Deep learning model fusion-based infrasound recognition of debris flows

Abstract

Infrasound generated during the formation and movement of debris flows exhibits long propagation distance and slow energy attenuation, rendering it ideal for remote monitoring of debris flows. However, accurately identifying the infrasound signals of debris flows is challenging because of the complexity of background noise interference and variations in signal characteristics linked to the event’s specific characteristics. In this study, the infrasound signal is preprocessed using high and low-pass filters, wavelet soft threshold denoising to mitigate noise interference, and then time-frequency transformed into a two-dimensional image, which is then input into a deep learning model fused with ResNet18 and Vision Transformer for training. The fusion model offers a potent feature extraction capability and the ability to generalize models, which leads to a better understanding of the details of the infrasound signals from debris flows. The experimental results show that the recognition accuracy of the proposed method is up to 88.60%, which is able to effectively predict and warn about the upcoming debris flow events.