Segment Anything Model 2 (SAM2) is a recent model designed to segment objects across frames in a video or time series. It takes prompts in the first frame and tracks the object in following frames. This works well in natural videos where objects move smoothly and remain visible over time. Interestingly, seismic data shares a similar structure to video—it is a stack of slices through space, forming a 3D volume. Each slice can be thought of as a frame, and structures like faults or salt bodies appear across multiple slices.
However, unlike natural video, seismic objects do not always follow smooth, continuous motion. A salt body might disappear and reappear, or faults might shift slightly across slices. This makes tracking-based methods unreliable. At the same time, manually labeling each slice is slow and difficult. Experts must scan through hundreds of slices to mark repeating structures. This creates a strong need for tools that can find and segment all instances of a geological structure from a single example.
In this work, we adapt SAM2 to the seismic domain using embedding-based prompt transfer. We begin by prompting one slice of the seismic volume to segment a target object. We extract the image embedding from this prompted slice and compare it to the embeddings of all other slices. Using cosine similarity, we find the regions in other slices that are most like the example. These matched regions are then used as new prompt locations, allowing SAM2 to segment multiple instances of the object across the entire volume.
This method treats seismic data like a video but overcomes the problem of object discontinuity. Unlike the original SAM2, which expects a single object to persist across frames, our method allows multi-instance segmentation across space, not just time. It works even when the object appears in non-adjacent slices or changes shape. Because it relies only on embeddings, it avoids the need for training or manual prompting in every slice.
We apply this method to seismic volumes containing salt bodies and fault systems. Starting from one prompt, the model automatically finds and segments similar structures throughout the volume. This cuts down on annotation time and enables rapid mapping of complex subsurface features.
In summary, we extend SAM2 for seismic data by treating the volume like a video and using embedding-based prompt transfer to support multi-instance segmentation. Our method is simple, training-free, and leverages the strength of SAM2’s embedding space to discover repeating geological patterns. This makes seismic interpretation faster, easier, and more scalable for experts working with large datasets.
