BirdNET: applications, performance, pitfalls and future opportunities

Abstract

Automated recognition software is paramount for effective passive acoustic monitoring. BirdNET is a free and recently developed bird sound recognizer. I performed a literature review to evaluate the current applications and performance of BirdNET, which is growing in popularity but has been subject to few assessments, and to provide recommendations for future studies using BirdNET. Prior research has employed BirdNET for a wide range of purposes but few studies have linked BirdNET detections to ecological processes or real-world monitoring schemes. Among evaluated studies, average precision (% detections correctly identified) usually ranged around 72-85%, and recall rate (% target species vocalizations detected) ranged around 33-84%. Some studies did not assess BirdNET performance, which hampers the interpretation of the ecological results and may provide poorly informed decisions. Recommendations on how to evaluate Bird-NET efficiency are provided. The impact of the confidence score threshold, a user-selected parameter as the minimum score for detections reported, on BirdNET output although variable among species is consistent. The use of high confidence score thresholds increases the percentage of detections correctly classified but lowers the proportion of calls and bird species detected. The selection of an optimal score may depend on the priorities of the user and research goals. BirdNET is a great tool for automated bird monitoring but it should be used with caution due to inherent challenges for automated bird identification. The continued refinement of BirdNET suggests further improvements in the coming years.

A First Assessment of Birdnet Performance at Varying Distances: A Playback Experiment

“…The number of misclassifications varied over distances and did not follow a linear pattern. This study provides valuable information that may contribute to improved surveys and for expanding the use of BirdNET for surveying bird communities using passive acoustic monitoring.—Pérez-Granados, C. (2023). A first assessment of BirdNET performance at varying distances: a playback experiment.…”

“…El número de clasificaciones erróneas varió con las distancias y no siguió un patrón lineal. Ese estudio proporciona información valiosa que puede contribuir a mejorar futuros muestreos y a expandir el uso de BirdNET para censar comunidades de aves usando monitoreo acústico pasivo.—Pérez-Granados, C. (2023). Un primer análisis del rendimiento de BirdNET a distancias variables: un experimento con playback .…”

“…Zhao et al ., 2017). Among recently developed algorithms for automated bird identification on sound recordings is BirdNET (reviewed by Pérez-Granados, 2023). BirdNET has already proved its utility for bird monitoring purposes and as a reliable research resource (Cole et al ., 2022; Manzano-Rubio et al ., 2022; Wood et al ., 2022).…”

“…Kahl et al . (2021) assessed BirdNET's ability (precision) to correctly classify bird vocalisations of 984 species in focal recordings and found an overall precision of 0.791 (see an extended review in Pérez-Granados, 2023). However, bird vocalisations, as with any other acoustic signal, attenuate over distance and therefore inevitably degrade progressively during propagation from the signaller to the receiver (Naguib & Wiley, 2001).…”

“…Likewise, signal degradation over distance also has an impact on the ability of automated signal recognition software, such as BirdNET, to detect and correctly classify bird vocalisations (see Knight et al ., 2020 for a relationship between recognizer score and the energy of acoustic signals). Therefore, it is expected that BirdNET performance will decrease with distance, but to date, there is no information available about the impact of bird distance from the recorder on BirdNET performance (Pérez-Granados, 2023).…”