Computationally-Efficient Call Classification of New Zealand Birds Using Texture-Based Features

Bird call classification - both manual and automatic - commonly relies on spectrograms, visual time-frequency representations of the signal power distribution of the audio signal. These spectrogram “images” can be processed using deep learning approaches like convolutional neural networks (CNNs) for call identification. However, such methods typically require significant computational resources. Using labelled data from the calls of three New Zealand bird species, we compare this approach with computationally lightweight approaches that extract texture-based features from both spectrograms and audio waveforms, specifically local binary pattern (LBP) and local phase quantisation (LPQ) features, paired with classical machine learning classifiers. We find that LBP-based features outperform other feature-classifier combinations and achieve superior accuracy compared to the standard CNN approach. Given their significantly lower computational cost, texture-based techniques prove particularly well-suited for resource-constrained environments. Experimental results demonstrate that lightweight feature extraction methods can serve as efficient and effective alternatives to deep learning approaches for automated bird call classification, offering practical advantages for deployment in field conditions or embedded systems.