LaNCoR Improves Earthquake Arrival Time Picking with Noisy Labels

In supervised machine learning, the presence of inaccurately labeled training data, commonly known as "label noise," poses a significant threat to model integrity and performance. This corruption directly leads to models learning erroneous mappings, resulting in poor generalization and reduced accuracy on unseen data. In seismology, current applications typically rely on either very large training datasets or extensive data augmentation techniques to mitigate the impact of such label noise, which can be both labor-intensive and costly. This paper introduces a novel method called Label Noise-Contrastive Robust Learning (LaNCoR), designed to effectively manage noisy labels in seismic signal processing tasks. A key advantage of LaNCoR is its ability to achieve this without the need for massive training datasets. The approach works by aligning the input waveform feature and label representation distributions within the feature space. This alignment mechanism helps to correct mislabeling and substantially reduce its detrimental effects on the training process. The researchers demonstrated LaNCoR's effectiveness on the critical task of P-phase arrival-time picking using real microseismic data. When applied to two baseline models and training approaches, LaNCoR showed remarkable improvements, boosting performance by up to 28.8% across various metrics. This robust learning approach holds considerable promise for enhancing model training not only in seismology but also across broader geosciences.