Full seismic waveform analysis combined with transformer neural networks improves coseismic landslide prediction

Abstract

Seismic waves can shake mountainous landscapes, triggering thousands of landslides. Regional-scale landslide models primarily rely on shaking intensity parameters obtained by simplifying ground motion time-series into peak scalar values. Such an approach neglects the contribution of ground motion phase and amplitude and their variations over space and time. Here, we address this problem by developing an explainable deep-learning model able to treat the entire wavefield and benchmark it against a model equipped with scalar intensity parameters. The experiments run on the area affected by the 2015 Mw7.8 Gorkha, Nepal earthquake reveal a 16% improvement in predictive capacity when incorporating full waveforms. This improvement is achieved mainly on gentle (~25°) hillslopes exposed to low ground shaking (~0.2 m/s). Moreover, we can largely attribute this improvement to the ground motion before and much after the peak velocity arrival. This underscores the limits of single-intensity measures and the untapped potential of full waveform information.