Mangrove Birdsong Classification with Unbalanced Data

Birds are an important component of mangrove ecosystems, reflecting the changes in mangrove ecological diversity. Most existing birdsong classification algorithms do not consider the data imbalance, resulting in low recognition rates for endangered bird species with limited samples, and the accuracies of these algorithms are also challenged by environmental noise. To address these issues, a birdsong classification algorithm based on deep learning is proposed. Band-pass filter and wavelet transform were applied to reduce the noise of the audio signal, which effectively improved the audio quality. The weighted-random-sampler was adapted to balance the dataset considering the scarcity of endangered bird samples. The EfficientNet model was chosen to focus on capturing characteristic differences in the frequency distribution and sound wave intensity of birdsong signals. Finally, UAR, confusion matrix and training time were used as evaluation metrics in the experiments, which made the evaluation more rigorous and scientific for imbalanced datasets. The experimental results show that the proposed algorithm achieves a highest recall rate of 96%, a UAR of 70%, and an excellent confusion matrix; in terms of training efficiency, compared to the comparison model, the EfficientNet model used in this algorithm takes 45.08% and 11.62% less training time on CPU and GPU, respectively.