Sequence-constrained multitask horizon tracking

Abstract

Deep learning-based automatic horizon tracking has achieved promising results but still faces serious cross-horizon phenomena. Therefore, based on multitask learning, seismic data characteristics, and the seismic horizon sequence (or position) relationship, we have developed a sequence-constrained multitask horizon tracking (SMHT) method for high-precision automatic horizon tracking. SMHT contains the horizon label automatic enrichment, the multitask horizon tracking network (MHTN), and the horizon sequence-constrained loss function. Horizon label automatic enrichment aims to automatically generate the upper and lower auxiliary horizon labels of the target horizon label, and then the horizon region labels corresponding to the auxiliary and target horizon labels for MHTN. MHTN contains the shared layer, the auxiliary task, the main task, and the horizon sequence-constrained horizon correction. In MHTN, the shared layer generates multiscale feature maps of the input seismic data. The auxiliary task uses the concept of object detection to extract the horizon region (or probability), and the main task extracts the high-precision horizon within the extracted horizon region. The horizon sequence-constrained horizon correction with the horizon sequence-constrained loss function aims to avoid the cross-horizon phenomenon and finally obtain precise horizon tracking results. Application of MHTN to two field 3D seismic data sets finds that SMHT performs well in automatic horizon tracking.