Assistant Professor University of Pittsburgh Pittsburgh, Pennsylvania, United States
Abstract: Rana sierrae (the Sierra Nevada yellow-legged frog) is an endangered resident of high-elevation aquatic habitats in the Sierra Nevada mountains. Though the frog has received research attention for decades, gaps in the understanding of its life history have proven difficult to fill in due to the frogs’ remote habitat. Sampling occurs almost exclusively in summer during daylight hours, which limits knowledge of the species’ behaviors outside of these time windows. Here, we fill some of these gaps by recording wild R. sierrae with autonomous hydrophone recorders. Acoustic monitoring allows us to sample the lakes at times of the day and season when humans cannot be present in the habitat. Acoustic monitoring is a particularly effective tool for monitoring frogs because most frogs produce species-specific vocalizations that are associated with breeding activity. We deployed hydrophones in a lake to monitor R. sierrae activity over the course of a breeding season. We first manually reviewed a subset of the data to characterize the diversity of sounds in the aquatic soundscape. We then developed an automated recognition algorithm to detect R. sierrae vocalizations. Using the automated recognizer, we measured vocal activity patterns across 24-hour cycles and across the duration of the breeding season.
Inspecting a subset of the audio dataset revealed undocumented diversity of R. sierrae vocalizations, including 3 previously undescribed call types. Our machine learning classifier accurately detected R. sierrae vocalizations in a validation set (area under ROC 0.95). Using the classifier, we measured diel and seasonal patterns of vocal activity at the study lake. Contrary to prior expectations, we found that vocal activity was higher at night than during the day. Seasonal activity patterns exhibited strong spatiotemporal heterogeneity, whereby the dates containing vocal activity depended on the location of the recorder. These findings suggest that the addition of acoustic monitoring to R. sierrae survey efforts could provide valuable information about breeding activity that is not captured by visual surveys alone. We share audio recordings and analysis scripts to support further study of R. sierrae vocalizations and aquatic soundscapes more broadly. Our study demonstrates that combining acoustic monitoring with machine learning can provide unique insight into ecological processes that are otherwise difficult to observe.
