Toward User-Guided Seismic Facies Interpretation With a Pre-Trained Large Vision Model

Recent advancements in computer vision and deep learning have made a paradigm shift in seismic facies interpretation. Various supervised and unsupervised deep learning-based seismic facies research using Convolutional Neural Network (CNN) as the baseline architecture have been proposed and achieved promising results. However, these approaches are often limited by the lack of available high-quality labelled dataset and minimum explainability of the model. This is further compounded by the fact that these models do not allow user-guided interaction, which limits the ability of seismic interpreters to perform localized seismic facies segmentation through prompts. To optimize the already existing approach, we present a time-efficient and high precision alternative by reformulating seismic facies segmentation as a Segment All or Segment One (SASO) task. In this paper, we propose FaciesSAM, based on Fast Segment Anything Model, for seismic facies segmentation. For the first time, our paper highlights that FaciesSAM can help to decouple seismic facies identification into all-instance segmentation (segment all) and prompt-guided selection (segment one) for broad and localized facies interpretation, respectively. The effectiveness of our proposed method was evaluated on the benchmark F3 geological dataset. Experimental results shows that our method is effective, even when trained on fewer seismic images, achieving a mAP0.5 of 83.3% with 3.1% increase in pixel accuracy and 4.6% increase in mean class accuracy when compared to the CNN benchmark results. Furthermore, our method meets real-time processing speed of 1.14ms per seismic section. These results underscore the capacity of prompt-based CNN detectors to solving specific geological challenges such as seismic interpretation accuracy and processing speed.