A method for downscaling satellite precipitation based on actual measurement sites

Abstract

This invention discloses a method for downscaling satellite precipitation based on measured stations. It extracts watershed DEM data, including satellite precipitation, wind speed, and temperature. Weighted interpolation of precipitation data from measured stations yields grid data. A deep learning model learns the nonlinear, spatiotemporal coupling relationship between multiple factors and precipitation distribution, and predicts low-resolution satellite precipitation using dynamic weights. High-resolution satellite precipitation is predicted by upsampling satellite precipitation, temperature, wind speed, and station precipitation using a deep learning model and dynamic weights. The residual between the low-resolution precipitation and the measured station precipitation is calculated and upsampled. The high-resolution satellite precipitation is then added to the sampled residual to obtain the downscaled precipitation product. This invention downscales satellite precipitation products based on measured station data to obtain high-resolution precipitation products, achieving refined precipitation data and providing data support for inputting refined hydrological models.

Description

Technical Field

[0001] This invention relates to the fields of hydrology and meteorology, specifically to a method for downscaling satellite precipitation based on actual measurement stations. Background Technology

[0002] Precipitation is the core input variable of hydrological models, and the spatiotemporal resolution and accuracy of its data directly determine the reliability of simulation results and the timeliness of forecasts and warnings.

[0003] Currently, mainstream precipitation observation methods (ground stations, weather radar, and satellite remote sensing) all suffer from inherent systematic errors and uncertainties. While ground-based observation data is considered a benchmark, its spatial representativeness is severely insufficient. Sparse station distribution makes it difficult to capture the spatial heterogeneity of precipitation, especially in complex terrain or areas with few stations; discrete station observations cannot characterize the continuous spatial distribution of precipitation within a region. Furthermore, high station maintenance costs, instrument bias, and missing data records limit its widespread application. Radar quantitative precipitation estimation offers advantages in high spatiotemporal resolution and real-time performance, but its accuracy is susceptible to interference from various factors, including coverage blind spots caused by ground cover obstruction, beam attenuation, ground clutter, vertical profile representativeness errors, and radar parameter calibration deviations. These factors often lead to systematic overestimation or underestimation of precipitation by radar, and the limited coverage of radar observations makes it difficult to meet the needs of large-scale precipitation monitoring. Satellite quantitative precipitation estimation provides valuable wide-area coverage and quasi-continuous observation capabilities, becoming a core means of large-scale, cross-regional precipitation monitoring. However, its raw data suffers from a key deficiency of low spatiotemporal resolution, and its accuracy is generally lower than that of ground-based and radar observations. Errors mainly stem from limitations in remote sensing instrument sensitivity, insufficient sub-cloud / intra-cloud precipitation detection capabilities, difficulties in inverting complex surfaces, limitations in the inversion algorithms themselves, and the lack of ability to characterize small-scale precipitation processes at coarse resolution. Although satellite precipitation products (such as the GPM, IMERG, and FY series) are continuously upgraded, the spatial resolution of their raw data still falls short of the application requirements of high-precision distributed hydrological models, and the data often contains significant regional or systematic biases.

[0004] The technical limitations of the aforementioned observation methods make it difficult for raw satellite precipitation data to directly meet the stringent requirements of hydrological models (especially high-precision distributed models) for high spatiotemporal resolution and strong spatial heterogeneity representation of input data. Coarse-resolution satellite precipitation input without downscaling cannot accurately characterize the spatial distribution differences of precipitation within small-scale watersheds, significantly amplifying the uncertainties in hydrological simulation and forecasting, and even leading to false alarms or missed alarms.

[0005] Therefore, developing effective satellite precipitation downscaling techniques is crucial. The core objective is to leverage the wide-area coverage advantage of satellite precipitation products, constrained by high-precision observation data from a high-density ground station network, and employ advanced statistical / physical fusion methods to improve the spatial resolution of satellite precipitation data while correcting systematic biases and random errors. This achieves a refined spatial mapping of precipitation information, providing hydrological models with inputs that combine wide coverage, high resolution, and high precision, ultimately enhancing flood disaster simulation, forecasting, and early warning capabilities.

