Summary Distributed Acoustic Sensing (DAS) data often contain various types of noise, including random noise, coherent noise (e.g., coupling or linear noise), and common mode noise, which significantly degrade seismic signal quality. Conventional denoising methods struggle to effectively suppress diverse noise components while preserving important seismic signals. To address this issue, we propose a denoising self-supervised cascade network (DAS-DSCnet), a multi-stage neural network designed to progressively denoise DAS data without requiring external labels or synthetic training data generation. The network consists of three stages: Stage 1 targets random noise using a Noise2Noise-based approach; Stage 2 suppresses dataset-specific coherent noise using a denoising convolutional neural network (DnCNN)-based network trained with internally extracted noise patches; and Stage 3 predicts and removes common mode noise through trace shuffling and a Noise2Noise-based model. Training data for each stage are generated directly from the input DAS data by exploiting the data’s inherent characteristics, enabling efficient learning that reflects field-specific noise features. The model was evaluated using two distinct field DAS datasets with different noise patterns. The results demonstrate that DAS-DSCnet achieves superior noise suppression compared to conventional approaches, enhancing signal continuity while minimizing leakage. The denoising performance remains stable across different stacking configurations and hyperparameters, confirming the model’s robustness. Therefore, DAS-DSCnet offers a scalable and practical framework for improving seismic data quality in DAS applications, demonstrating the potential for fully automated, data-driven denoising in large-scale seismic monitoring.
Jun et al. (Thu,) studied this question.