A deep learning approach for infrasound signal detection in snow-covered terrain

Infrasound signal detection using arrays is an established method for detecting signals from a variety of cultural and geophysical sources. Array processing requires analysis from two or more sensors at an infrasound station. By employing array techniques, signals are isolated from various sources of noise by computing coherence between recorded signals across spatially distributed sensors at a station. Array processing requires multiple sensors and intra-array comparisons that potentially increases time required for event detection. This study presents a signal detection method utilizing deep learning to identify high-quality signals in infrasound waveforms with data from a single infrasound sensor, which might permit event discrimination without using arrays. The approach involves using audio signal processing techniques, applied to infrasound frequencies, to extract a feature space, which is then used to train a deep learning model. A 25-day continuous dataset featuring diverse infrasonic events that include potential avalanches, vehicle traffic, and explosions, collected from Little Cottonwood Canyon, Utah, USA was analyzed for this study. The data provide a convenient testbed for high-quality signal identification using single channel infrasound. The sensor network was motivated by the goal of identifying snow avalanches. Performance was assessed for an optimized model over a range of test data.