AI System Identifies Bat Species Beyond Trained Taxonomy

Reliable monitoring of bat populations is crucial for conservation efforts, but their nocturnal and cryptic nature makes automated identification challenging. Echolocation calls vary significantly with behavior and environment, often overlapping between species, complicating classification. This research introduces ChiroEcho, a novel deep learning framework designed to improve automated bat vocalization classification. ChiroEcho operates by jointly predicting both the species and genus of a bat from its calls. Crucially, it then combines these genus-level predictions with geographic species distribution data during inference. This innovative approach allows the framework to resolve and identify specific species that were not explicitly part of its initial learned taxonomy, particularly when only one species of a predicted genus is known to occur in a given region. Evaluated on recordings spanning 35 European bat species, the system demonstrated that geographic information can extend, rather than merely constrain, a classifier's effective taxonomic coverage. This method increased operational coverage from 35 to 41 of the 48 native European bat species, boosting coverage from 73% to 85%. This represents the broadest operational coverage reported for automated European bat classification to date, offering a significant advancement for ecological monitoring and conservation.