Abstract:

The establishment of an accurate, reliable and efficient landslide susceptibility assessment method is a key tool for pre-disaster scientific warning and comprehensive prevention and control. However, the traditional landslide susceptibility evaluation method fails to effectively address the prediction bias caused by the spatial heterogeneity and redundant features. To address this problem, this paper proposes a method for evaluating landslide susceptibility (SF-Stacking) that takes into account spatial heterogeneity and feature optimization. The method first uses AGNES (agglomerative nesting) to divide the global raster cells into several local regions, then uses a strategy which takes into account feature optimization to select the optimal combination of feature factors for each sub-region, and finally uses Stacking integration technology to couple multiple machine learning algorithms to achieve landslide susceptibility evaluation. Using Yibin city as the study area, the SF-Stacking method is compared with seven state-of-the-art methods based on the landslide hazard susceptibility zoning map and statistical indicators. Results show that the SF-Stacking method has the best accuracy, the highest robustness and the best interpretability.

Key words: landslide susceptibility, spatial heterogeneity, feature optimization, Stacking ensemble learning