Evaluation of Deep Learning-Based Monitoring of Frog Reproductive Phenology

Abstract

To evaluate the utility of a deep-learning approach for monitoring amphibian reproduction, we examined the classification accuracy of a trained model and tested correlations between calling intensity and frog abundance. Field recording and count surveys were conducted at two sites in Kyoto City, Japan. A convolutional neural network (CNN) model was trained to classify the calls of five anuran species. The model achieved 91-100% precision and 75-98% recall per species, with relatively lower performance on less abundant species. Computational experiments investigating the effects of the number and seasonality of the training samples showed that models trained on larger datasets from broader recording seasons performed better. Calling activity was high when males were abundant (Pearson's r 5 0.45-0.66), although correlations between the calling activity and the number of pairs in amplexus were generally weaker. Our results suggest that deep learning is an effective tool for reconstructing the reproductive phenology of male anurans from field recordings. However, caution is required when applying to rare species and when inferring female reproductive activity.

Deep learning-based detector of invasive alien frogs,Polypedates leucomystax and Rhinella marina, on an island at invasion front

“…We applied the trained model to the entire recordings collected on Ishigaki Island (sites 1-3) to infer the breeding phenology of P. leucomystax and R. marina. Reproductive activity was quantified by counting the number of detections (Kimura and Sota 2023). Give that each recording was 15 minutes in length and the BirdNET analysis unit was 3 seconds, the number of detections per recording ranged from 0 (no detection) to 300 (continuous detections throughout the recording).…”