Machine learning-based sediment connectivity surrogates for RUSLE and MUSLE in ungauged watersheds

Abstract Sediment yield prediction is vital for sustainable watershed management, particularly in data-scarce regions. This study, conducted in the Göksun Çayı Karaahmet sub-basin, Türkiye, evaluated whether sediment connectivity indices can reproduce outputs from the Revised Universal Soil Loss Equation (RUSLE) and Modified Universal Soil Loss Equation (MUSLE). Sediment yield was modeled for 196 sub-catchments and 69 rainfall events over 10 years using GIS-based factors: rainfall erosivity, soil erodibility, slope length-steepness, land cover, and hydrological parameters. Despite different assumptions (rainfall erosivity versus runoff and peak discharge), RUSLE and MUSLE showed strong agreement (R² = 0.87 at the event scale; R² = 0.93 at the sub-catchment scale). Predicted sediment yields ranged from 0.02 to 16.46 t ha -1 (MUSLE) and 0.04–10.63 t ha -1 (RUSLE/SDR), with mean values of 0.89 and 0.96 t ha -1 , respectively. Sediment connectivity indices-including the Index of Connectivity (IC), Sediment Delivery Ratio (SDR), and Topographic Wetness Index (TWI), were applied as inputs to five machine learning (ML) models (XGBoost, Random Forest, k-NN, SVR, and ANN). XGBoost and Random Forest achieved the best performance (R² = 0.912–0.942, RMSE = 0.065–0.089, MAE = 0.047–0.055), reproducing empirical outputs. IC, SDR, and TWI were dominant predictors. These results demonstrate that connectivity metrics integrated with ML can emulate empirical erosion models, offering a scalable, data-efficient alternative for ungauged basins. However, because the models were trained on RUSLE/MUSLE outputs from 69 events under static land use and climate, they may underpredict extreme sediment events and require field validation before operational use.