[0006] To better describe the spatial variation patterns of precipitation, fully leverage the advantages of wide coverage of satellite precipitation and high precision of ground station observations, and improve the spatial resolution and data quality of satellite precipitation products, it is urgent to adopt a technical solution that uses satellite quantitative precipitation estimation as the background field and ground station precipitation observation data as the true value constraint to perform downscaling fusion and residual refinement of satellite precipitation, thereby realizing the transformation of coarse-resolution satellite precipitation into high-resolution, high-precision precipitation grid data. Summary of the Invention

[0007] The purpose of this invention is to provide a method for downscaling satellite precipitation based on actual measurement stations. This method utilizes precipitation data from actual measurement stations to transform the station data from point to area, and downscales the precipitation data of satellite precipitation products. The resulting precipitation products provide data support for input into refined hydrological models.

[0008] To achieve the above functions, this invention designs a method for downscaling satellite precipitation based on actual measurement stations, performing the following steps S1-S9 to complete the error correction of satellite precipitation data:

[0009] Step S1: For the target area with actual measurement stations, extract the target watershed, the DEM raster data of the target watershed, and the mask file of the target watershed, and proceed to step S2;

[0010] Step S2: Obtain slope data based on DEM raster data, and classify the terrain type of the target area by combining the slope data with the geographical location, then proceed to step S3;

[0011] Step S3: Using the mask file of the target watershed, extract the satellite precipitation data, temperature and wind speed data of the target watershed, and proceed to step S4;

[0012] Step S4: The precipitation data, temperature, wind speed and normalized slope of the measured stations are interpolated to the spatial resolution of the satellite precipitation data through inverse distance weighted interpolation to obtain the measured station grid precipitation data with the same spatial resolution as the satellite precipitation data, and then proceed to step S5.

[0013] Step S5: Construct the LightSRCNN model and the LightGBM model respectively. Take the original precipitation data from the measured stations, the measured station grid precipitation data, temperature, wind speed and normalized slope as input, and the predicted low-resolution precipitation data of the grid points as output. Train the LightSRCNN model and the LightGBM model respectively to obtain the trained LightSRCNN model and the LightGBM model, and proceed to step S6.

[0014] Step S6: Based on the trained LightSRCNN model and LightGBM model, output low-resolution precipitation data respectively, calculate dynamic weights, obtain dynamically weighted low-resolution precipitation data, and proceed to step S7.

[0015] Step S7: Calculate the precipitation residual between the dynamically weighted low-rate precipitation data and the measured station grid precipitation data, and resample to 1km resolution to obtain high-resolution residuals, then proceed to step S8.

[0016] Step S8: Resample the original precipitation data from the measured stations, the measured station grid precipitation data, temperature, wind speed, and normalized slope to 1km. Use the 1km resolution data as input data for the trained LightSRCNN model and LightGBM model to predict and infer high-resolution precipitation data. Calculate the dynamic weights to obtain the dynamically weighted high-resolution precipitation data, and proceed to step S9.

[0017] Step S9: Overlay the high-resolution residuals onto the dynamically weighted high-resolution precipitation data to obtain precipitation data after residual scaling, and output the final precipitation data.

[0018] Beneficial effects: Compared with the prior art, the advantages of the present invention include:

[0019] This invention designs a method for downscaling satellite precipitation based on measured stations. Based on the watershed's DEM (Digital Elevation Model), a watershed mask file is obtained. Satellite precipitation, wind speed, and temperature data are extracted from the watershed mask file. Combining satellite precipitation, station precipitation, slope, wind speed, and temperature, the satellite precipitation data is downscaled. This method uses high spatiotemporal resolution precipitation product data as its base data. The data source is stable and reliable, and the training data process is smooth. Furthermore, it utilizes the relationship between station and satellite precipitation, slope, wind speed, and temperature to downscale the satellite precipitation, ensuring the objectivity and rationality of the results. High-resolution and high-precision precipitation products are obtained as input for a refined hydrological model. Attached Figure Description

