Wed, Aug 04, 2021:On Demand
Background/Question/Methods
Acoustic monitoring is an effective strategy for gathering ecological data because many animals produce species-specific sounds. Recently, the availability of low-cost autonomous recording units and the development of automated acoustic recognition technology have enabled more extensive acoustic surveys than is possible with human point counts. Automated acoustic monitoring is particularly valuable for species that may not produce sound during a 10 minute point count, such as Ruffed Grouse (Bonasa umbellus). However, existing automated recognition methods for birds are not designed to identify the low-frequency, accelerating “drumming” signals produced by Ruffed Grouse. Therefore, we develop an automated method to detect Ruffed Grouse vocalizations in audio recordings. Our approach involves applying a non-linear transformation to the time domain of an audio signal, which stretches the accelerated drumming of Ruffed Grouse into a periodic signal which can be detected with an existing pulse-repetition detector. To test the performance of our detector, we performed point counts and collected audio recordings from 28 sites with deer exclosures and 27 control sites with no exclosure. We compared measurements of site occupancy across the treatments based on the automated method versus point count data.
Results/Conclusions By manually reviewing the 10 top scoring 8-second files at each of 57 monitoring sites, we confirmed the presence of Ruffed Grouse at 10 unique sites. In comparison, human point counts which visited each site twice during the breeding season detected Ruffed Grouse at 9 unique sites. Only 5 sites had detections both from the automated approach and point counts, meaning that the automated approach located Ruffed Grouse at 5 additional points where it was not detected during point counts. This demonstrates that automated recognition of data from autonomous recording units can augment species-presence information from point count data. In either detection method, the results did not provide evidence to reject the null hypothesis that occupancy and deer exclosure treatment are independent. We note that human point count observations of Ruffed Grouse at sites where it was not detected by the automated method always had detection distances of over 50 meters. This suggests that the automated recognition model is comparable to human point counts for close-by drumming, but could be improved by increasing its sensitivity to distant drumming. Future work could use this automated recognition method to monitor the response of Ruffed Grouse populations to management practices across large landscapes.
Results/Conclusions By manually reviewing the 10 top scoring 8-second files at each of 57 monitoring sites, we confirmed the presence of Ruffed Grouse at 10 unique sites. In comparison, human point counts which visited each site twice during the breeding season detected Ruffed Grouse at 9 unique sites. Only 5 sites had detections both from the automated approach and point counts, meaning that the automated approach located Ruffed Grouse at 5 additional points where it was not detected during point counts. This demonstrates that automated recognition of data from autonomous recording units can augment species-presence information from point count data. In either detection method, the results did not provide evidence to reject the null hypothesis that occupancy and deer exclosure treatment are independent. We note that human point count observations of Ruffed Grouse at sites where it was not detected by the automated method always had detection distances of over 50 meters. This suggests that the automated recognition model is comparable to human point counts for close-by drumming, but could be improved by increasing its sensitivity to distant drumming. Future work could use this automated recognition method to monitor the response of Ruffed Grouse populations to management practices across large landscapes.