A Climate-Compatible Machine-Learning Framework for Future Rainfall Erosivity Projections Using Standardized Climate Model Outputs

<title>Abstract</title> <p> Rainfall erosivity (R ₃₀ ) is a key driver of soil erosion, yet projecting its future evolution remains challenging because standard climate-model outputs generally lack the sub-daily rainfall information required by conventional erosivity estimation methods. This study develops a climate-compatible machine-learning framework that estimates rainfall erosivity directly from predictor variables routinely available in climate-model outputs, providing a transferable approach for large-scale climate impact assessments. Historical rainfall erosivity observations from 209 meteorological stations across Türkiye were used to train a Random Forest emulator using strict station-wise GroupKFold spatial cross-validation. The validated model was then driven by annual precipitation projections from three global climate models (GFDL, HadGEM, and MPI) under the RCP4.5 and RCP8.5 scenarios to project rainfall erosivity for 2025–2095. The emulator explained approximately 54% of the observed variability under spatial validation (R² = 0.54). National projections indicate a moderate decline in mean rainfall erosivity across all climate model–scenario combinations, while regional and station-level analyses reveal pronounced spatial heterogeneity, including localized increases in several coastal and transitional climatic regions. Multi-model ensemble analyses further demonstrate increasing divergence among climate projections toward the end of the century, emphasizing the importance of accounting for climate-model uncertainty. By relying exclusively on predictors consistently available from climate projections, the proposed framework provides a practical and computationally efficient bridge between standardized climate-model outputs and rainfall erosivity assessment, supporting large-scale soil erosion risk evaluation and climate adaptation planning. </p>