[0020] Figure 1 This is a flowchart of a method for downscaling satellite precipitation based on actual measurement stations, provided by an embodiment of the present invention;

[0021] Figure 2 This is a schematic diagram comparing the distribution of evaluation indicators for raw satellite data and downscaled data provided in an embodiment of the present invention;

[0022] Figure 3 This is a schematic diagram comparing evaluation indicators of raw satellite data and downscaled data in different rainfall ranges according to embodiments of the present invention. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0024] This invention provides a method for downscaling satellite precipitation based on measured stations, referring to... Figure 1 Perform the following steps S1-S9 to complete the error correction of the satellite precipitation data:

[0025] Step S1: For the target area with actual measurement stations, extract the target watershed, the DEM (Digital Elevation Model) raster data of the target watershed, and the mask file of the target watershed, and proceed to step S2;

[0026] The specific steps of step S1 are as follows:

[0027] Step S1.1: Fill the depressions in the target area;

[0028] Step S1.2: Calculate the flow direction of the watershed in the target area;

[0029] Step S1.3: Calculate the runoff of the watershed;

[0030] Step S1.4: Determine the outlet station of the watershed;

[0031] Step S1.5: Extract the target watershed, as well as the DEM raster data and mask file of the target watershed.

[0032] In this embodiment, the target watershed is the Yellow River-Huai River-Hai River Basin. The raw DEM data of the target watershed is obtained from digital elevation data jointly provided by the United States Geological Survey (USUG) and the National Geomatics Center of China. The DEM and mask file of the target watershed are obtained.

[0033] Step S2: Obtain slope data based on DEM raster data, and classify the terrain type of the target area by combining the slope data with the geographical location, then proceed to step S3;

[0034] The specific steps of step S2 are as follows:

[0035] Step S2.1: Extract the slope factor from the DEM raster data of the target watershed and normalize the slope. The specific formula is as follows:

[0036] ;

[0037] In the formula, The normalized slope value. This is the original slope value. The maximum value of all valid original slope values. The minimum of all valid original slope values;

[0038] Step S2.2: Based on the slope factor and combined with the geographical location, classify the terrain type of the target area, including plains, mountainous areas and coastal areas.

[0039] Step S3: Using the mask file of the target watershed, extract the satellite precipitation data, temperature and wind speed data of the target watershed, and proceed to step S4;

[0040] The specific steps of step S3 are as follows:

[0041] Step S3.1: Download satellite precipitation data, temperature, and wind speed, and determine the coverage, spatial resolution, and temporal resolution of each data point;

[0042] Step S3.2: Based on the mask file of the target watershed obtained in step S1, and using it as the boundary of the target watershed, extract the satellite precipitation data, temperature, and wind speed of the target watershed.

[0043] Step S4: The precipitation data, temperature, wind speed and normalized slope of the measured stations are interpolated to the spatial resolution of the satellite precipitation data through inverse distance weighted interpolation to obtain the measured station grid precipitation data with the same spatial resolution as the satellite precipitation data, and then proceed to step S5.

[0044] Inverse distance weighting (IRW) is a distance-based spatial interpolation method. Its core idea is that the value of an unknown point is influenced by nearby known points, with the influence increasing the closer the points are. Raw, measured precipitation data from various stations is interpolated using IW to obtain gridded precipitation data with the same spatial resolution as the precipitation product.

[0045] The specific steps of step S4 are as follows:

[0046] Step S4.1: Assume These are the coordinates of the measured station i. Let j be the coordinates of grid point j. Calculate the distance matrix between the measured station coordinates and the grid point coordinates, as shown in the following formula:

[0047] ;

[0048] In the formula, This represents the distance between the measured station i and grid point j.

[0049] Step S4.2: Calculate the weight of the measured station based on the distance between the measured station and the grid point, as shown in the following formula:

[0050] ;

[0051] In the formula, The weight of the measured site i;

[0052] Step S4.3: Based on the weights of the measured sites, perform interpolation calculations, as shown in the following formula:

[0053] ;

[0054] In the formula, N is the number of sample points. The observed value is at station i. The interpolated data for grid point j is used to obtain the measured precipitation data for the grid station.

