Nearest‐neighbor techniques for automated monitoring of nocturnal flight calls

Abstract

Flight‐calls are short vocalizations used primarily during nocturnal flight. Their observation provides a means for studying the timing, location, and composition of nocturnal migrations. As part of a three‐year study the Cornell Lab of Ornithology is using autonomous recorders to sample flight‐calls of nocturnal migrants in the Northeastern US. The resulting tens of thousands of hours of recording, make software‐assisted detection and classification essential. Automatic processing and human evaluation have yielded a considerable collection of flight‐calls, 5‐1000 examples for 100 species. The many‐class classification problem, along with the availability of many examples from most of the classes, and established (condensation and editing) and recent (metric‐trees) techniques used in prototype‐based classification nearest‐neighbor techniques, have led us to develop nearest‐neighbor based techniques and software to assist in the analysis of this data. We will present classification results on two examples, a set of thrushes (genera Catharus and Hylocichla, family Turdidae) consisting of six species and wood‐warblers (family Parulidae) consisting of 48 species. The thrush flight‐calls are visually and aurally distinctive, usually 100‐400 ms in duration and occupy and the 2‐5 kHz band. Wood‐warbler flight‐calls are typically between 20‐100 ms in duration and occupy the 5‐10 kHz, and are difficult for many experienced observers to distinguish.