Toward a Neural Network-Based Approach for Improved Atmospheric Infrasound Localization

Accurate infrasound localization represents a challenging problem in acoustics that yields many important applications. Current methods for distance and altitudinal localization have only been evaluated in limited short-range scenarios and fail to take full advantage of real-time atmospheric surroundings to maximize accuracy. We develop a neural network architecture that uses both the local acoustic pressure and atmospheric data to produce accurate localization predictions including the lateral distance and altitude of an infrasonic source, using only a singular-sensor arrangement for a more flexible and practical implementation that can be transferred to real-time aerial detection. To improve generalizability, we train our neural network in a variety of atmospheric conditions and conduct validation testing on samples generated from designated hold-out atmospheric profiles. We generated a comprehensive infrasonic raypath dataset using an experimentally validated infrasound raytracing tool, a technique inspired by other fields where controlled data is difficult to obtain in real life. A preliminary comparison with existing infrasound localization methods in similar conditions demonstrated that our neural network-based approach approximately halved both median distance and altitudal localization errors. These promising results highlight the potential of our neural network-based approach to significantly enhance infrasound localization accuracy across diverse atmospheric conditions. Future steps would involve more thorough validation and benchmarking of the model in real-world scenarios to detect any trends in localization error that were not captured in this work.