P-wave polarity determination via ensemble deep learning models

P-wave first-motion polarities play a central role in understanding earth dynamics. Manual or classical automated procedures for determining polarities face several challenges. To address these issues, recent advanced studies leverage deep learning techniques, particularly Convolutional Neural Networks (CNNs). This paper explores the efficacy of ensemble deep learning approach, combining predictions from multiple CNN models. Ensemble methods exhibit improved overall performance and enhanced capabilities in managing waveforms showing no polarity. Additionally, a specific augmentation procedure known as time-shift, enhances the ability to evaluate the uncertainty on noise-only waveforms.