Machine learning an audio taxonomy: Quantifying biodiversity and habitat recovery through rainforest audio recordings

Abstract

We present a set of tools for semi-supervised classification of ecosystem health in Meso-American tropical dry forest, one of the most highly endangered habitats on Earth. Audio recordings were collected from 15-year-old, 30-year-old and old growth tropical dry forest plots in the Guanacaste Conservation Area, Costa Rica, on both nutrient rich and nutrient poor soils. The goals of this project were to classify the overall health of the regenerating forests using markers of biodiversity. Semi-supervised machine learning and digital signal processing techniques were explored and tested for their ability to detect species and events in the audio recordings. Furthermore, multi-recorder setups within the same vicinity were able to improve detection rates and accuracy by enabling localization of audio events. Variations in species' and rainforest ambient noise detection rates over time were hypothesized to correlate to biodiversity and hence the health of the rainforest. By comparing levels of biodiversity measured in this manner between old growth and young dry forest plots, we hope to determine the effectiveness of reforestation techniques and identify key environmental factors shaping the recovery of forest ecosystems.