Deep Learning Reveals Debris Flow Impact Forces from Seismic Signals

Debris flows are rapid mass movements that move down steep mountain creeks and are a major threat to human life, properties, and infrastructure.  As the debris flow travels down the channel, the impact force of the debris on the channel bed generates ground vibrations that propagate to and can be recorded by the seismometer. The impact force is an important parameter in the design of debris-flow damage mitigation, such as check dams. Direct measurements of impact force from debris flows are limited by the high cost of instrumentation and the risk of instruments being destroyed in the process. Installing and maintaining a seismic network outside the debris flow channel keeps it protected from the hazard and can be a suitable alternative to direct measurements of the impact force. Connecting seismicity to the generating impact force is complex due to the complicated environment. Bridging this gap using deep learning could help estimate physical information to improve debris flow warning.   

In this study, we train an extended-LSTM (xLSTM) model to invert impact force from seismic signals generated by debris flows in the Illgraben catchment, in Switzerland. We chose the xLSTM model ahead of others due to its ability to process long and complex sequences of data. We used seismic signals generated by debris flows as they pass through CD 27 and impact force signals recorded at CD 29 by an 8m² force plate. The xLSTM model is compared to the LSTM model architecture as a baseline, and we show that the xLSTM model performs better at capturing the distribution of the impact force and producing lower mean error. Along with this, it inverts the peak impact force with an absolute error of less than 1kN to the measured impact force.  Furthermore, we find a strong correlation between the volume and the cumulative impact (CIF) force for debris flows, showing that the xLSTM inverted impact force can be used to derive an initial constraint on the volume of a debris flow event. This method can support early warning systems for debris flow by allowing for quick impact force analysis and providing initial constraints on some physical characteristics, for example, debris-flow volume.