Estimation of rainfall erosivity on the Chinese Loess Plateau: A new combination of the ERA5 dataset and machine learning

Abstract

Determining the characteristics of rainfall erosivity at different times and locations is significant for global and regional soil erosion prevention and control. However, accurate quantification of rainfall erosivity remains a challenge due to the limitations of high-precision and long-time series precipitation data. Here, compared with the rainfall erosivity of 291 meteorological stations on the Chinese Loess Plateau, we evaluated the distribution characteristics and accuracy of rainfall erosivity on different temporal and spatial scales calculated by the ERA5 precipitation dataset. A random forest model was constructed to generate a bias dataset to correct the annual rainfall erosivity. Besides, the relationship between erosivity density (ED) and rainfall patterns was quantified to clarify the characteristics of rainfall erosivity. Conducted analysis suggested that the rainfall erosivity calculated from the ERA5 precipitation dataset has high accuracy on average annual, seasonal and monthly scales, and the R2 were 0.62, 0.86, and 0.85, respectively. The addition of the random forest model further optimized the result of average annual rainfall erosivity of ERA5, with root mean square error (RMSE) and absolute value of relative bias (BIAS) decreasing to 185.7 MJ mm ha−1h−1 a−1 and 2.17%, respectively. Meanwhile, ERA5 can well capture the average annual rainfall erosivity in different spaces of the study area. The high correlation between erosivity density and rainstorms (R2 = 0.87) indicated that a larger proportion of rainstorms in erosive rainfall will lead to a greater potential risk of soil erosion. These findings provide a reliable approach to accurately obtain the variable rainfall erosivity in areas with complex terrain and sparse stations, and have important implications for soil sustainability.