Optimizing deep-learning first-break picking

ABSTRACT Accurate first-break picking is a key requirement for high-quality velocity model building, but manual picking remains time-intensive and error-prone. Existing single-station automatic approaches neglect spatial coherence across neighbouring traces and other deep-learning–based first-break picking approaches sacrifice temporal resolution through data downsampling, limiting their reliability. We present DeepFB, a U-Net–based neural network designed for robust automatic first-break picking in active-source seismic data. The model operates on overlapping chunks of multiple traces, preserving temporal resolution while exploiting spatial correlations. Automated hyperparameter optimization using the evolutionary algorithm Propulate yielded optimal model configurations without manual tuning. Among other hyperparameters, we tested whether model performance improves when training is performed with a reduced traveltime data set or when noise augmentation techniques are used to improve first-break picking, particularly in noisy ocean-bottom seismometer records. Application to the HIPER2 experiment at the coastline of Ecuador demonstrates that DeepFB achieves picking accuracy comparable to manual picking, with residuals close to the manual picks from our test data set. Tomographic inversions based on automatic and manual picks produce consistent velocity models, with only localized discrepancies near complex geological structures. DeepFB thus enables accurate, efficient, and scalable first-break picking, reducing manual workload while maintaining reliability.