A lightweight deep-learning detector for the real-time monitoring of the invasive frogs Rhinella marina and Polypedates leucomystax

Detecting invasive frog species swiftly and efficiently is vital for managing biological invasions. Passive acoustic monitoring, assisted by deep learning, has proven to be an efficient means of identifying anurans based on their unique mating calls. However, existing systems often require manual retrieval of recordings and inference processing on a dedicated workstation.

This delay in detection is particularly problematic on remote islands, where frequent travel is impractical. In response, researchers have developed a compact convolutional neural network capable of detecting two invasive frog species, Rhinella marina and Polypedates leucomystax, using low-power microcontrollers. This innovative approach allows for continuous, real-time monitoring of these invasive species.

The model demonstrated an impressive mean invasive-species F1 score of 0.87, with precision surpassing 0.88 for both species. This performance is on par with larger models, but with a significantly reduced size and computational requirements. Further validation through cross-domain evaluation confirmed the model's reliability in detecting these species at two out-of-training locations - Iriomote Island in Japan, and an independent dataset from Australia.

While some localized threshold calibration was necessary for Polypedates leucomystax at Iriomote, the model exhibited robust generalization across varying recording conditions and biogeographic contexts. These findings highlight the feasibility of effective acoustic detection of invasive species on self-contained, battery-powered devices.

By eliminating the need for server infrastructure, this technology enables autonomous early warning systems in remote protected areas, empowering conservation efforts to respond swiftly and effectively to invasive frog invasions.