Reducing Local Correlations Among Causal Factor Classifications as a Strategy to Improve Landslide Susceptibility Mapping

Landslide susceptibility map (LSM) is the basis of hazard and risk assessment, guiding land planning and utilization, early warning of disaster, etc. Researchers are often overly keen on hybridizing state-of-the-art models or exploring new mathematical susceptibility models to improve the accuracy of the susceptibility map in terms of a receiver operator characteristic curve. Correlation analysis of the causal factors is a necessary routine process before susceptibility modeling to ensure that the overall correlation among all factors is low. However, this overall correlation analysis is insufficient to detect a high local correlation among the causal factor classes. The objective of this paper is to identify three questions: 1) Is there a high correlation between causal factors in some parts locally? 2) Does it affect the accuracy of landslide susceptibility assessment? 3) How to eliminate this influence? To this aim, taking Wanzhou County as the test site, where landslide susceptibility assessment based on 12 causal factors has been previously finished using frequency ratio (FR) model and random forest (RF) model. In this work, we conducted a local spatial correlation analysis of the "Altitude" and "Rivers" factors and found a sizeable spatial overlap between Altitude-class1 and Rivers-class1. The "Altitude" and "Rivers" factors were reclassified, then, the FR model and RF model were used to re-evaluate the susceptibility and analyze the accuracy loss caused by the local spatial correlation of the two factors. The results demonstrated that the accuracy of LSMs was markedly enhanced after reclassification of "Altitude" and "Rivers", especially for the RF model-based LSM. This research shed new light on the local correlation of causal factors arising from particular geomorphology and their impact on susceptibility.