Towards Artificially Intelligent Landslide Susceptibility Mapping: A Critical Review and Open Questions

Abstract

Since the 1970s, the scientific community has dedicated significant efforts to the development of landslide susceptibility models through various approaches, with the current spotlight firmly on artificial intelligence techniques. Despite their unique advantages, these cutting-edge tools have introduced significant challenges, the solution of which hinges on critical user decisions. These decisions chiefly revolve around selecting landslide conditioning factors and designing the optimal configuration of internal mechanisms of susceptibility modeling approaches—both critical determinants influencing model predictive accuracy. To address origin of these issues, a systematic review of literature spanning seven years, from 2015 to 2021, was conducted. The results revealed the utilization of 151 various landslide conditioning factors, highlighting a clear dearth of consensus on the selection of geospatial covariates in the literature. Nonetheless, only about one-third of the reviewed articles considered the feature selection techniques to seek the optimal factor subset. The review also showed that 54 distinct machine learning algorithms were used, with logistic regression being the most commonly applied susceptibility modeling approach, featured in 70 articles. Notably, deep learning algorithms were marginally employed, appearing in a mere 7.08% of the reviewed articles since 2018. However, a significant proportion (64.32%) of the articles used non-optimized predictive models with default settings, while a trial-and-error approach was adopted in 10.81% of the reviewed literature. Beyond the comprehensive literature review, this chapter delves into a series of ill-explored open questions and reveals opportunities that can serve as potential research roadmaps, potentially guiding the trajectory of future studies in landslide susceptibility mapping. © 2024 Elsevier B.V., All rights reserved.