Detecting geological features in seismic data using segment anything model 2 across multiple datasets

There is a growing interest in applying computer vision models to seismic interpretation, as manually segmenting seismic facies is often a time-consuming task. Foundation models like Segment Anything (SAM) have shown their effectiveness in segmentation tasks. This work explores SAM and SAM 2 models for segmenting geobodies and seismic layers across three different datasets: Parihaka, F3 Netherlands and Penobscot. We conducted experiments using Meta’s pre-trained SAM and SAM 2 models and fine-tuned them on seismic data. Fine-tuning considerably improved pre-trained models on segmenting geological features, with SAM 2 showing slightly better results than SAM. This highlights how it is possible to adapt visual Foundation Models to address seismic data analysis