
Uncertainty Estimation of Subsurface Fault Interpretation Leveraging Multi-Quality Annotations
In this work show that while fault delineation is hampered by weak seismic reflectors, degraded image quality, and interpreter variability, non-expert annotations can approximate expert quality in certain cases and offer valuable insights for ambiguous images. We propose an algorithm (details omitted) that incrementally integrates additional interpreters in reverse quality order, which enhances disagreement and uncertainty modeling compared to approaches relying solely on scarce expert-labeled data.

Multi-Instance SAM2 for Seismic Segmentation
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

Line SAM: a Line-based Prompting for Seismic Foundation Models
In this work we propose a line-based prompting method for seismic foundation models that leverages the inherent geophysical line features—such as reflections and horizons—to better delineate target geobodies.

Synthetic pretraining is not enough: domain shift through normalization
In this work, we train eight segmentation models on three fault segmentation datasets, Thebe, synthetic, and CRACKS, to study domain shifts. Our results show that domain adaptation in seismic images is closely tied to normalization, as dataset-specific intensity differences strongly affect model performance.