Spatial modeling of gully density on the Qinghai-Tibet Plateau: Application of hyperparameter optimization in interpretable machine learning

Machine learning has indeed become an important method for gully erosion modeling, but its accuracy in simulating gully density remains significantly lower than that of gully erosion susceptibility assessment. Predicting gully density and detecting the dominant factors of gully erosion on a large scale are significant challenges. In this study, a gully density length survey was carried out in a random area of 14187 qudarats (1*1 km2) of the Qinghai-Tibet Plateau to construct a basic gully density lengthdataset. Combined with the multiple collinearity test, 17 environmental factors were selected to construct the characteristic evaluation indexes. Various machine learning models were used, and different hyperparameter optimization algorithms were selected to train the models to obtain the best model. The results showed that the performance of the Bayesian-optimized XGBoost model was the best in general, and its prediction accuracy for gully density grade reached 73.479%. The current model demonstrated suboptimal predictive performance for gully density units as moderate and abovegrades, respectively. The primary source of error originates from units within the elevation range of 4000-5000 m a.s.l and slope gradients below 20°. Through the SHapley Additive exPlanations(SHAP), it was found that slope, elevation and vegetation cover were the main controlling factors affecting gully density in this area. The results provide a scientific reference for further understanding gully erosion development mechanism and soil and water conservation management measures in the Qinghai-Tibet Plateau, and provide the possibility for accurate prediction of global gully erosion intensity in the future.