Inspired by the geophysical features of the seismic data that comes in the form of the lines —such as reflections and horizons—, we develop a line-based prompting method for a seismic foundation model. Seismic foundation models have been an attractive research topic due to their ability to generalize across different domains []. However, they still require not just a lot of diverse data to generalize, but also substantial computational resources for training and inference. Current works use the same concept of Segment Anything model (SAM), which is a foundation model for general image segmentation. SAM is prompt-based model that uses user-defined point prompts to isolate desired regions. However, some researchers show that SAM is biased to the edges that exists in the image, making it essentially an edge detector. This problem as we show in Fig. 1, can affect the results of SAM or any point-based model because points are very sparse and lack granularity. Hence we address this gap by developing a new method of prompting that emphasize the edges of the target geobody. To achieve this the interpreter identifies edges of the target geobody using lines rather sparse points. The lines will be stored in a mask that will be fed to the model. In addition to that, consecutive points will be sampled from the given line which will form a another set of prompts to the model. As a results, two forms will of prompts will be given to the model, a mask and series of points.
| Model | mIoU | Dice |
| SAM | 0.418 | 0.582 |
| Finetuned SAM | 0.270 | 0.395 |
| Line based SAM (Ours) | 0.803 | 0.888 |
For implementation, we utilize the image encoder of SAM, and train the prompt encoder and the mask decoder to adapt to this concept of prompting. We train the model using F3 block dataset to segment the salt dome region. For comparison, we also finetune SAM on the same dataset. Moreover, we take the pretrained SAM and test all three models on the test section of the F3 block. It is worth noting that the test section is notably different from the training section, and is regarded as a hard segmentation problem. We should our results in Table 1.
