Enhancing debris flow warning through seismic feature selection and machine learning model comparison

Machine learning can improve the accuracy of detecting mass movements in seismic signals and extend early warning times. However, we lack a profound understanding of the limitations of different machine learning methods and the most effective seismic features especially for the identifcation of debris flows. This contribution explores the importance of seismic features with Random Forest and XGBoost models. We find that a widely used approach based on more than seventy seismic features, including waveform, spectrum, spectrogram, and network metrics features, suffers from redundant input information. Our results show that six seismic features are sufficient to perform binary debris flow classification with equivalent or even better results., e.g., the Random Forest and XGBoost models achieve improvements over the benchmark of 0.09% and 1.10%, respectively, when validated on the ILL12 station. Considering models that aim to capture patterns in sequential data rather than information in the current time window, using the Long Short-Term Memory algorithm does not improve the binary classification performance over Random Forest and XGBoost models. However, in the early warning context, the Long Short-Term Memory model performs better and more consistently detects the initiation of debris flows. Our proposed framework simplifies seismic signal-driven early warning for debris flows and provides a proper workflow that can be used for detecting also other mass movements.