Scenario-based analysis of rainfall erosivity trends in Taiwan

This study proposes a two-level stacking machine learning approach for predicting rainfall erosivity ( R m ) in Taiwan, providing a flexible alternative to traditional empirical methods. Conventional models rely on limited high-resolution rainfall data and are often region-specific, which limits their accuracy elsewhere. In contrast, the proposed ensemble framework captures complex, non-linear interactions among climatic and topographic variables to improve prediction accuracy. In the first level, six base models were combined, and in the second level, each base model was used as a meta-model to form the ensemble structure. Twenty-eight predictor variables, including climatic and topographic factors, were derived from Coupled Model Intercomparison Project Phase 6 (CMIP6) high-resolution global climate data and a digital elevation model (DEM). To ensure robustness, the modeling procedure was repeated five times using different train–test splits, and final performance metrics were calculated as averages across five datasets. Feature selection using Boruta identified rainfall-related variables as the most important contributors. The ensemble approach significantly improved predictive performance, achieving a root mean square error (RMSE) of 5317 . 92 ± 261 . 23 MJ ⋅ mm ⋅ ha − 1 ⋅ hour − 1 ⋅ year − 1 and a Nash–Sutcliffe efficiency (NSE) of 0 . 67 ± 0 . 02 . The analysis revealed an increasing trend in R m , particularly under higher emission scenarios (SSP3-7.0 and SSP5-8.5), with increases projected in the latter half of the century. These findings highlight the importance of targeted climate mitigation and adaptation strategies for soil conservation and watershed management. This study supports Sustainable Development Goals 13 (Climate Action) and 15 (Life on Land) by improving R m prediction to reduce land degradation and enhance climate resilience. • Two-level stacking ensemble ML framework predicts rainfall erosivity ( R m ) in Taiwan. • Combined six base and meta models with 28 climate and DEM predictors. • Random forest (RF) meta-model achieved best accuracy (NSE = 0.67, RMSE = 5317.92 MJ ⋅ mm ⋅ ha −1 ⋅ hour −1 ⋅ year −1 ). • R m shows increasing trends under high-emission scenarios in late 21st century.