An SPCL-Stacking ensemble learning framework for landslide susceptibility assessment

Background Machine learning has been increasingly applied to data driven landslide susceptibility assessment. However, false negative label noise introduced during negative sample selection and the limited predictive capability of single models remain unresolved issues. To address these issues, this study proposes an SPCL-Stacking landslide susceptibility assessment method that integrates Self-Paced Curriculum Learning (SPCL) with heterogeneous ensemble learning. Results Using slope units as the basic assessment units, SPCL-Stacking organized candidate negative samples through a fixed curriculum threshold sequence. With $$T=10$$ , the highest mean cross validation AUC occurred at curriculum level $$t=8$$ . In the fixed reference experiment using random seed 42, SPCL-Stacking achieved an AUC of 0.8444, an Accuracy of 0.7827, and an F1 score of 0.5903. Across 50 paired repetitions, its mean AUC was 0.8386, compared with 0.8290 for traditional Stacking-LR, with a mean paired difference of 0.0096 (95% CI 0.0075–−0.0116). The high and very high susceptibility classes occupied 44.06% of the study area, contained 88.45% of the positive samples, and reached a relative distribution density of 2.05. SHAP ranked Rainfall, PLC, CEV, DTRI, and DTRO as the five conditioning factors with the largest mean absolute contribution values. Conclusions The proposed method integrates candidate negative sample optimization with heterogeneous ensemble learning, achieving a modest but consistent improvement in predictive performance and a susceptibility classification with increasing positive sample concentration across classes. It provides a data driven approach for county scale landslide susceptibility screening and geoenvironmental disaster prevention in Fengjie County.