Lightweight acoustic classification for cane-toad monitoring

We propose a light weight algorithm to classify cane-toads, a non-native invasive amphibian species in Australia as well as other native frog species, based on their vocalizations using sharply resource-constrained acoustic sensors. The goal is to enable fast in-network frog classification at the resource-constrained sensors so as to minimize energy consumption of the sensor network by reducing the amount of data transmitted to a central server. Each sensor randomly and independently samples a signal at a sub-Nyquist rate. The vocalization envelopes are extracted and matched with the original signal envelopes to find the best match. The computational complexity of the algorithm is O(n). It also requires less than 2KB of data memory. Our experiments on frog vocalizations show that our approach performs well, providing an accuracy of 90% and a miss rate of less than 5%.