Evaluation of Debris-Flow Vibration Signals Recorded at the Aiyuzi Stream in Shenmu Taiwan Using Machine Learning Methods

Abstract

In recent years, due to the frequent occurrence of extreme weather due to climate change, the Taiwan region has often suffered from landslides and debris flows in the past 20 years. This study used the ground surface vibration signals collected by the geophone from seven debris flow events in the Shenmu area. Data were processed to represent the time series of velocity and accumulated energy per second. Datasets were established for model training and validation. In this study, Support Vector Machine (SVM) and Random Forest (RF) algorithms were used for comparison. After analyzing the data through balance processing (Synthetic Minority Oversampling Technique, SMOTE), a signal model of debris flow was established. The research results showed that the models using SVM and RF training had good accuracy, recall, and AUC values when choosing input data average of every 6 s and the 10-s time interval within which the data were marked as the occurrence of debris flow. The performance of SVM was better than that of RF after validation. Through the aforementioned research, the vibration signals of debris flow can be regarded as a reference factor, and the model established by the SVM method had acceptable performance and can be used for early-warning of debris flow.

Keywords

• debris flow
• machine learning
• vibration signal
• Support Vector Machine
• Random Forest