Isolating Vocalizations Beats Denoising: Spectrogram Granularity and Preprocessing in Ecoacoustic Classification

Passive Acoustic Monitoring (PAM) offers invaluable ecological insights, but classifying animal vocalizations via deep learning is severely hampered by pervasive noise, overlapping sounds, and critical mislabeling in manually annotated training data. This study systematically evaluates the impact of spectrogram generation strategies, diverse background noise reduction techniques, and a novel approach to identify and remove potentially mislabeled training instances on ecoacoustic classification performance. Experiments with ResNet and ViT architectures reveal that spectrograms with tightly bounded timefrequency windows around vocalizations dramatically improve classification. Surprisingly, both background noise removal and the exclusion of potentially noisy training labels yielded only a limited impact. These findings critically redirect future ecoacoustic efforts towards optimizing vocalization isolation.