Automatic identification of landslide events from continuous seismic recordings using particle swarm optimization-optimized random forest

The absence of on-site monitoring measures or witnesses has delayed catastrophic landslide relief in remote areas. Recent advancements have revealed that existing seismic networks can record seismic signals generated by large landslides, which holds potential for predicting landslides. Nevertheless, automatically recognizing these landslide events from enormous seismic recordings is still challenging. Here, we develop an automatic recognition model using a particle swarm optimization-optimized random forest algorithm that automatically scans continuous seismic recordings within a 200 s window and identifies landslide events within 1.013 s, with an accuracy exceeding 98%. Our results demonstrate that stations within an epicentral distance of 250 km are more sensitive in recording landslide-generated seismic signals. Our work delivers the potential for a timely disaster response to catastrophic landslides in remote mountainous areas. Key findings also suggest the model’s potential to capture events preceding the main landslide event.