Performance Analysis of SegFormer for Fault Prediction

Abstract

Accurately identifying faults in seismic images is critical to reservoir characterization, structural geological interpretation, and well placement. While traditional methods rely on horizontal discontinuities in seismic reflectivity, they are often plagued by artifacts and require manual correction as post-processing. SegFormer, based on a backbone transformer, has successfully tackled this challenge by utilizing the open-source dataset from the Thebe gas field located in Australia's Carnarvan basin. The experiment involved comparing three primary backbone models, namely, MiT-B1, MiT-B3, and MiT-B5, along with hyperparameter tuning. Tests demonstrate that the lowest cross-entropy loss is MiT-B5 with 98.8% accuracy. In addition, it is noteworthy that SegFormer displays a high degree of accuracy in predicting faults through inference in 3D post-stack seismic migration while producing minimal artifacts, suggesting that the technology has the potential to supplant traditional seismic attributes in fault interpretation processes.