[0055] Step S4.4: For the normalized slope, temperature and wind speed interpolation, perform steps S4.1-S4.3 to obtain grid data consistent with satellite precipitation resolution.

[0056] Step S5: Construct the LightSRCNN model and the LightGBM model respectively. Take the original precipitation data from the measured stations, the measured station grid precipitation data, temperature, wind speed and normalized slope as input, and the predicted low-resolution precipitation data of the grid points as output. Train the LightSRCNN model and the LightGBM model respectively to obtain the trained LightSRCNN model and the LightGBM model, and proceed to step S6.

[0057] The specific steps of step S5 are as follows:

[0058] Step S5.1: Construct the training dataset. The input feature X is multi-source two-dimensional spatial grid data, including filtered original measured precipitation data, measured precipitation data from gridded stations, temperature data, wind speed data, and normalized slope data. The training target Y is the predicted low-resolution precipitation data, which is interpolated from the grid points. composition;

[0059] Step S5.2: Construct the LightSRCNN model, which consists of an input layer, three convolutional layers, and an output layer. The input layer receives 5-channel two-dimensional spatial grid data. The first convolutional layer has 32 5×5 convolutional kernels, the second convolutional layer has 16 3×3 convolutional kernels, and the first two convolutional layers use the ReLU activation function. The third convolutional layer has one 3×3 convolutional kernel. The output layer directly outputs low-resolution precipitation data for each grid in a single channel.

[0060] Step S5.3: Train the LightSRCNN model. The loss function during training consists of a weighted average mean squared error (MSE) loss and a gradient L1 loss. The gradient L1 loss weight is α=0.3, and the MSE loss weight is β=0.7. No early stopping strategy is set. The LightSRCNN model is trained for 15 epochs. 80% of the data is used as the training set, and 20% of the data is used as the validation set. The validation set R² score is used as the performance evaluation metric. The model with the best validation set R² score is saved as a pth file.

[0061] Step S5.4: Construct the LightGBM gradient boosting tree model. The input data is the same as the input of the LightSRCNN network. The task type is regression, and the output is the predicted low-resolution precipitation data. The evaluation metric is RMSE. Set 31 leaf nodes, a learning rate of 0.05, and a feature bag ratio / sampling bag ratio of 0.8 to prevent overfitting. Enable multi-threaded training and use a fixed random seed to ensure reproducibility.

[0062] Step S5.5: Using mean squared error (MSE) loss as the loss function, without designing an early stopping strategy, train the boosting tree for 100 rounds; use 80% of the data as the training set and 20% of the data as the validation set, filter out NaN values ​​in the validation set, calculate the R² score to evaluate the model performance, and finally save the trained model pkl file.

[0063] Step S6: Based on the trained LightSRCNN model and LightGBM model, output low-resolution precipitation data respectively, calculate dynamic weights, obtain dynamically weighted low-resolution precipitation data, and proceed to step S7.

[0064] The specific steps of step S6 are as follows:

[0065] Step S6.1: Perform inference using the trained LightSRCNN model and LightGBM model respectively, and output LightSRCNN low-resolution precipitation data and LightGBM low-resolution precipitation data respectively;

[0066] Step S6.2: Based on the divided region type, set the basic weights. Specifically, for plains, the basic weight of LightSRCNN low-resolution precipitation data is set to 0.5, and the basic weight of LightGBM low-resolution precipitation data is set to 0.5; for mountainous regions, the basic weight of LightSRCNN low-resolution precipitation data is set to 0.8, and the basic weight of LightGBM low-resolution precipitation data is set to 0.2; for coastal regions, the basic weight of LightSRCNN low-resolution precipitation data is set to 0.7, and the basic weight of LightGBM low-resolution precipitation data is set to 0.3.

