Susceptibility assessment of rainfall-induced clustered landslides integrating ensemble machine learning and negative sample sampling strategies

Rainfall-triggered clustered landslides frequently occur in the hilly and mountainous regions of southern China, posing severe threats to local lives and property. Such geohazards feature short-term abrupt occurrence, multi-point concurrent failures, and strong spatial aggregation. Critically, landslide sites share highly similar geo-environmental conditions with adjacent stable slopes, which brings great challenges to landslide susceptibility modeling. Conventional negative sample sampling strategies often generate oversimplified non-landslide samples from environmentally distinct stable areas. These easy negative samples make machine learning models achieve seemingly high statistical scores while failing to capture subtle differences between unstable and stable slopes within landslide-prone zones, consequently producing widespread spatial misjudgments. Taking the clustered landslide event triggered by extreme rainfall in October 2023 across the Xinyi-Luoding border area, western Guangdong Province, China as the study object, this study aims to explore how coupling different negative sample sampling strategies with machine learning models affects the performance of landslide susceptibility assessment.

Based on pre- and post-event Planet satellite images, visual image interpretation was implemented to acquire 3792 positive landslide samples. Twelve evaluation factors were selected from five dimensions: meteorology and hydrology, topography and geomorphology, geology and soil, vegetation cover, and human activities. Spearman correlation analysis was applied to detect multicollinearity, and the elevation factor was eliminated due to strong collinearity. Under a 1∶1 positive-to-negative sample ratio and a unified 100-m buffer constraint around landslide boundaries, three types of negative sample datasets were established via buffer-based random sampling, low-slope constrained sampling and K-means unsupervised clustering sampling. The whole dataset was partitioned into a 70 % training subset and a 30 % testing subset. Grid search combined with five-fold cross-validation was adopted for hyperparameter optimization of random forest (RF), extreme gradient boosting (XGBoost), and Stacking ensemble learning models (using RF and XGBoost as base learners and logistic regression as the meta-learner). Model performance was quantified using AUC-ROC, accuracy (ACC), and F1-score, and the natural-breaks method was adopted to generate multi-grade landslide susceptibility maps.

The results demonstrated that Stacking ensemble learning algorithms could achieve relatively high baseline performance, yet negative sample sampling strategies had a prominent influence on model outputs. Noticeably, ACC and F1-score values only showed slight discrepancies across different sampling schemes. Nevertheless, obvious differences were observed in AUC values and spatial pattern rationality of susceptibility maps. Benefiting from its capacity to characterize environmental heterogeneity within non-landslide domains, the K-means clustering sampling combined with the Stacking ensemble model obtained the optimal overall performance with an AUC of 0.93, accompanied by the highest ACC and F1-score. The derived high- and very-high-susceptibility zones were mainly distributed in the northwestern transition zone between low-to-middle mountains and hills, showing good agreement with the spatial pattern of actual landslide clusters. By comparison, buffer-based random sampling yielded moderate performance, while low-slope constrained sampling performed worst and tended to produce overestimated susceptibility and scattered mapping noise.

This study shows that the K-means-based unsupervised clustering sampling can sufficiently capture the environmental heterogeneity of non-landslide areas in clustered landslide areas. When integrated with the Stacking ensemble learning, it improves both statistical accuracy and spatial-geological rationality of susceptibility outputs. The proposed workflow provides a reliable methodological reference for refined susceptibility evaluation of rainfall-induced clustered landslides in hilly and mountainous regions across southern China.