Enhancing Debris Flow Warning via Machine Learning Feature Reduction and Model Selection

Abstract

The advent of machine learning has significantly improved the accuracy of identifying mass movements through the seismic waves they generate, making it possible to implement real-time early warning systems for debris flows. However, we lack a profound understanding of the effective seismic features and the limitations of different machine learning models. In this work, we investigate eighty seismic features and three machine learning models for single-station-based binary debris flow classification and multi-station-based warning tasks. These seismic features, derived from physical and statistical knowledge of impact sources, are grouped into five sets: Benford's law, waveform, spectra, spectrogram, and network. The machine learning models belong to two families: two ensemble models, Random Forest and eXtreme Gradient Boosting (XGBoost); one recurrent neural network model, Long Short-Term Memory (LSTM). We analyzed feature importance from the ensemble models and found that the number and even the types of seismic features are not critical for training an effective binary classifier for debris flow. When using models designed to capture patterns in sequential data rather than focusing on information only in one given window, using the LSTM does not significantly improve the performance of binary debris flow classification task over Random Forest and XGBoost. For the multi-station-based debris flow warning task, the LSTM model predicts debris flow probability more consistently and provides longer warning times. Our proposed framework simplifies machine learning-driven debris flow classification and lays the foundation for affordable seismic signal-driven early warning using a sparse seismic network.