Using image processing to detect and classify narrow-band cricket and frog calls

Abstract

An automatic call recognition (ACR) process is described that uses image processing techniques on spectrogram images to detect and classify constant-frequency cricket and frog calls recorded amidst a background of evening sounds found in a lowland Costa Rican rainforest. This process involves using image blur filters along with thresholding filters to isolate likely calling events. Features of these events, notably the event's central frequency, duration and bandwidth, along with the type of blur filter applied, are used with a Bayesian classifier to make identifications of the different calls. Of the 22 distinct sonotypes (calls presumed to be species-specific) recorded in the study site, 17 of them were recorded in high enough numbers to both train and test the classifier. The classifier approaches 100% true-positive accuracy for these 17 sonotypes, but also has a high false-negative rate (over 50% for 4 sonotypes). The very high true-positive accuracy of this process enables its use for monitoring singing crickets (and some frog species) in tropical forests.