Lahar Early Warning at Volcano Santiaguito, Guatemala: A Standard and a Deep Learning Approach

Abstract Santiaguito is an active volcano located in Guatemala. Among its many hazards, frequent lahar flows impact the infrastructure and the population living and working downstream among the wide river channel network. Lahars are flows comprising large amounts of water and pyroclastic debris which can rapidly initiate and reach speeds of tens of meters per second down the channels making them highly destructive. Lahars are commonplace in the long rainy season. Recently, the agency responsible for monitoring earthquake and volcanic activity in Guatemala (INSIVUMEH), has built and operates a dense real-time seismic network around the volcano. This study shows how seismic data can be used to provide early warning for lahars. We used the 2022 and 2023 seismic data to create a lahar catalog with a total of 50 lahars. We use the seismic records of lahars from 2022 (91 waveforms from 25 events) to train a Siamese Neural Network (SNN) that uses 5 min waveforms to produce a lahar detection. We use the same data to optimize a classical lahar detector that uses ratios between two different short-term/long-term average (STA/LTA) characteristic functions. We test and compare both approaches on continuous data from 2023. Both algorithms operate as single-station detectors but require detection from at least two stations. The tests show that the STA/LTA method misses fewer events, whereas the SNN provides earlier warnings. We show that the two methods could be combined for an optimal lahar early warning that could be faster than the individual methods and operate on a single station. In a move toward operational use, the STA/LTA method (implemented through SeisComP) is used to operate a real-time early warning messaging system at INSIVUMEH. Furthermore, we demonstrate the usefulness of SNNs for developing deep learning models when little training data is available.