Influence of index contribution rate and machine learning models on Benggang susceptibility at the slope unit scale

To evaluate how feature contribution thresholds and machine learning algorithms affect Benggang susceptibility prediction, this study used Huichang County, Jiangxi Province, China, as the study area. The multi-scale segmentation (MSS) technique was used to delineate 8,473 slope units as basic mapping units. Twenty-two environmental indicators related to rainfall, lithology, vegetation, terrain, soil, and hydrology were ranked by GeoDetector q-values, and seven feature subsets were constructed according to cumulative contribution-rate thresholds of 100%, 95%, 90%, 85%, 80%, 75%, and 70%. Logistic regression (LR), eXtreme Gradient Boosting (XGBoost), and random forest (RF) models were then applied to predict Benggang susceptibility. The results show that model accuracy did not increase monotonically with the inclusion of additional indicators. Instead, the area under the receiver operating characteristic curve (AUC) first increased and then decreased as low-contribution indicators were introduced. The optimal threshold was 90%, at which redundant information and statistical noise were most effectively reduced. At this threshold, RF achieved the best overall performance, with an AUC of 0.849, followed by XGBoost and LR. The proposed GeoDetector-MSS-RF framework improves slope-unit-scale Benggang susceptibility mapping and provides a practical basis for erosion prevention, spatial planning, and ecological management in similar red-soil hilly regions.