Landslide susceptibility assessment through multi-model stacking and meta-learning in Poyang County, China

Abstract

This study aims to evaluate the effectiveness of various individual machine learning and their ensemble techniques such as Stacking, Voting and Meta-learning in landslide susceptibility assessment taking Poyang, Jiangxi, China as an example. Multi-source geo-environmental data including field surveys, Sentinel-2A/B satellite images, Digital Elevation Models (DEM), and geological and hydrological data were utilized to construct and validate landslide susceptibility models. Results show that the Stacking Classifier outperformed other models, achieving the highest F1 Score of 0.846 and AUC (Area Under ROC Curve) of 0.923, demonstrating its strong predictivity, followed by the Voting Classifier with the F1 Score of 0.829 and AUC of 0.922. Among the individual models, the Multi-Layer Perceptron (MLP) performed best with the F1 Score of 0.828 and AUC of 0.904. Furthermore, the explainable Artificial Intelligence (XAI) technique was applied to better understand the mechanism of classifiers in predicting landslide susceptibility and it suggests a significant correlation between land use, distance to fault, and landslide occurrences. In conclusion, Stacking and Voting hybrid learning models show clear advantages over the individual ones for landslide risk zoning. The results of study may provide technical support for disaster mitigation efforts and future urban planning in areas prone to landslides in Poyang.