Wavelet Scattering Operators for Multiscale Processes: The Case Study of Marine Mammal Vocalizations

Marine mammals vocalization pose challenges in understanding animal communication due to signal diversity and environmental factors. Researchers leverage machine learning (ML) to characterize vocalizations, monitor movements, and enhance comprehension of vocalizations. The Watkins Marine Mammal Sound Database, a crucial resource, spans decades but poses challenges for ML classification. Addressing interpretability issues in deep learning, we employ the Wavelet Scattering Transform (WST), offering invariance and stability. WST’s application to this dataset contributes to understanding complex natural sounds. Our study presents a statistical analysis using WST, emphasizing class dispersion, demonstrating high accuracy compared to existing preprocessing methods.