BirdNova: A Web-Based Bird Species Identification Platform Using Separate Image and Audio Deep Learning Models
Year: 2026
Type: preprint
Abstract: Bird species identification supports biodiversity research and wildlife monitoring, but manual identification can be difficult when species have similar visual characteristics or vocalisations. This paper presents BirdNova, a web-based application that provides bird identification through two independent deep-learning workflows: image classification and audio classification. The image workflow resizes input photographs to 224 × 224 pixels and processes them using a convolutional neural network (CNN). The audio workflow converts recordings to mono audio at 22,050 Hz, standardises their duration to five seconds by trimming or padding, and generates Mel-spectrogram representations for a separate audio classifier. The two models are accessed through a common Flask-based interface but do not perform feature-level or decision-level fusion. The image classifier was evaluated on an independent test set of 1,753 images covering 25 bird species. It achieved 80.03% test accuracy, 81.49% weighted precision, 80.03% weighted recall, and 79.93% weighted F1-score. Validation accuracy was 95.96%, a difference of 15.93 percentage points from independent test accuracy. This gap indicates that generalisation requires further investigation. The current manuscript documents the audio workflow but does not report a complete quantitative evaluation of that module; therefore, no audio performance claim is made. BirdNova is presented as an application-level integration of separate image and audio classifiers, with its evaluation limitations explicitly identified.
Source: Research Square
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236608704
