Monthly satellite and ground precipitation fusion method based on random forest and land surface environmental variables

Abstract

This invention discloses a monthly-scale satellite-to-ground precipitation fusion method based on random forest and land surface environmental variables, comprising the following steps: First, temporal matching is performed between ground station precipitation data and satellite precipitation grid data; second, the locations of ground rain gauge stations are identified and spatially matched with the satellite precipitation grid; third, on grids with rain gauge stations, a random forest model is established with ground precipitation data as the dependent variable and satellite precipitation data and relevant land surface environmental variables as independent variables; finally, on grids without rain gauge stations, satellite precipitation data and relevant land surface environmental variables are used to calculate the fused precipitation data through the established random forest model. This invention combines precipitation-related land surface environmental variables to fuse ground precipitation and satellite precipitation, obtaining high-precision fused precipitation data, which can provide a more accurate precipitation input source for hydrological models, etc.

Description

Technical Field

[0001] This invention relates to a multi-source precipitation data fusion method, and more particularly to a monthly-scale terrestrial precipitation fusion method based on random forest and land surface environmental variables that can reduce errors. Background Technology

[0002] Precipitation, a common natural phenomenon, plays a vital role in the global water cycle, ecosystems, climate system, and energy balance. The spatiotemporal distribution of precipitation directly or indirectly affects terrestrial hydrological processes such as surface runoff, groundwater dynamics, soil moisture, and evapotranspiration, making it an important parameter in hydrology, meteorology, and ecology. Simultaneously, precipitation allows atmospheric moisture to return to the ground, providing freshwater resources essential for life on Earth. Therefore, reliable and accurate precipitation measurement is not only crucial for understanding climate change but also indispensable for water resource management, hydrological forecasting, ecological environment monitoring, and drought and flood early warning. Due to the influence of topography, geographical location, and underlying surface factors, precipitation exhibits strong spatiotemporal variability and is easily affected by natural and anthropogenic interference, making high-precision precipitation measurement relatively difficult.

[0003] With the development of sensors and inversion algorithms, satellite remote sensing has become a common method for measuring precipitation data, providing spatially continuous and temporally complete precipitation data on a global scale. Therefore, satellite precipitation products have brought unprecedented opportunities to hydrological and climatological research on a global scale. However, due to limitations imposed by topographic factors, inconsistencies between electromagnetic signals and cloud cover, and the precipitation inversion algorithms themselves, satellite precipitation estimation suffers from significant systematic and random errors at both global and regional scales, requiring appropriate corrections at local scales. Satellite precipitation products are spatially continuous but relatively low-precision "area" data, while ground-based precipitation data are high-precision, discontinuous "point" data. Therefore, ground-based precipitation data can be fused with satellite precipitation products to obtain high-precision satellite-ground fused precipitation data. Currently, multiple linear regression models are commonly used to establish the relationship between ground precipitation and environmental variables and satellite precipitation, thereby achieving the fusion of satellite and ground precipitation data. However, this method may suffer from multicollinearity of independent variables, leading to significant model errors. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a precipitation fusion method that can avoid errors caused by multicollinearity of independent variables.

[0005] Technical solution: The monthly-scale terrestrial precipitation fusion method based on random forest and land surface environmental variables of the present invention includes the following steps:

[0006] (1) Time-based matching of precipitation data from ground stations with satellite precipitation grid data;

[0007] (2) Identify the locations of ground rainfall stations and spatially match them with the satellite precipitation grid;

[0008] (3) On the grid with rainfall stations, a random forest model is established with ground precipitation data as the dependent variable and satellite precipitation data and related land surface environmental variables as independent variables;

[0009] (4) On grids without rain gauge stations, satellite precipitation data and relevant land surface environmental variables are used to calculate the fused precipitation data through a well-established random forest model.

[0010] In step (1), firstly, unqualified data is deleted. Unqualified data refers to files in the original ground data files where all precipitation records are -999. Then, the precipitation data measured by satellite using Coordinated Universal Time is matched with the precipitation records of the corresponding time zone observed by the ground station.

[0011] In step (2), the latitude and longitude of the rainfall station are first determined, and then the grid with the center of the satellite precipitation grid data closest to the rainfall station is matched with the rainfall station.

[0012] Step (3) includes the following steps:

[0013] (31) The precipitation data from ground stations and the satellite precipitation data are accumulated in the same month to obtain monthly-scale precipitation data from ground stations and satellites;

[0014] (32) On the grid with stations, a random forest model is established for each month, with ground station precipitation data as the dependent variable and satellite precipitation data and various land surface environmental variables as independent variables.

[0015] In step (32), the land surface environmental variables are remote sensing data on a monthly scale.

[0016] The land surface environmental variables include NDVI, LST-day, LST-night, DEM, Latitude, and Longitude data.

