A comparative study of various combination strategies for landslide susceptibility mapping considering landslide types

• Explores the link between landslide types and environmental factors. • Tailored LSMs for each landslide type capture development patterns. • “Divide-and-conquer” strategies outperform traditional methods. • Proposes a novel procedure for creating a reliable susceptibility map. Landslide susceptibility mapping (LSM) assists planners, local administrations, and decision-makers in preventing, mitigating and managing associated risks. This study proposes a novel DES-based framework that effectively captures the spatial developmental patterns of different landslide types, leading to higher precision LSM. The Wanzhou district (administrative division) of Chongqing Province, southwestern China, was selected as the test area, encompassing 881 landslides classified into rockfalls, reservoir-affected (RA) landslides, and non-reservoir-affected (NRA) landslides. Subsequently, three inventory maps and sixteen environment factors were used as inputs, with multicollinearity and importance analyses used to select the best factor combination for three types of landslides. Finally, the susceptibilities of rockfalls, RA and NRA landslides were combined by six combination strategies: Maximum, Mean, Probability, Voting, Stacking, and Dynamic Ensemble Selection (DES) models, and the optimal strategy was identified by area under the receiver operating characteristic curves (AUC), confusion matrix, and landslide distribution statistic. For LSM of individual landslide types, ResNet consistently outperformed traditional machine learning models, achieving testing AUC values of 0.8925, 0.9427, and 0.6754 for rockfalls, RA, and NRA landslides, respectively. The evaluation of the combination strategies revealed that the DES model achieved the highest testing AUC value of 0.8779, followed by Stacking (0.8728), Maximum (0.8704), Probability (0.8669), and Voting (0.8653), whereas the widely-used Mean method performed the worst (0.8503), even lower than the non-classified LSM (0.8587). The findings offer a robust approach for mitigating future landslide risks and minimizing their adverse impacts, providing valuable insights for geohazard management and decision-making.