[S-SS07] Environmental Seismology: from deep earth to surface process

[SSS07-15] Detection of debris flow based on seismic signals: an integrated approach combining machine learning and signal processing

*Shuaixing Yan¹, Dongpo Wang¹, Hui Chen^1,2, Xiaopeng Li², Yanjie Fan² (1.State Key Laboratory of Geohazard Prevention and Geoenvironment Protection, Chengdu University of Technology, 2. School of Mathematical Sciences, Chengdu University of Technology)

Keywords:debris flow, seismic signal, random forest, STA/LTA, early warning

Rapid and accurate detection of debris flow events is crucial for constructing high-resolution monitoring and early warning system of debris flow based on seismic signals, whereas the complexity of debris flow signal and the dynamic change of background noise bring great challenges to the development of automatic detection technology. This study proposes an integrated approach combining machine learning and signal processing to identify and pick up the occurrence time of debris flow events in complex environmental conditions. We first proposed an improved random forest-based model to enhance accuracy and generalization performance across diverse geographical contexts. A comprehensive dataset of historical debris flow events from 12 global regions were compiled and the Boruta feature selection algorithm were employed to select five key features. To augment the model's generalization capacity, we incorporated preprocessing techniques and Bayesian hyperparameters optimization. Furthermore, by analyzing the time-frequency energy evolution and statistical characteristics of debris flow events, the detection standard is dynamically adjusted by using the improved Short Time Average / Long Time Average (STA/LTA) to adapt to diverse time monitoring scenarios. Results show that the proposed model exhibited superior performance in identifying debris flow events across various regions, capturing the occur time of debris flow signal, and judging its duration, with relatively low false positive rate and missing rate. The enhanced model contributes to the advancement of early warning systems for debris flow hazards and provides a robust framework for hazard assessment in previously unexamined regions.