Publication

Abstract : Distributed acoustic sensing (DAS) enables dense seismic monitoring; however, event detection is challenged by limited labeled data and noise. This letter introduces a semisupervised framework based on multiloss temporal convolutional network, where hybrid masking and multiobjective loss enhance signal-to-noise ratio (SNR) and improve label efficiency. The method achieves 98.75% classification accuracy and 36.55 dB SNR, significantly surpassing semisupervised baselines. To further illustrate adaptability, transfer learning experiment on an external dataset confirms the model’s generalization capability. This label-efficient method advances scalable and robust DAS-based seismic event detection with minimal labeled data requirements.