Investigating misclassification in bird sounds: The adverse effect of label noise on BirdNET

Understanding the causes of misclassification in AI bird sound recognition is important for improving the robustness and reliability of bioacoustic monitoring systems. Birds are ideal bioindicators because of their ecological value, abundance, and unique vocalizations. Advances in deep learning have enabled tools such as BirdNET, using EfficientNet as its foundation, to classify bird species based on vocalizations. However, misclassification remains a challenge; often, it results from model limitations, dataset imbalances, signal noise, or incorrect labels. This study investigates the causes of misclassification using a ResNet-based model for comparative analysis alongside BirdNET. We employ Explainable AI (XAI) techniques, such as LIME, to interpret model decision-making and low-dimensional mapping (t-SNE and PCA) to visualize latent space patterns. Dataset analyses reveal significant class imbalances and the presence of label noise, particularly in challenging species, where there is a higher level of false positives in the prediction. Additionally, we systematically analyze the impact of label noise on model performance using controlled experiments on two bird classes with different levels of label noise. Our results predictably demonstrate that reducing label noise in the training sets can improve the F1 scores of the model. These findings emphasize the importance of robust dataset preparation and noise management for bioacoustic monitoring technologies. • Identifies key causes of misclassification in bird sound recognition, including model limitations, dataset issues, and label noise, in BirdNET. • Uses dimensionality reduction (t-SNE, PCA) to visualize the latent space to emphasize the latent space entanglement and infeasibility of determining a decision boundary for systematic label noise removal. • Investigates the class imbalance, SNR, and cross-correlation of the dataset as the sources of misclassification in BirdNET. • Demonstrates the critical role of label noise, showing its effect on improving F1 scores.