[0017] In step (4), satellite precipitation data and NDVI, LST-day, LST-night, DEM, Latitude, and Longitude are used as independent variables, and the fused precipitation data is calculated using the random forest model established in step (3).

[0018] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: it can obtain the optimal precipitation data of the study area, the obtained precipitation data has high accuracy and the error is smaller than that of conventional precipitation fusion methods, and the obtained satellite-to-ground precipitation data has higher consistency with ground reference data; it is also easy to implement, requires less computation, and can provide a more accurate precipitation input source for hydrological models, etc. Attached Figure Description

[0019] Figure 1 This is a flowchart of the present invention;

[0020] Figure 2 This is a time series of the accuracy indicators of satellite-ground fusion precipitation in Embodiment 1 of the present invention. Detailed Implementation

[0021] See Figure 1 Embodiment 1 of the present invention includes the following steps:

[0022] Step 1: Perform temporal matching between ground station precipitation data and satellite precipitation grid data;

[0023] Step 2: Identify the locations of ground-based rain gauge stations and spatially match them with the satellite precipitation grid.

[0024] Step 3: On the grid with rainfall stations, establish a random forest model with surface precipitation data as the dependent variable and satellite precipitation data and related land surface environmental variables as independent variables.

[0025] Step 4: On grids without rain gauges, satellite precipitation data and relevant land surface environmental variables are used to calculate fused precipitation data using a pre-established random forest model.

[0026] Example: Daily precipitation data selected from satellites were TMPA daily data from 2011-2015, with a spatial resolution of 0.25°. Daily precipitation observation data from meteorological stations were obtained from the China Meteorological Administration, covering more than 2,000 meteorological stations in the study area. Environmental variables included: Digital Elevation Model (DEM), Normalized Difference Vegetation Index (NDVI), Land Surface Temperature (LST, including daytime and nighttime surface temperatures), and latitude and longitude. Specifically, DEM data came from SRTM, with a spatial resolution of 90 meters; NDVI data came from NASA's MOD13C2 product, a monthly composite NDVI index with a spatial resolution of 0.05°; and LST data came from NASA's MOD11C3 product, a monthly composite LST with a spatial resolution of 0.05°.

[0027] Since this invention uses all rainfall stations for fusion, and lacks additional ground station data or other forms of validation data to evaluate the applicability and effectiveness of its fusion method, the k-fold cross-validation method, widely used for evaluating model accuracy, is employed to verify the fusion accuracy of this invention. Cross-validation significantly reduces the randomness of a single random partition by dividing the sample data multiple times. Through multiple partitions and training, the model can traverse the sample data, thereby improving its generalization ability. Here, a 10-fold cross-validation method is used, where all ground rainfall stations are randomly divided into 10 parts. Each time, 9 parts of the ground stations are used in the fusion with satellite data, and 1 part of the ground station data is used to evaluate the fusion effect. This process is repeated ten times until every part of the ground stations is used to evaluate the fusion effect.

[0028] (1) Perform time matching between ground station precipitation data and satellite precipitation grid data; download daily-scale TMPA data and ground rain gauge data. First, perform quality control on the ground station precipitation data, deleting unqualified data. Next, since the satellite precipitation data TMPA uses Coordinated Universal Time (UTC), while the ground stations use Beijing Time, the time systems of the daily data are inconsistent, thus requiring time matching. The TMPA data used here is one day later than the ground station data, so it is necessary to match the satellite TMPA data of the current day with the ground station data of the next day. For example, the TMPA data of January 1st needs to be matched with the ground data of January 2nd.

[0029] (2) Since the vector "point" precipitation data and the raster satellite precipitation grid data are not in the same format, it is necessary to spatially match the stations with the grid. After determining the latitude and longitude of the precipitation stations, we adopted a simple "grid-station" matching method, that is, the grid whose center is closest to the precipitation station is matched with the precipitation station.

[0030] (3) On the grid with rainfall stations, a random forest model is established with ground precipitation data as the dependent variable and satellite precipitation data and related land surface environmental variables as independent variables.

[0031] (31) Since there is a certain delay in the response of land surface environmental variables to precipitation, especially for NDVI, we establish the relationship between land surface environmental variables and precipitation on a monthly scale. The land surface environmental variables are directly obtained from relevant remote sensing products on a monthly scale. The precipitation data from ground stations and satellite precipitation data are accumulated in the same month to obtain the monthly-scale precipitation data from ground stations and satellites;

[0032] (32) On the grid with stations, a random forest (RF) model was established using ground station precipitation data as the dependent variable and satellite precipitation data along with six land surface environmental variables (NDVI, LST-day, LST-night, DEM, Latitude, and Longitude) as independent variables. Since the study period was 2011-2015, a random forest model needed to be established for each month.

