A Self‐Training U‐Net Approach for First‐Break Picking With Minimal Annotations

ABSTRACT

First‐break picking is critical in seismic data processing. Traditional techniques such as short‐term average/long‐term average often fail in noisy environments, while deep learning methods like U‐Net require large labelled datasets. We propose a site‐specific self‐training framework with three contributions: (1) a windowed labelling strategy expanding point annotations into 11‐sample windows, reducing class imbalance by an order of magnitude; (2) a weighted loss function penalizing missed picks 100‐fold more than false positives; and (3) an iterative self‐training procedure augmenting training data with quality‐controlled predictions. Using only 1% of gathers for manual labelling, our U‐Net method achieves over 98% prediction coverage and above 97% accuracy ( samples) across four mining exploration datasets. Compared to cross‐site transfer approaches requiring fully annotated surveys, we achieve comparable first‐break picking accuracy with two orders of magnitude less labelled data, particularly when acquisition parameters differ between sites.