Item: AVALANCHE INFRASOUND SIGNAL CLASSIFICATION USING MACHINE LEARNING
-
-
Title: AVALANCHE INFRASOUND SIGNAL CLASSIFICATION USING MACHINE LEARNING
Proceedings: International Snow Science Workshop 2024, Tromsø, Norway
Authors:
- Jeffrey B Johnson [ Boise State University ]
- Scott Havens [ Snowbound Solutions ]
- Clark Corey [ Snowbound Solutions ]
- Evi Ofekeze [ Boise State University ]
- Skyler Chase [ Boise State University ]
- Owen Walsh [ Boise State University ]
- Jerry Mock [ Boise State University ]
- Hans Peter Marshall [ Boise State University ]
Date: 2024-09-23
Abstract: Infrasound monitoring is an established technology for detecting and locating snow avalanches using array processing techniques. Infrasound arrays are collections of three (or more) closely-spaced (10-20 m) sensors deployed in network configurations typically used for operational avalanche monitoring systems. Multiple arrays are collectively used to both identify avalanche signals and map their progression over time. Properly tuned networks of arrays, such as those operating in Little Cottonwood Canyon (Utah, USA), can be used to reliably identify avalanches with destructive indices as low as D2 and occurring at distances greater than 1 km from the sensors. Arrays are used to identify signals as avalanches and also minimize false positive detections from other signals like munitions, vehicles, aircraft, and earthquakes. An infrasound network dataset from Little Cottonwood Canyon spans three years of continuous records, starting in 2021-2023, and contains more than a thousand avalanche signals. A priority moving forward is to robustly identify avalanches using data from a single infrasound sensor. We are using the rich dataset from Little Cottonwood Canyon, and tens of thousands of avalanche signal windows, to train automated classifiers using speech recognition methodologies. Our efforts are to identify the smallest possible feature variable space to reliably identify an avalanche signal and discard non-interesting signals or noise. We show that the use of machine learning techniques to classify avalanches may permit more rapid (real-time) identification of avalanches without the need to integrate multiple channels of data. Although a single sensor is not capable of avalanche source localization it can be valuable for low-cost installations and mobile avalanche alert detection systems.
Object ID: ISSW2024_O8.2.pdf
Language of Article: English
Presenter(s): Jeffrey Johnson
Keywords: infrasound, snow avalanche dynamics, avalanche hazard mitigation, machine learning
Page Number(s): 997 - 1004
-