Landslides forecasting using satellite rainfall estimations and machine learning in the Colombian Andean region

Abstract

Due to its high topographic gradients and climatic conditions, the Colombian Andean Region (CAR) has been associated with human and economic losses caused by rainfall-triggered landslides over time; for this reason, the definition of the rainfall conditions that trigger landslides has received significant research interest. However, numerous areas that constitute this region do not have ground-rainfall gauge networks to correlate rainfall and landslides. In this sense, the information provided by Satellite Rainfall Estimations (SREs) can help address this deficiency. Therefore, this work aimed to propose a methodology to spatialise the rainfall-triggered landslide hazard over the Andean Region using conditioning variables, SREs (CHIRPSv2 and MSWEPv2.6), and commonly used ground-rainfall gauges. Thus, the performance of the selected SREs is assessed using the modified Kling Gupta efficiency Metric (KGE’) and its components. Subsequently, the Random Forest Model (RFM) was trained for the 1981–2019 period, and its performance was assessed in terms of different skill scores and the Area Under the Receiver Operating Characteristic Curve (ROC-AUC). The model's predictive capacity for rainfall-triggered landslides was also tested using input data from recent events outside the training period (2020–2021). The KGE's results showed a good correlation between rainfall estimates (CHIRPSv2 and MSWEPv2.6) and observations (ground-rainfall gauges) for accumulated rainfall scales greater than three days, being better at the monthly and annual scales (KGE'>0.5). In addition, the performance of the RFM was good in all evaluated cases (ROC-AUC>0.78), with the ground-rainfall database outperforming the satellite databases. The results indicated that CHIRPSv2 is the best-performing SRE for landslide prediction in the Colombian Andean Region. Finally, the landslide forecast model results proved that the RFM could predict rain-triggered landslides using only rainfall variables, providing a suitable option for Landslides Early Warning Systems (LEWS). Lastly, the model results indicated an accurate probability of occurrence using CHIRPSv2 estimations, suggesting that the model would also yield good results in areas with a scarcity of ground-rainfall data.