The enhanced global landslide nowcast and recent major events

Abstract:

The Landslide Hazard Assessment for Situational Awareness (LHASA) model identifies potential landslide activity as a “nowcast” from sixty degrees north to south by merging satellite precipitation with a susceptibility map. Recently, LHASA has been enhanced to incorporate new landslide inventories and inputs into a new model structure. Machine learning is at the core of LHASA Version 2. The model was trained with XGBoost, a commonly used tool in data science. Lesser known options of XGBoost, interaction constraints and monotonicity constraints, were invaluable for ensuring physical realism of model outputs. The revised nowcast maintains the same daily 1-km resolution as LHASA version 1.1, but it offers a continuous probability estimate in addition to a discrete output.

In order to verify the nowcast performance, a retrospective model run was completed for the years 2015-2020. The model successfully identified the potential for rainfall-triggered landslides associated with several major events, such as the impact of Cyclone Harold on Vanuatu. However, not every incident was assigned a high probability of slope failure. This indicates the limitations of existing globally available datasets on major predictors of landslide occurrence, such as rainfall. Overall, LHASA Version 2 represents a significant improvement over the previous global landslide nowcast.