Noise-resistant automatic seismic framework for monitoring rockslide slope

National seismic networks provide high-quality data for monitoring large-scale landslides within tens of kilometers, but are difficult to detect small-scale slope failures due to signal attenuation. Denser local networks can enhance monitoring capabilities focusing on prone slopes that potentially cause fatalities and economic losses, but being close to human settlements introduces significant noise interference. This study developed a noise-resistant automatic algorithm including three stages: detection, noise elimination, and classification, for a local seismic network deployed near villages to monitor an active rockslide slope. The main concept is to effectively filter diverse and abundant surrounding noise and purify the dataset before feeding it into the machine learning classifier. During a one-year examination period, 98.6% of non-target sections, including numerous calm ambiences and random noise, were filtered out by STA/LTA, signal-to-noise ratio, and cross-correlation in the detection and noise elimination stages. As a result, the remaining dataset primarily consisted of earthquake and rockslide signals in approximately a 5:1 ratio, with only a few vehicle passages and random noise. This denoised dataset was subsequently used to train a Random Forest classifier with two attribute clusters, achieving good recall rates of 78% for rockslides and 99% for earthquakes. However, approximately 20% of manually labeled rockslides were misclassified as earthquakes due to their overlapping attribute ranges that cause certain distinctive attributes to resemble earthquake characteristics. This study establishes an applicable framework for monitoring slope hazards near vulnerable villages, demonstrating that effective noise filtering can significantly improve the reliability of classification in seismic monitoring implemented in high-noise environments.