Learning Earthquake Wave Arrival Time Picking from Labels with Inaccuracies

The article presents the Label Noise-Contrastive Robust Learning (LaNCoR) approach, designed to address the challenges posed by inaccurately labeled training data in supervised machine learning for seismic signal processing. LaNCoR aligns input waveform feature and label representation distributions to correct mislabeling, achieving performance improvements of up to 28.8% on P-phase arrival-time picking tasks using real microseismic data. This method reduces reliance on large-scale training datasets, making it a significant advancement for practitioners in seismology and geosciences dealing with label noise in their models.