Assessing the Spatial Transferability of Annual Soil Erosion Models Across Mediterranean Catchments Using Hierarchical and Ensemble Approaches

Spatial transferability remains a major limitation in erosion modelling across environmentally heterogeneous catchments. This study evaluated climatic, vegetation, and morphometric predictors and compared a linear model, Extreme Gradient Boosting (XGBoost), and a hybrid stacking ensemble using 93 catchment–year records from three semi-arid Mediterranean catchments in southern Spain. The methodological contribution lies in jointly evaluating hierarchical variance partitioning and targeted predictor interactions, comparing the linear and XGBoost models under leave-one-catchment-out (LOCO) validation, and examining an internally fitted stacking ensemble based on pooled LOCO base-model predictions. Precipitation was the strongest statistical predictor of the Universal Soil Loss Equation (USLE)-derived annual erosion estimates. Targeted linear interactions did not consistently improve LOCO performance, and the estimated catchment-level variance approached zero after inclusion of the environmental predictors. Under pooled LOCO validation, the linear model yielded a squared Pearson correlation of 0.39, and root mean square error (RMSE) of 47.9 t ha−1 yr−1. XGBoost reduced RMSE to 27.9 t ha−1 yr−1, while the squared correlation was 0.35. he internally fitted stacking meta-model yielded an apparent RMSE of 23.6 t ha−1 yr−1 and a squared correlation of 0.53 on the same pooled predictions used for meta-model estimation. Transferability was catchment-dependent and poorest for Casasola. Because the response was model-derived, only three catchments were available, and the meta-learner lacked an outer spatial validation layer, the stacking results represent an internal comparison rather than definitive evidence of regional transferability. Lower prediction error did not necessarily imply broader spatial generalization or improved process understanding.