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

This study aimed to develop and evaluate a deep learning-based detector for Southeast Asian treefrog (Polypedates leucomystax) and cane toad (Rhinella marina) on Iriomote Island, a Natural World Heritage Site located 30 km from established populations of these alien species on the nearby Ishigaki Island. Although a deep learning model typically requires local training data to be most accurate, the alien frogs have been eradicated from Iriomote Island, making such data unavailable. To address the data gap, we first trained the BirdNET model with acoustic data collected on Ishigaki Island, where these species were common, as well as native frog calls from Iriomote Island. Next, we evaluated model performance using (1) wild sounds on Ishigaki Island, and (2) sounds obtained by playing back the calls of the alien species on Iriomote Island. Model precision and recall for wild sounds were both 0.987 for P. leucomystax and 1 for R. marina. For playback sounds, model recall values decreased (0.629 for P. leucomystax and 0.906 for R. marina), while precisions remained nearly identical (1 for both P. leucomystax and R. marina). Despite the lower recall particularly for P. leucomystax, playback survey dates were mostly identifiable from the high number of detections. These results suggest that data from Ishigaki Island enabled training a model with adequate, though not complete, generalization across invasion front.