Evaluating the Reliability of Landslide Susceptibility Models Using Random Forest in Heterogeneous Landslide-Occurrence Region of the Indian Himalayas

The task of modelling landslide susceptibilityLandslide susceptibility becomes complex when landslides manifest unique behaviours influenced by distinct sets of conditioning factors within a specific region. A landslide susceptibilityLandslide susceptibility model that omits the landslide types can increase its bias. This paper generates a landslide inventoryInventory database encompassing all landslide types within a study area. Additionally, the inventoryInventory underwent a classification process that resulted in its division into five distinct categories: (i) the entirety of landslides, (ii) debris slides in their entirety, (iii) debris slides induced by natural slope conditions, (iv) debris slides induced by cut slopes, and (v) rockfall-slide events. The interrelationship between these landslide groups and conditioning factors is analysed through the relationship index (RI). We then evaluated and compared different landslide susceptibilityLandslide susceptibility models using a random forestRandom forest (RF) algorithm. Each model is based on a different set of landslide inventoryInventory. The inventoryInventory map covers the Kalimpong region (a part of Eastern HimalayaEastern himalaya). The investigation revealed notable distinctions between the conditioning factors and the responses exhibited by natural slope-induced debris slides, cut slope-induced debris slides, and rockfall-slide events. Further, susceptibility values based on these groups of landslides were found to be different from each other. One of the research's interesting outcomes is that although the RF susceptibility models are based on all landslides, despite having an accuracy level of more than 0.9, they still cannot reflect the susceptibility of cut slope-induced landslides and the group of rockfall and slide. Therefore, the predictive performance of models in terms of statistics is necessary for building a susceptibility model but cannot be the only metric to justify these models’ reliability. The study recommends that the landslide susceptibility model should consider the type of each landslide separately rather than holistic way.