AI-Driven Sub-seasonal Landslide Forecasting in Nepal for Disaster Preparedness (Papers Track)
Kelsey Doerksen (University of Oxford); Sihan Li (Sheffield University); Yarin Gal (University of Oxford); Freddie Kalaitzis (Aspia Space); Alexander Densmore (Durham University); Alexandre Dunant (Eurac Research); Nick Rosser (Durham University); Simon Dadson (University of Oxford)
Abstract
Landslides can be deadly natural disasters, particularly in Nepal, caused by large earthquakes along the India-Asia collision zone and intense monsoon rainfall. It is well understood that the reliability of monsoon seasonal landslide forecasting is heavily reliant on the qualities of the rainfall data and the landslide data archive. However, the link between precipitation thresholds and landslides in the region has been derived through linear correlation or regression methods that are thought to be oversimplifying the relationship between the two, and often a single threshold is applied over the whole country. Risk maps have been generated from historical data, but do not provide forecasts of landslide occurrence and are done on seasonal timescales, limiting their usefulness in short-term disaster preparedness efforts and anticipatory actions. We propose the use of Machine Learning and Deep Learning techniques to forecast landslides across the entirety of Nepal using a combination of geomorphic data, precipitation observations and sub-seasonal precipitation forecasts from an ensemble of dynamical forecast models on a rolling daily basis. We present two methods using open-source Earth Observation data in a tabular and multi-channel array format for landslide forecasting on a District-level across the entirety of Nepal. We further explore the relative skills of landslide prediction using precipitation forecasts from three distinct dynamical forecast models, a comparison that has not been done before. We achieve our highest F1-score of 0.79 with our UNet architecture and show consistent good performance of landslide forecast on a 14-day lead time throughout the monsoon season.