An ensemble neural network approach for space-time landslide predictive modelling

There is an urgent need for accurate and effective Landslide Early Warning Systems (LEWS). Most LEWS are currently based on a single temporally-aggregated measure of rainfall derived from either in-situ measurements or satellite-based rainfall estimates. Relying on a summary metric of precipitation may not capture the complexity of the rainfall signal and its dynamics in space and time in triggering landslides. Here, we present a proof-of-concept for constructing a LEWS that is based on an integrated spatio-temporal modelling framework. Our proposed methodology builds upon a recent approach that uses a daily rainfall time series instead of the traditional cumulated scalar approximation. Specifically, we partition the study area into slope units and use a Gated Recurrent Unit (GRU) to process a satellite-derived rainfall time series and combine the output features with a second neural network (NN) tasked with capturing the effect of terrain characteristics. To assess if our approach enhances accuracy, we applied it in Vietnam and compared it against a standard modelling approach that incorporates terrain characteristics and cumulative rainfall over 14 days. Our protocol leads to better performance in hindcasting landslides when using past rainfall estimates (CHIRPS), as compared to the standard modelling approach. While not tested here, our approach can be extended to rainfall obtained from weather forecasts, potentially leading to actual landslide forecasts.