Automatic Capture and Classification of Frog Calls

Abstract

Global frog populations are threatened by an increasing number of environmental threats such as habitat loss, disease, and pollution. Traditionally, in-person acoustic surveys of frogs have measured population loss and conservation outcomes among these visually cryptic species. However, these methods rely heavily on trained individuals and time-consuming field work. We propose an end-to-end workflow for the automatic recording, presence-absence identification, and web page visualization of frog calls by their species. The workflow encompasses recording of frog calls via custom Raspberry Pis, data-pushing to Jetstream cloud computer, and species classification by three different machine learning models: Random Forest, Convolutional Neural Network, and Recursive Neural Network.