This study compares three automatic methods for detecting and locating induced earthquakes in Lu County, Sichuan, where events are shallow, weak, and frequent due to human activities. These characteristics make detection challenging due to overlapping signals and low energy.

The first method, Multi-Station STA/LTA with absolute location, efficiently identifies over 33,000 events, but is sensitive to noise. The second, S-SNAP, uses deep learning and waveform stacking to detect 8,690 high-quality events, offering better accuracy and robustness. The third, SUGAR, applies AI-based image recognition on brightness videos from waveform stacks to detect events, showing strong adaptability to dense local seismicity.

Overall, the study shows that combining phase and energy features, deep learning, and physical models can improve earthquake detection and support hazard assessment.