Using commercial machine-learning software to conduct bird species inventories

Abstract

Bird species inventories are an effective way to measure changes in an ecosystem. They can be conducted using automated sound recorders. Birds vocalizing in recordings can be found using manual detection (an observer finds and identifies the sounds), or automatic detection (machine-learning software finds and identifies the sounds, and an observer verifies). Here, I present a method of training software to identify regional birds and a method to efficiently verify its identifications. I then compared the number of species detected by manual and automatic detection using ~625 hours of field recordings/site over 29 sites. Automatic detection found ~45% more species/site (average: 28 vs. 19 species/site, P<0.01), but each method detected species that the other didn't. Automatic detection finds more species when effort constraints limit manual detection to a small portion of the audio, but for the most complete species list I recommend using a combination of manual and automatic detection.