[0033] (4) On grids without rainfall stations, using satellite precipitation data and NDVI, LST-day, LST-night, DEM, Latitude, and Longitude as independent variables, the merged precipitation data is calculated using the random forest model established in step three. The specific steps are as follows:

[0034] (41) Random forest modeling was performed in the R language integrated development environment Rstudio;

[0035] (42) Random forest modeling was performed using the randomForest package in R language in conjunction with the caretb package. randomForest contains two important model parameters: the number of features (number of dependent variables) used in each decision regression tree, mtry, and the number of trees in the random forest model, ntree.

[0036] (43) First, it is necessary to determine the parameter `mtry` in the random forest modeling. The parameter optimization function of the `caret` package is used to optimize the parameters of `mtry` in the random forest. In the `caret` package, the `trainControl` is first used to set the resampling strategy for the training samples. The resampling strategy used in this invention is 10-fold cross-validation. Next, the `train` function of the `caret` package is used to optimize the parameters of `mtry` in the random forest. That is, using station data as the dependent variable, satellite precipitation data and land surface environmental variables as independent variables, and minimizing the root mean square error (RMSE) as the target metric, random forest modeling is performed. Finally, the `mtry` that minimizes the RMSE can be obtained, and then this `mtry` is used to determine the number of smallest numbers `ntree` in the regression process. At this point, the parameter optimization of the random forest model is complete.

[0037] (44) Random forest modeling is performed using the randomFores function of the randomForest package. Ground station data is used as the dependent variable, satellite precipitation data and land surface environmental variables are used as independent variables, and the mtry and ntree obtained in step 3 are used as the optimal random forest parameters. These parameters are then passed to the randomForest function, and the random forest model is finally established.

[0038] (45) Prediction is performed using the predict function of the randomForest package. That is, satellite precipitation data and land surface environmental variables at locations without station grids are used as independent variables, and the merged precipitation data at locations without station grids is predicted using the random forest established in step 4. Finally, the merged precipitation data at locations without station grids is obtained.

[0039] In this way, on grids with stations, ground station data is used as the optimal precipitation data, while on grids without stations, fused precipitation data is used as the optimal precipitation data, thus obtaining the optimal precipitation data for this study area. Figure 2 Figure 2 presents the time series of accuracy indicators for satellite-to-ground fused precipitation data implemented in this invention, and compares them with a multiple linear regression model. Accuracy indicators include relative bias (RB), root mean square error (RMSE), and correlation coefficient (CC). As shown in Figure 2, the accuracy indicators of this invention are higher than those of the commonly used multiple linear regression model, mainly reflected in lower relative bias (RB) and root mean square error (RMSE), and higher correlation coefficient (CC). This indicates that after sampling the fusion method of this invention, there is a higher consistency between satellite-to-ground precipitation and the ground reference.

Claims

1. A method for fusing monthly-scale terrestrial precipitation based on random forest and land surface environmental variables, characterized in that, The method includes the following steps: (1) Time-based matching of precipitation data from ground stations with satellite precipitation grid data; (2) Identify the locations of ground rainfall stations and spatially match them with the satellite precipitation grid; (3) On a grid with rainfall stations, a random forest model is established with surface precipitation data as the dependent variable and satellite precipitation data and relevant land surface environmental variables as independent variables; the land surface environmental variables include NDVI, LST-day, LST-night, DEM, Latitude, and Longitude data; including the following steps: (31) The precipitation data from ground stations and the satellite precipitation data are accumulated in the same month to obtain monthly-scale precipitation data from ground stations and satellites; (32) On the grid with stations, a random forest model is established for each month, with ground station precipitation data as the dependent variable and satellite precipitation data and various land surface environmental variables as independent variables. (4) On grids without rain gauge stations, satellite precipitation data and relevant land surface environmental variables are used to calculate the fused precipitation data through a well-established random forest model.

2. The method for integrating monthly-scale terrestrial precipitation based on random forest and land surface environmental variables according to claim 1, characterized in that, In step (1), unqualified data is first deleted, and the precipitation data measured by satellite using Coordinated Universal Time is matched with the precipitation records of the corresponding time zone observed by ground stations.

3. The method for integrating monthly-scale terrestrial precipitation based on random forest and land surface environmental variables according to claim 1, characterized in that, In step (2), the latitude and longitude of the rainfall station are first determined, and then the grid with the center of the satellite precipitation grid data closest to the rainfall station is matched with the rainfall station.

4. The method for integrating monthly-scale terrestrial precipitation based on random forest and land surface environmental variables according to claim 1, characterized in that, In step (32), the land surface environmental variables are remote sensing data on a monthly scale.

5. The method for integrating monthly-scale terrestrial precipitation based on random forest and land surface environmental variables according to claim 1, characterized in that, In step (4), satellite precipitation data and NDVI, LST-day, LST-night, DEM, Latitude, and Longitude are used as independent variables, and the fused precipitation data is calculated using the random forest model established in step (3).

Images