A Dynamic Model for Rainfall-Triggered Landslides in Central America

Thursday, 18 December 2014
John Mordecai Doane Simmons, Columbia University of New York, Earth Institute, Palisades, NY, United States
Rainfall-triggered landslides are a consistent threat to life and property around the world and are extremely challenging to physically model over large areas. The Dynamic Landslide Assessment model combines a static landslide susceptibility map with a decision tree approach that considers satellite-derived rainfall and soil moisture thresholds to designate landslide “warning” or null value for the evaluated pixel. This model is used in order to facilitate the development and implementation of an online landslide ‘nowcast’ system for the region of Central America. For initial diagnostics, receiver operating characteristic (ROC) curves were generated from a 3-year model run of 125 landslide points in Central America.

To optimize the threshold necessary to trigger a landslide, normalized rainfall was evaluated against landslide occurrence in each model run and the True Positive Rates (TPR) and False Positive Rates (FPR) recorded. Each curve is evaluated against other instances by locating the minimum distance from a perfect fit (TPR = 1and FPR = 0) on a calculated ROC curve. The preliminary runs of the model are promising; the smallest R value is associated with a TPR of 0.79 and a FPR of 0.21. In addition, the inclusion of soil moisture data improves the skill of the model when rainfall is considered within a temporal window of a single day. However, consideration of larger precipitation windows (e.g. 3-day) somewhat diminishes the effectiveness of the inclusion of soil moisture data. Nevertheless, the simplicity of the model will continue to facilitate evaluation and inclusion of new data sources and metrics, as well as rapid integration of the model components within the developing online system.