Correlation of biological and acoustical measurements of the endangered Golden Cheeked Warbler with applications of machine learning

Abstract

Acoustic measurements on Setophaga chrysoparia (Golden-cheeked warbler) have been made within the 40,000 acre Balcones Canyonland Preserve (BCP) in Austin, TX, where the long-term goal is to understand the potential effects of anthropogenic noise on breeding success. The anthropogenic sources of noise include road traffic, jet aircraft, helicopters, and urban development and utilization. During the breeding season from March through May, acoustical recordings were made in dense Ash Juniper forests for 12 hours per day at 12 locations within the BCP. The evolution of the number of songs of Type A and B and their variants are signatures of breeding success. The study attempts to apply principle component analysis for feature selection input into a feedforward neural network classifier of the A and B signatures, their variants, and the ability to automate the identification of individual birds. A feedforward neural network attempts to parameterize a mapping y = f(x,θ) for input features and category pairs of the A and B signatures and parameterization θ. Explored is the sensitivity of the number of hidden layers required to learn the optimal parameterization θ approximating f. [Work supported by City of Austin.]