Multi‐Source Rainfall Erosivity Fusion Based on Machine Learning Models

ABSTRACT High‐accuracy rainfall erosivity (RE) data are essential for understanding soil erosion processes. However, considerable differences exist among various precipitation datasets in terms of RE estimation. This study aimed to develop a machine learning‐based framework for multisource RE fusion to increase the accuracy and spatial consistency of RE estimation. To achieve this goal, RE data derived from station observations were used as a benchmark to systematically evaluate the performance characteristics of five commonly used gridded precipitation datasets (Climate Hazards Group InfraRed Precipitation with Station data [CHIRPS], CPC, CN05.1, CHM, and fifth‐generation atmospheric reanalysis data from the European Centre for Medium‐Range Weather Forecasts [ERA5]) in estimating RE across mainland China from 1983 to 2020, thereby identifying their biases. A machine learning‐based multisource RE fusion framework was subsequently established by employing six algorithms, namely, linear regression (LR), decision tree regression (DTR), random forest (RF), extra trees regressor (ETR), extreme gradient boost (XGB), and gradient boosting machine (GBM), to integrate RE estimates from the above five data sources. The results revealed that (1) the station‐interpolated precipitation products (CHM and CN05.1) outperformed the reanalysis dataset (ERA5), yet all five gridded datasets (CHIRPS, CPC, CN05.1, CHM, and ERA5) generally exhibited RE underestimation; (2) the fused RE product provided a significantly increased estimation accuracy, with the average root mean square error (RMSE) decreasing from 1826.21 to 1077.06 MJ mm ha −1 h −1 and the Nash–Sutcliffe efficiency (NSE) coefficient increasing from 0.67 to 0.89. The fusion method effectively corrected the common issues of the underestimation of high‐intensity RE events (large and heavy RE) and the overestimation of moderate RE events observed in the five original datasets (CHIRPS, CPC, CN05.1, CHM, and ERA5); and (3) the fused RE dataset achieved significant improvements in complex and observation‐sparse regions, such as the Tibetan Plateau (the average NSE increased from −0.30 to 0.60). Overall, the potential of machine learning‐based fusion methods in enhancing RE estimation was demonstrated, and a high‐resolution fused RE dataset that could serve as a reliable foundation for soil erosion assessment and land management was provided.