[0067] Step S6.3: Divide the flood season into flood season and non-flood season, and further adjust the basic weights. If the time is June, July and August, it is the flood season. Adjust the basic weights of LightSRCNN low-resolution precipitation data and LightGBM low-resolution precipitation data in mountainous areas to 0.9 and 0.1 respectively, adjust the basic weights of coastal areas to 0.8 and 0.2 respectively, and adjust the basic weights of plain areas to 0.7 and 0.3 respectively.

[0068] Step S6.4: Based on the defined basic weights, weighted fusion is performed to obtain low-resolution precipitation data, as follows:

[0069] ;

[0070] In the formula, For in the grid Low-resolution precipitation data, For in the grid Weights of the LightSRCNN model For in the grid LightSRCNN low-resolution precipitation data, For in the grid Weights of the LightGBM model This is low-resolution precipitation data from LightGBM.

[0071] Step S7: Calculate the precipitation residual between the dynamically weighted low-rate precipitation data and the measured station grid precipitation data, and resample to 1km resolution to obtain high-resolution residuals, then proceed to step S8.

[0072] The specific steps of step S7 are as follows:

[0073] Step S7.1: Based on the measured station grid, read the low-resolution precipitation data corresponding to the measured station grid;

[0074] Step S7.2: Subtract the corresponding low-resolution precipitation data from the measured station grid precipitation data to obtain the precipitation residual;

[0075] Step S7.3: The precipitation residual is upsampled to 1km using the interpolation method in step S4 to obtain a high-resolution precipitation residual.

[0076] Step S8: Resample the original precipitation data from the measured stations, the measured station grid precipitation data, temperature, wind speed, and normalized slope to 1km. Use the 1km resolution data as input data for the trained LightSRCNN model and LightGBM model to predict and infer high-resolution precipitation data. Calculate the dynamic weights to obtain the dynamically weighted high-resolution precipitation data, and proceed to step S9.

[0077] Step S9: Overlay the high-resolution residuals onto the dynamically weighted high-resolution precipitation data to obtain precipitation data after residual scaling, and output the final precipitation data.

[0078] The method for calculating downscaled precipitation data is as follows:

[0079] ;

[0080] In the formula, the final precipitation is the downscaled precipitation data. For high-resolution precipitation data, This represents the high-resolution precipitation residual.

[0081] A diagram comparing the evaluation metrics of raw satellite data and downscaled data is provided. Figure 2 , Figure 2 In the figure, (a), (b), (c), (d), and (e) represent the BIAS, CC, RMSE, KGE, and CSI indices of precipitation products GPM, PERSIANN-CDR, CMORPH, GSMAP, GPCP, MSWEP, CHIRPS, EAR5, and GLDAS, respectively. The performance of each precipitation product after downscaling was analyzed by calculating BIAS, CC, RMSE, KGE, and CSI. It can be seen that after downscaling, all indicators of each precipitation product are improved, and the consistency with precipitation data from actual observation stations is higher, significantly improving precipitation forecasting. A comparison diagram of the evaluation intervals of raw satellite data and downscaled data in different rainfall ranges is provided. Figure 3 After downscaling, the precipitation products are closer to the actual surface rainfall at the measurement stations for different rainfall intensities, resulting in more accurate predictions.

