← Back to Sessions

Identification of Lahar Signals: A Supervised Learning Model Applied to Monitoring Data of Volcan De Fuego, Guatemala

Description: 

The rise in popularity of machine learning models has provided improved tools for identifying and classifying seismic signals of different natures. Here, we showcase a novel approach that uses a supervised K-nearest neighbor model to detect tremor signals associated with rain-triggered lahars at Volcan de Fuego, Guatemala. For this, we used a set of three permanent and five temporary broadband seismometers installed between ~10–350 m from the thalweg of two active lahar channels around Fuego. The recorded signals display durations between 0.5–6 hours, high signal-to-noise ratios, and overall high-frequency activity with lower frequencies active during denser, high-energy flow portions. We established a binary classifier trained with features that describe the seismic record in the time and frequency domains, split into 10-minute overlapping windows. The best features for this method were chosen based on the model's performance. They included measures of the signal's amplitude, frequency content, and statistical functions of the prior, such as kurtosis, skewness, and entropy. A training set of 5 confirmed lahars and equally long background noise was enough to achieve greater than 90% precision at stations with a comparatively low sampling rate (50 sps). Cross-validation shows that two of our datasets with higher sampling rates (200 sps) require a higher number of neighbors to achieve similar precision results: n<10 at 50 sps vs. n>50 at 200 sps. A post-processing step converts the resulting discrete sample predictions into prediction intervals and discriminates false positive results. This model detected 161 of 172 observed lahars during the 2018-2022 period at the longest-running seismic station in one of Volcan de Fuego's active channels. While these detections did not always match the total duration of observed lahars, they still correctly classified ~95% of the signal portion attributed to lahar activity during this period. We highlight the potential of our method for real-time monitoring applications, providing a valuable tool for the timely identification of lahar signals.

Session: Detecting, Characterizing and Monitoring Mass Movements - I

Type: Oral

Date: 5/2/2024

Presentation Time: 08:45 AM (local time)

Presenting Author: Gustavo

Student Presenter: Yes

Invited Presentation: Yes

Authors