Abstract Birds are widely recognized as vital bioindicators of ecosystem health for high sensitivity to environmental changes, and automated bird call analysis using deep learning has demonstrated significant potential for large-scale ecological monitoring; however, its effectiveness is often limited by environmental noise in real-world acoustic recordings. This paper presents an experimental investigation and evaluation of NR-BirdNet, a noise-robust deep learning framework designed for bird call analysis under acoustically challenging conditions. The proposed NR-BirdNet framework employs a hybrid and joint use of Convolution Neural Network (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) architecture to jointly capture the spectral and temporal characteristics of bird vocalizations, demonstrating improved robustness over existing denoising frameworks and hybrid architectures. Two experimental scenarios are considered: (i) a baseline evaluation using noise-free recordings, achieving a classification accuracy of 98.75 % with limited misclassifications among acoustically similar species, and (ii) a noise-corrupted evaluation incorporating environmental interference, where accuracy decreases to 92.50 %, reflecting the impact of background noise, signal overlap, and low signal-to-noise ratios. NR-BirdNet enhances reliability in realistic, noise-prone environments. The results confirm that environmental noise significantly degrades classification performance and increases learning complexity. The contribution of this study is the development of NR-BirdNet, a noise-robust hybrid CNN–BiLSTM framework, with the advantage of robust, high-accuracy bird call classification in noisy real-world conditions.
Sharma et al. (Thu,) studied this question.