基于随机森林模型的黄土高原典型小流域不同样本密度的沟蚀风险预测OA

[Objective]The study explores the optimal sample density for gully erosion risk prediction in typical small watersheds of the Loess Plateau and analyzes the main feature factors influencing gully occurrence,thereby providing a reference for gully erosion prevention and control.[Methods]The Random Forest(RF)model combined with SHAP algorithm was used to predict gully erosion risks at sample densities of 25%,50%,75%,and 100%of the total gully samples,and the contributions of the dominant factors to the model output were quantified.[Results]The 50%sample density achieved the best predictive performance,with accuracy,precision,and Kappa coefficient reaching 0.901,0.894,and 0.802 respectively.These values significantly exceeded those of the 25%density(0.871,0.851,0.743),and were higher than the 75%(0.898,0.882,0.795)and 100%(0.899,0.880,0.798)density.The AUC values were 0.924,0.956,0.956,and 0.959 respectively across the four densities,and the recall rates were 0.887,0.910,0.917,and 0.924 respectively.Notably,the 50%density showed negligible differences in AUC value and recall rate compared to the 75%and 100%densities,so it was considered the optimal choice under the premise of ensuring accuracy.[Conclusion]Among all influence factors,land use type contributes the most to the gully erosion risk prediction,followed by slope and planar curvature.The coupling analysis of gully erosion risk and erosion severity indicates that moderate and severe erosion may occur in low-risk areas,suggesting that the gully erosion risk levels cannot fully represent the erosion severity,which provides a reference for gully erosion assessment and prevention in watersheds with similar conditions across the Loess Plateau.