SVM classification-driven correction for spatial distribution errors in remote sensing precipitation and its application

Accurate estimation of precipitation is crucial for watershed hydrological cycle simulation, water resource management, and drought/flood disaster early warning. Current remote sensing precipitation products can provide continuous spatial distribution but are limited by inversion algorithms and information source quality, making their accuracy insufficient for refined applications. This study proposes a method for correcting spatial distribution errors of remote sensing precipitation based on Support Vector Machine (SVM) classification-regression, taking the Upper Hanjiang River Basin as the study area. The method aims to integrate the local accuracy of ground observations with the spatial continuity of remote sensing data, while considering the impact of underlying surface changes. A database containing daily precipitation from 392 rain gauges, MSWEP remote sensing products, and underlying surface factors (elevation, slope, aspect) in 2016 was constructed. A 40km bandwidth correction domain was determined via cross-validation, and an SVM model set was built using the training set to correct MSWEP data for the test set. Results show that after correction, the multi-station average ME, MAE, and SRMSE of the test set decreased by 72.26%, 32.93%, and 23.26%, respectively, with CC increasing from 0.70 to 0.83. For typical precipitation events on June 1 and July 13, 2016, the corrected data showed better spatial consistency with observations, with more accurate positioning of high-value centers. The 2km×2 km spatial downscaling reconstruction achieved significant effects. This study provides a reliable method to improve remote sensing precipitation quality, supporting watershed hydrological forecasting and flood control decisions.