PINN-AU-Net: physics-informed attention U-Net for spatiotemporal segmentation of DAS seismic wavefields

To address the challenges of low signal-to-noise ratio (SNR) and data processing difficulties in the application of Distributed Acoustic Sensing (DAS) for seismic monitoring, this paper proposes a two-dimensional wavefield segmentation method integrating physical information and attention mechanisms. Unlike traditional onedimensional phase picking, this method treats DAS spatiotemporal data as images, aiming to achieve pixel-level semantic segmentation of different phases (such as P-waves and S-waves), with effectiveness verification primarily focusing on P-waves. We designed an improved U-Net architecture by: 1) Introducing an attention mechanism to enhance the model’s sensitivity to features of different phases; and 2) Adopting the Physics-Informed Neural Network (PINN) framework, incorporating the wave equation residual as a penalty term into the loss function to guide the model in generating segmentation results consistent with physical laws. Experiments were conducted using the distributed fiber optic sensing observation dataset from the seabed off Cape Muroto in Japan. The results show that compared to the baseline U-Net, our model performs better on segmentation metrics such as mIoU, and the physical residual is reduced by an order of magnitude. This strongly verifies the effectiveness and robustness of this hybrid-driven method in breaking through the generalization bottleneck of pure data models and ensuring the physical consistency of the results.