Executive Summary

The presence or absence of certain animal species in an environment indicates the health of the ecosystem. It is one measurement, among many others, of how human activities impact our natural environment. With dense forest canopies making visual identification difficult, we looked into the capabilities of audio identification to determine biodiversity: how can we determine the biodiversity of a rainforest by classifying the sounds of the species heard in forest audio recordings?

Using data from an ongoing Kaggle competition to identify 24 bird and frog species from audio files, we cleaned and processed data using AWS architecture and the Dask library to handle a large volume of data (57.32 GB of audio files) — with S3 for storage and an EC2 instance running Dask and PySpark. From the audio files, we generated mel spectrogram images per animal species and used them as input to our deep learning model.

We achieved 85% accuracy using ResNet-50 to classify and identify species from mel-spectrograms, averaged across five iterations of training with a stratified 5-fold train-test split. The resulting model was then used to infer species from raw audio files.

Full text article and source codes can be provided upon request.