Development of multiclass alternating decision trees based models for landslide susceptibility mapping

Abstract

Worldwide landslide occurrences imply the need for intelligence tools to identify the most susceptible areas toward adopting efficient mitigation strategies and reaction plans. In this study, we developed three spatially explicit ensemble predictive models for the prediction of landslide susceptibility in the Muong Nhe district of the Dien Bien Province, Vietnam. The Multiclass Alternating Decision Trees (MADT) method was used as the base classifier with the Dagging, MultiboostAB, and Random Subspace (RSS) as the ensemble learners. The location of past landslides was identified through field surveys and the interpretation of Google Earth images, aerial photographs, and historical archives of the Muong Nhe district. The landslide locations were liked to twelve landslide conditioning factors (slope, aspect, elevation, curvature, topographic wetness index (TWI), stream power index (SPI), geology, flow accumulation, normalized difference vegetation index (NDVI), and distance to rivers, roads, and faults) to investigate the spatial patterns of landslide susceptibility across the study area. The results showed that the RSS-MADT model achieved the highest performance in terms of predicting future landslides (AUC = 0.878), followed by DG-MADT (AUC = 0.857), MAB-MADT (AUC = 0.854), and the single MADT model (AUC = 0.828), respectively. Approximately 13% and 10% of the Muong Nhe district were identified as having moderate and severe (high/very high) susceptibility to landslide occurrences. These areas that extend along the rivers, primarily in the central parts of the Muong Nhe district, should be treated with the highest priority to mitigate the negative impacts of future landslides.