Enhanced Identification of Wild Bird Species Using Ensemble Learning and Frequency Dynamic Convolution Methods

The paper presents the problem of correct species recognition of birds in ecological and conservation work using acoustic means. Classical CNNs do not work well in complex correlations between spectrograms and become unsuitable in application to outdoors with very high computational cost. The current work attempts to reduce such drawbacks by proposing a low computational-cost model, which can extract high-frequency features of bird calls in many frequency bands using Frequency Dynamic Convolution. Additionally, Coordinate Attention allows us to gain further global context information, hence more reliable predictions. The framework shows great improvements in terms of deep learning architectures, ranging from large ResNet50 to more lightweight MobileNet variants, especially for challenging acoustics. Furthermore, the ensemble model that uses ResNet50 and Random Forest enhances prediction performance, reaching an overall classification accuracy of 97.72%, which is substantially better than classical CNN models. It has shown the strengths of compact models for species identification, a compact framework deployable in the field setting and also very useful for contributing to population ecology and conservation biology.