Artificial Neural Networks to Classify Bird Acoustic Data from Costa Rica

Abstract

Biodiversity assessment plays an important role in allowing scientists to understand the structure and composition of biological communities. There are many approaches to assessing biodiversity– one such approach being the use of passive recorders to collect acoustic data from a region, and then analyzing the acoustic data to determine which species are present. However, processing large amounts of acoustic data using traditional methods–identifying species from acoustic data manually–is slow and laborious, and effective automatic classifiers are desired. In this project, deep neural networks are used to approach classifying a dataset of 28 bird species from El Area de Conservacion Guanacaste, a region in Costa Rica. The experiments run in this project attempt to determine the ideal parameters for creating, training, and testing deep neural networks on this specific dataset. The effectiveness of the neural networks on this dataset will be gauged and also be compared to the effectiveness of similar neural networks on a separate, larger dataset that experimented with by Hendrik Vincent Koops–who developed the codebase that was used in this project.