First Break Detection: A Self-Improving Deep Learning Framework

Abstract

This work presents an iterative deep learning framework for first-break detection in seismic data processing that prioritizes dataset-specific adaptability over universal generalization. The methodology combines a U-Net architecture with domain-specific quality control mechanisms to expand the training dataset from limited initial annotations. Experiments conducted on two Canadian mining seismic datasets demonstrate the framework’s effec- tiveness, with the Halfmile dataset showing improvements through iterative training - reducing mean absolute error from 16.16 to 9.16 samples while maintaining high window accuracy (0.9677). Results from the Sudbury dataset reveal the framework’s sensibility to data quality and underlying annotation consistency. These findings suggest potential benefits in developing workflows that adapt to specific survey characteristics.