Dual seismic image collaborative recognition algorithm based on deep learning

Abstract

To address the inefficiencies and subjectivity of traditional manual interpretation in seismic data analysis, this paper introduces a deep learning-based dual image collaborative recognition (DICR) model. The model is based on an enhanced you only look once version 8 (YOLOv8) architecture with a dual-stream feature extraction network. A multi-task-optimized Cross Stage Partial Darknet-Path Aggregation Network(CSPDarknet-PANet) backbone processes seismic stacked velocity spectra and seismic trace set data in parallel. The multi-class detection head estimates the probability distribution of energy clusters in the velocity spectrum, while the geometric morphology analysis module analyzes the geometric morphology of seismic reflection events. A novel cross-modal correction mechanism implements a bidirectional feedback system using a velocity-time domain transformation matrix. Iterative parameter optimization continuously aligns detected energy clusters with corrected seismic reflection events. Real seismic datasets were employed for end-to-end evaluation experiments. Across 728 images affected by strong noise interference and waveform distortions, the DICR model achieves an average absolute localization error of 4.7 % (±1.3 %) for energy cluster centers. Furthermore, the structural similarity index measure (SSIM) for seismic reflection event reconstruction reaches 0.912, while processing efficiency is approximately 30 times higher than that of manual interpretation. By incorporating domain knowledge into the deep learning framework via a confidence fusion (a decision-level integration of velocity spectra and gather features using weighted fusion), this model develops an intelligent recognition system with physical interpretability. The error rate is maintained within a strict 5 % confidence interval, ensuring compliance with practical engineering accuracy requirements for seismic exploration.