Comparative Evaluation of Deep Learning-Based Seismic Signal Denoising on an Edge Computing Platform Muhammad Azhar Raihan (1a), Ahmad Kadarisman (1,2) Djati Handoko (1), Martarizal (1,b*)
1 Department of Physics Faculty of Mathematics and Natural Sciences, Universitas Indonesia
2 Direktorat Instrumentasi dan Kalibrasi, BMKG, Jl. Angkasa I/2 Kemayoran - Jakarta, Indonesia
Email: a) muhammad.azhar51[at]ui.ac.id, b)martarizal[at]sci.ui.ac.id*
Urban earthquake monitoring is critically hindered by anthropogenic noise that spectrally overlaps with seismic signals, rendering conventional filtering methods insufficient. This study evaluates two deep-learning-based denoising methods, UrbanDenoiser and U-Net to identify the most effective approach for cleaning noisy seismic signals from two distinct stations in Indonesia while enabling on-site implementation on a low-power edge computing platform. UrbanDenoiser was trained on data from station TNGI and U-Net on data from station BKJI, after which both models were cross-evaluated on each other^s station data to assess generalization capability across different noise environments. Both models were compiled into Hailo Executable Format (HEF) and deployed on a Raspberry Pi 5 equipped with the Hailo-8L AI accelerator (13 TOPS) for real-time inference. Evaluation of UrbanDenoiser across 60,473 inference windows from station TNGI yielded a Pearson correlation of r = 0.9759 and an SNR of 6.89 dB, with a 5.00% anomaly rate at the 95th-percentile threshold, indicating strong denoising performance across low-to-moderate urban noise conditions (input SNR range: 3.7-11.4 dB). Initial results demonstrate that the Raspberry Pi 5 + Hailo-8L platform is capable of running real-time seismic denoising inference, validating its feasibility as an on-site solution for urban seismic monitoring networks in Indonesia.