[0082] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for downscaling satellite precipitation based on measured stations, characterized in that, Perform the following steps S1-S9 to complete the error correction of the satellite precipitation data: Step S1: For the target area with actual measurement stations, extract the target watershed, the DEM raster data of the target watershed, and the mask file of the target watershed, and proceed to step S2; Step S2: Obtain slope data based on DEM raster data, and classify the terrain type of the target area by combining the slope data with the geographical location, then proceed to step S3; Step S3: Using the mask file of the target watershed, extract the satellite precipitation data, temperature and wind speed data of the target watershed, and proceed to step S4; Step S4: The precipitation data, temperature, wind speed and normalized slope of the measured stations are interpolated to the spatial resolution of the satellite precipitation data through inverse distance weighted interpolation to obtain the measured station grid precipitation data with the same spatial resolution as the satellite precipitation data, and then proceed to step S5. Step S5: Construct the LightSRCNN model and the LightGBM model respectively. Take the original precipitation data from the measured stations, the measured station grid precipitation data, temperature, wind speed and normalized slope as input, and the predicted low-resolution precipitation data of the grid points as output. Train the LightSRCNN model and the LightGBM model respectively to obtain the trained LightSRCNN model and the LightGBM model, and proceed to step S6. Step S6: Based on the trained LightSRCNN model and LightGBM model, output low-resolution precipitation data respectively, calculate dynamic weights, obtain dynamically weighted low-resolution precipitation data, and proceed to step S7. Step S7: Calculate the precipitation residual between the dynamically weighted low-rate precipitation data and the measured station grid precipitation data, and resample to 1km resolution to obtain high-resolution residuals, then proceed to step S8. Step S8: Resample the original precipitation data from the measured stations, the measured station grid precipitation data, temperature, wind speed, and normalized slope to 1km. Use the 1km resolution data as input data for the trained LightSRCNN model and LightGBM model to predict and infer high-resolution precipitation data. Calculate the dynamic weights to obtain the dynamically weighted high-resolution precipitation data, and proceed to step S9. Step S9: Overlay the high-resolution residuals onto the dynamically weighted high-resolution precipitation data to obtain precipitation data after residual scaling, and output the final precipitation data.

2. The method for downscaling satellite precipitation based on actual measurement stations according to claim 1, characterized in that, The specific steps of step S1 are as follows: Step S1.1: Fill the depressions in the target area; Step S1.2: Calculate the flow direction of the watershed in the target area; Step S1.3: Calculate the runoff of the watershed; Step S1.4: Determine the outlet station of the watershed; Step S1.5: Extract the target watershed, as well as the DEM raster data and mask file of the target watershed.

3. The method for downscaling satellite precipitation based on actual measurement stations according to claim 1, characterized in that, The specific steps of step S2 are as follows: Step S2.1: Extract the slope factor from the DEM raster data of the target watershed and normalize the slope. The specific formula is as follows: ； In the formula, The normalized slope value. This is the original slope value. The maximum value of all valid original slope values. The minimum of all valid original slope values; Step S2.2: Based on the slope factor and combined with the geographical location, classify the terrain type of the target area, including plains, mountainous areas and coastal areas.

4. The method for downscaling satellite precipitation based on actual measurement stations according to claim 1, characterized in that, The specific steps of step S3 are as follows: Step S3.1: Download satellite precipitation data, temperature, and wind speed, and determine the coverage, spatial resolution, and temporal resolution of each data point; Step S3.2: Based on the mask file of the target watershed obtained in step S1, and using it as the boundary of the target watershed, extract the satellite precipitation data, temperature, and wind speed of the target watershed.

5. The method for downscaling satellite precipitation based on actual measurement stations according to claim 1, characterized in that, The specific steps of step S4 are as follows: Step S4.1: Assume These are the coordinates of the measured station i. Let j be the coordinates of grid point j. Calculate the distance matrix between the measured station coordinates and the grid point coordinates, as shown in the following formula: ； In the formula, This represents the distance between the measured station i and grid point j. Step S4.2: Calculate the weight of the measured station based on the distance between the measured station and the grid point, as shown in the following formula: ； In the formula, The weight of the measured site i; Step S4.3: Based on the weights of the measured sites, perform interpolation calculations, as shown in the following formula: ； In the formula, N is the number of sample points. The observed value is at station i. The interpolated data for grid point j is used to obtain the measured precipitation data for the grid station. Step S4.4: For the normalized slope, temperature and wind speed interpolation, perform steps S4.1-S4.3 to obtain grid data consistent with satellite precipitation resolution.

