An Optimized Random Forest Model for Debris-Flow Event Detection from Seismic Signals

Accurate identification of debris-flow events from seismic records is essential for developing high-resolution monitoring and early-warning systems. Here we develop an optimized Random Forest (RF) classifier designed to improve detection accuracy and, critically, to generalize across diverse geographic and environmental settings. We compile a global dataset of historical debris-flow events from 12 representative regions and construct an RF-based workflow that combines interpretable feature selection and automated model tuning. The Boruta algorithm is used to identify five informative predictors, improving interpretability while reducing redundancy in the feature set. In parallel, Bayesian optimization is employed to tune RF hyperparameters and enhance out-of-sample performance. We conduct three comparative experiments to quantify the contribution of each component. Results show that the combined Boruta–Bayesian RF consistently outperforms conventional RF approaches, achieving an accuracy of 96.25%, an F1 score of 0.9714, and an AUC of 0.9819. To further assess transferability, we apply the trained model to independent seismic data collected at Tianmo Gully in southeastern Tibet, China. The model successfully distinguishes debris-flow signals from background noise across the study period, demonstrating stable performance beyond the training regions. Overall, the proposed optimized RF framework offers an efficient, interpretable, and transferable solution for debris-flow detection using seismic signals, providing practical methodological support for the development of operational debris-flow early-warning systems.