Automatic Earthquake Detection in High‑Noise Regions of North Aceh Using Deep Learning

Abstract

The Lhokseumawe and Nisam segments of the Sumatran Fault cross North Aceh and generate shallow crustal earthquakes that threaten nearby communities and infrastructure. Monitoring this low-magnitude seismicity is difficult: most viable station sites lie close to roads, residential neighborhoods, or small industries, and STA/LTA detectors fail under that background noise. From January to December 2025, we ran a 66-station temporary network across North Aceh. We processed the continuous three-component records with a deep-learning phase picker within an open-source seismological machine-learning framework, followed by probabilistic phase association and analyst repicking in seismicity-management software. The final catalog contains 53 events, ML 0.02 to 3.69, with depths between 0 and 20.6 km, all in the upper crust. Epicenters track subparallel to the two fault segments, and 18 events cluster in September 2025. The workflow needs no model retraining and no specialized hardware. It worked here, and we expect it to work in other noisy, sparsely instrumented regions facing similar hazard pressure.