SeisMamba: Selective state space models for DAS seismic denoising and phase picking

초록

Distributed acoustic sensing (DAS) produces dense, high-rate channel–time arrays that challenge deep-learning models for continuous monitoring. This study presents SeisMamba, a dual-axis bidirectional selective state-space model that scans DAS features along channel and time axes within a U-Net-style encoder–decoder. SeisMamba is compared with parameter-matched ResUNet and SwinDAS under shared protocols for supervised synthetic denoising and an exploratory channel-mean P/S phase-picking task constructed from the available DeepSubDAS files. An unchanged checkpoint is also examined qualitatively on Utah FORGE field data; separate analyses assess scalability, loss sensitivity, and architectural ablations. For denoising, ResUNet gives the highest SNR (29.19 ± 0.12 dB), while SeisMamba reaches 26.98 ± 0.05 dB with high structural similarity (SSIM 0.9954 ± 0.0001) and is strongest on colored noise. For the within-release DeepSubDAS task, SeisMamba gives the strongest parameter-matched weighted-CE result: 73.03±0.33% accuracy, F1P=0.509±0.008, and the lowest peak-time P-arrival error (0.286 ± 0.001 s). Loss-sensitivity results are reported separately from the matched architecture ranking. In the full-record harness at Nc=256, Nt=65,536, SeisMamba runs in 181.7 ms, SwinDAS in 289.1 ms, and ResUNet is OOM on the 24 GB benchmark GPU unless chunked (575.6 ms). Seed-42 ablation diagnostics indicate that both scan axes matter for denoising, whereas the time-axis branch is especially important for the channel-mean picking target. Within the tested protocols, SeisMamba provides a competitive accuracy–scalability trade-off for large DAS processing rather than a universal replacement for convolutional denoisers. © 2026 Elsevier B.V.

키워드