Machine-learning based landslide susceptibility modelling with emphasis on uncertainty analysis

Landslide susceptibility maps are vital tools used by decision-makers to adopt mitigation strategies for future calamities. In this context, research on landslide susceptibility modelling has become a topic of relevance and is in constant evolution. Though various machine-learning techniques (MLTs) have been identified for landslide susceptibility modelling, the uncertainty inherent in the models is rarely considered. The present study attempts to quantify the uncertainty associated with landslide prediction models by developing a new methodological framework based on the ensembles of the eight MLTs. This methodology has been tested at the highlands of the southern Western Ghats region (Kerala, India), where landslides have frequently been occurring. Fourteen landslide conditioning factors have been identified as part of this study, and their association was correlated with 671 historic landslides in the study area. The study used four ensemble models such as the mean of probabilities, the median of probabilities, the weighted mean of probabilities, and the committee average. The weighted mean of probability was proved to be the best model based on the average of 800 standalone MLTs, viz., receiver operating characteristics, true skill statistics, and area under curve with corresponding validation scores. Thereafter, an uncertainty analysis was carried out on the coefficient of variation. A confident map was generated to represent the distinct zonation of landslide susceptibility areas with definite uncertainty scales. Nearly 74% of the past landslides fall in the higher susceptibility-low uncertainty category. It is also inferred that such micro-level zonation based on MLTs may improve the efficiency of landslide susceptibility maps and may help in accurately identifying landslide-prone areas in the future. The confident maps thus generated can be used as a ready reference to the planners for the formulation of landslide adaptation strategies at micro-scales.