[0060] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A multi-source variable random forest VOD downscaling method considering vegetation optical thickness memory, characterized in that: The following steps are involved: (1) Obtain vegetation optical thickness data, SMAP soil moisture data, MODIS data, and SRTM digital elevation data in the target area. MODIS data include surface temperature data, vegetation index data, terrain data, and WorldClim precipitation data. Vegetation index products include normalized difference vegetation index products, enhanced vegetation index products, and leaf area index data. In addition to vegetation index data, MODIS data also includes surface temperature data and land cover categories in the area. SRTM digital elevation data include elevation, aspect, and slope data. (2) Design the vegetation optical thickness memory data as one of the auxiliary factors for vegetation optical thickness downscaling, and select the vegetation optical thickness with a lag of seven days as the auxiliary data; (3) Data quality control of the above auxiliary data products; projection conversion, splicing, resampling and cropping of MODIS data, soil moisture data and precipitation data; projection conversion, resampling and cropping of SRTM digital elevation data; obtaining MODIS data, soil moisture data, precipitation data and SRTM digital elevation data of the target area with the same spatial resolution as the vegetation optical depth data; then performing spatiotemporal matching with the vegetation optical depth data as the input data of the vegetation optical depth downscaling model; the relevant data preprocessing is involved: all MODIS data are resampled using ArcGIS, ENVI and Matlab software; bilinear interpolation is selected as the resampling method to reduce the impact of geographic reference errors; the resampled MODIS data are geo-corrected to obtain the same geographic registration as the vegetation optical depth data; then all data are cropped to ensure that the images have exactly the same coverage area; all data are preprocessed to obtain their pixel values; the preprocessed auxiliary data and the vegetation optical depth data have the same projection, spatial resolution and consistent coverage; (4) A vegetation optical thickness downscaling model is constructed based on random forests, and the vegetation optical thickness downscaling model is trained using lagged vegetation optical thickness data, soil moisture data, processed MODIS data, precipitation data, and SRTM digital elevation data; wherein the general expression of the vegetation optical thickness downscaling model is: ； In the above formula, is the residual, are dynamic variables, corresponding to NDVI, 、PRE、EVI、SM、7-day delayed VOD( ); They are non-dynamic variables, corresponding to DEM, slope (SLOPE), aspect (ASPECT), longitude (LON), and latitude (LAT); Modeling functional relationships between dependent and explanatory variables; (5) Obtain the feature importance ranking of the VOD spatial downscaling (RFDM-VOD) model considering vegetation optical thickness memory (i.e., lagged VOD); the features in the feature importance ranking of the downscaling (RFDM-VOD) model include NDVI, 、PRE、EVI、SM、7-day delayed VOD( ), DEM, slope (SLOPE), aspect (ASPECT), longitude (LON), latitude (LAT); (6) The lagged vegetation optical thickness data, soil moisture data, MODIS data and SRTM digital elevation data were resampled to 0.25° / 0.1° spatial resolution respectively and input into the vegetation optical thickness downscaling model for hierarchical downscaling, and finally the vegetation optical thickness data with a spatial resolution of 0.01° in the target area were obtained; among them, obtaining the vegetation optical thickness data with a spatial resolution of 0.01° in the target area requires hierarchical downscaling, including downscaling from 0.25° to 0.1° and from 0.1° to 0.01°; (7) Application and evaluation of the VOD spatial downscaling model considering vegetation optical depth memory (RFDM-VOD); the evaluation includes the spatiotemporal correlation analysis between the downscaled vegetation optical depth and the vegetation index; in addition, the indicators for model accuracy verification include the root mean square error (RMSE), the positive Pearson correlation coefficient (R), and the mean absolute error (MAE); the formulas are as follows: ； ； ； In the above formula, is the predicted value of the i-th VOD; is the original value of the i-th VOD; is the arithmetic mean of the i-th VOD prediction value; is the arithmetic mean of the original value of the i-th VOD.

Images