Riding the Wave: One-Touch Automatic Salt Segmentation by Coupling SAM and SegGPT

Abstract

AbstractDeep learning has transformed the way geological interpretations are conducted for subsurface energy exploration. Seismic image processing and interpretation are the most active areas where deep learning has been implemented to optimize the overall workflow. Among different seismic features, the identification and delineation of salt bodies often present a challenge in seismic interpretation. Salt boundary interpretation is important for understanding salt tectonics and velocity model building for seismic migration. Recent works have applied deep learning to help the identification of salt bodies with remarkable results. However, a large volume of high-quality labeled datasets is required to achieve good accuracy. Such a labeling task is costly, time-consuming, and prone to human error. This limitation hinders the progress of deep learning applications in seismic interpretation. With the rise of generative models, such as ChatGPT and zero-shot deep learning models, it is currently possible to train deep learning models with no or very minimum labeling and pre-training for a specific task. In this study, we utilized the recently introduced Segment Anything Model (SAM) to segment the salt bodies with only a few points or a line and coupled it with Segment Everything In-Context (SegGPT) for the surrounding features (i.e., non-salt bodies). The results show that the model could successfully identify and segment salt bodies with the one-touch method and show comparable accuracy with other conventional deep learning methods, achieving a mean intersection over union (mIoU) value of 0.85. For the first time, this study presents an application of combined generative and zero-shot models for seismic interpretation, particularly salt bodies identification. The proposed model has the potential to be applied to other features in seismic interpretation that would significantly optimize the process. The proposed model also allows the implementation of a greener deep learning model with a lower carbon footprint.