Debris Flow Early Warning Using the Patch Fourier Transformer

Machine learning techniques have been extensively applied to identify debris flow events in seismic signals and develop debris-flow early warning systems. However, several challenges persist. Traditional models find it difficult to directly process the raw waveform signals and instead rely heavily on manual feature extraction, which may result in redundant or insufficient features, potentially resulting in unreasonable generalization bias. Meanwhile, deep learning approaches, particularly those based on convolutional neural networks (CNNs), require multilayer stacking for dimensionality reduction. This may cause overfitting. To address these challenges, this research introduces an enhanced model based on the Transformer architecture: the Patch Fourier Transformer (PFT).

The Patch attention mechanism allows the model to focus on key regions of the seismic waveform, highlighting areas of significant energy fluctuations that correspond to debris flow events. Utilizing the Patch attention mechanism, our model effectively captures energy fluctuations in the time-frequency domain and exhibits a high level of consistency with the spatio-temporal distribution of attention weights. By mapping the attention distribution to specific time-frequency regions, the model provides insight into the seismic signal components that most influence its decision-making process.

The model was evaluated using seismic data from 12 debris flow events in the Illgraben, a Swiss catchment. The PFT model achieved over 96% accuracy in waveform identification. Furthermore, the early warning system provided warning times ranging from 24 minutes to 2 hours without generating any false alarms. These results highlight the considerable potential and advantages of the PFT model for debris flow identification and early warning applications.