Multi-label Bird Species Classification Using Ensemble of Pre-trained Networks

Birds are crucial to the ecosystem's optimum functioning since they maintain a stable climate, convert toxins into nutrients, and oxygenate the atmosphere. Identifying multiple bird species from audio recordings is a significant challenge. We propose ensemble fusion models followed by a sequential aggregation strategy, where output features of different pre-trained models are aggregated and then fused, which can improve the performance of multi-label bird call classification. The pro-posed work is also implemented using chromagram features for a comparative study. A sliding window analysis of melspectrograms is carried out to learn the bird's vocalizations. The normalised sigmoid probabilities from different transfer learning models undergo fusion. Those candidates with the highest likelihood are considered to be birds that exist in the recording. The paper reports an F1-score of 0.60 for the 2-Classifier model, and it outperforms the baseline and chromagram-based models. Experimental evaluation is done using the bird's sound database, Xeno-canto.