Mapping landslide susceptibility with the consideration of spatial heterogeneity and factor optimization

Spatial heterogeneity and information redundancy of landslide influencing factors (LIFs) greatly impair the generalizability of landslide susceptibility mapping (LSM) models. To this end, this paper proposes a new LSM method that takes into account spatial heterogeneity and factor optimization. Firstly, a method based on frequency ratio, coupled with buffer-controlled sampling, is developed to extract non-landslides from the non-landslide area. Then, the study site is divided into several homogeneous areas using agglomerative clustering based on LIFs and spatial locations. Next, the LIFs are optimized in each region based on a combination of variance inflation factor, the Boruta algorithm, and geographical detector so as to avoid information redundancy and noise from both statistical and spatial perspectives. Finally, the random forest (RF) model with the optimized LIFs is used for LSM at the regional scale. Taking 686 landslides and 15 LIFs in Yibin city, China as an example, the proposed method was compared with four state-of-the-art models for LSM including regional RF, global RF with factor optimization, global RF using clustering attribute as one of its inputs, global RF without factor optimization, and global RF using a combination of factor optimization and clustering attribute. Results indicate that compared to the four classical models, the proposed method increases the Accuracy, Recall, Precision, F1 score, and the area under the receiver operating characteristic (AUC) curve by 1.6–5.2%, 2.9–12.5%, 0.1–3.5%, 2.9–7.0%, and 1.8–4.7%, respectively. Additionally, the proposed method can produce more accurate and reasonable landslide susceptibility maps, with an increase in the disaster activity intensity index by 2.7–20.8%. Overall, the proposed method presents a viable alternative for the spatial forecast of landslide susceptibility.