Thu, Aug 18, 2022: 5:00 PM-6:30 PM
ESA Exhibit Hall
Background/Question/Methods: Amid recent declines in biodiversity and ecosystem health, autonomous acoustic recorders have become popular as a low-disturbance, large-scale option for monitoring sound-producing animals. One specialized use of this data is acoustic localization, which allows researchers to track animal movements, infer territory boundaries, and identify individuals. One of the most popular algorithms for acoustic localization comes from the R package SoundFinder, which uses a GPS algorithm as an analog for calculating the locations of sound sources based on observed time delays. An alternative model, proposed by Gillette and Silverman, restates the problem as a direct linear system. There has been little evaluation of the ability of these two methods to localize sound under real-world conditions and no direct comparison of their performance. Here, we generate simulated data mimicking real-world conditions and compare the performance of these algorithms. We tested each algorithm using a variety of configurations, both with and without error in recorder location and time measurement. We compared these algorithms using error in the pseudorange, the expected time delays based on the estimated source location compared against the observed delays.
Results/Conclusions: We computed results for sound sources in three dimensions, ranging from 0 and 100 meters above the ground on a 40x40 m square, with an array of 9 microphones evenly spaced throughout. Without errors in recorder location or timing, we found negligible differences between the algorithms, with all errors below 10-9 m. When we introduced realistic location and timing errors, SoundFinder maintained errors within 0.5m for ground-level sound sources, with most errors below 1m until the source was 50m or higher. The Gillette & Silverman model, however, had errors above 1m. Errors were approximately normally distributed. In general, large pseudorange errors appeared to result from errors in height estimates, with errors in the x,y-plane being comparable in magnitude and more consistently distributed than SoundFinder. Overall, we found SoundFinder to perform better in three-dimensional localization and the Gillette & Silverman algorithm to be better for ground-level sound sources. However, the performance of the two methods was closely comparable for the two-dimensional use case. We note that the Gillette & Silverman is faster and so may be better suited for the analysis of larger datasets, although in practice, it requires one additional recorder compared to Soundfinder.
Results/Conclusions: We computed results for sound sources in three dimensions, ranging from 0 and 100 meters above the ground on a 40x40 m square, with an array of 9 microphones evenly spaced throughout. Without errors in recorder location or timing, we found negligible differences between the algorithms, with all errors below 10-9 m. When we introduced realistic location and timing errors, SoundFinder maintained errors within 0.5m for ground-level sound sources, with most errors below 1m until the source was 50m or higher. The Gillette & Silverman model, however, had errors above 1m. Errors were approximately normally distributed. In general, large pseudorange errors appeared to result from errors in height estimates, with errors in the x,y-plane being comparable in magnitude and more consistently distributed than SoundFinder. Overall, we found SoundFinder to perform better in three-dimensional localization and the Gillette & Silverman algorithm to be better for ground-level sound sources. However, the performance of the two methods was closely comparable for the two-dimensional use case. We note that the Gillette & Silverman is faster and so may be better suited for the analysis of larger datasets, although in practice, it requires one additional recorder compared to Soundfinder.