6. The method for downscaling satellite precipitation based on actual measurement stations according to claim 1, characterized in that, The specific steps of step S5 are as follows: Step S5.1: Construct the training dataset. The input feature X is multi-source two-dimensional spatial grid data, including filtered original measured precipitation data, measured precipitation data from gridded stations, temperature data, wind speed data, and normalized slope data. The training target Y is the predicted low-resolution precipitation data, which is interpolated from the grid points. composition; Step S5.2: Construct the LightSRCNN model, which consists of an input layer, three convolutional layers, and an output layer. The input layer receives 5-channel two-dimensional spatial grid data. The first convolutional layer has 32 5×5 convolutional kernels, the second convolutional layer has 16 3×3 convolutional kernels, and the first two convolutional layers use the ReLU activation function. The third convolutional layer has one 3×3 convolutional kernel. The output layer directly outputs low-resolution precipitation data for each grid in a single channel. Step S5.3: Train the LightSRCNN model. The loss function during training consists of a weighted average mean squared error (MSE) loss and a gradient L1 loss. The gradient L1 loss weight is α=0.3, and the MSE loss weight is β=0.

7. No early stopping strategy is set. The LightSRCNN model is trained for 15 epochs. 80% of the data is used as the training set, and 20% of the data is used as the validation set. The validation set R² score is used as the performance evaluation metric. The model with the best validation set R² score is saved as a pth file. Step S5.4: Construct the LightGBM gradient boosting tree model. The input data is the same as the input of the LightSRCNN network. The task type is regression. The output is the predicted low-resolution precipitation data. The evaluation metric is RMSE. Set 31 leaf nodes, a learning rate of 0.05, and a feature bag ratio / sampling bag ratio of 0.

8. Step S5.5: Using mean squared error (MSE) loss as the loss function, without designing an early stopping strategy, train the boosting tree for 100 rounds; use 80% of the data as the training set and 20% of the data as the validation set, filter out NaN values ​​in the validation set, calculate the R² score to evaluate the model performance, and finally save the trained model pkl file.

7. The method for downscaling satellite precipitation based on actual measurement stations according to claim 1, characterized in that, The specific steps of step S6 are as follows: Step S6.1: Perform inference using the trained LightSRCNN model and LightGBM model respectively, and output LightSRCNN low-resolution precipitation data and LightGBM low-resolution precipitation data respectively; Step S6.2: Based on the divided region type, set the basic weights. Specifically, for plains, the basic weight of LightSRCNN low-resolution precipitation data is set to 0.5, and the basic weight of LightGBM low-resolution precipitation data is set to 0.5; for mountainous regions, the basic weight of LightSRCNN low-resolution precipitation data is set to 0.8, and the basic weight of LightGBM low-resolution precipitation data is set to 0.2; for coastal regions, the basic weight of LightSRCNN low-resolution precipitation data is set to 0.7, and the basic weight of LightGBM low-resolution precipitation data is set to 0.

3. Step S6.3: Divide the flood season into flood season and non-flood season, and further adjust the basic weights. If the time is June, July and August, it is the flood season. Adjust the basic weights of LightSRCNN low-resolution precipitation data and LightGBM low-resolution precipitation data in mountainous areas to 0.9 and 0.1 respectively, adjust the basic weights of coastal areas to 0.8 and 0.2 respectively, and adjust the basic weights of plain areas to 0.7 and 0.3 respectively. Step S6.4: Based on the defined basic weights, weighted fusion is performed to obtain low-resolution precipitation data, as follows: ； In the formula, For in the grid Low-resolution precipitation data, For in the grid Weights of the LightSRCNN model For in the grid LightSRCNN low-resolution precipitation data, For in the grid Weights of the LightGBM model This is low-resolution precipitation data from LightGBM.

8. The method for downscaling satellite precipitation based on actual measurement stations according to claim 1, characterized in that, The specific steps of step S7 are as follows: Step S7.1: Based on the measured station grid, read the low-resolution precipitation data corresponding to the measured station grid; Step S7.2: Subtract the corresponding low-resolution precipitation data from the measured station grid precipitation data to obtain the precipitation residual; Step S7.3: The precipitation residual is upsampled to 1km using the interpolation method in step S4 to obtain a high-resolution precipitation residual.

9. A method for downscaling satellite precipitation based on measured stations according to claim 1, characterized in that, The method for calculating downscaled precipitation data in step S9 is as follows: ； In the formula, the final precipitation is the downscaled precipitation data. For high-resolution precipitation data, This represents the high-resolution precipitation residual.

Images