—Seismic facies segmentation is a fundamental task in subsurface interpretation, yet the high cost and subjectivity of manual labeling severely limit the availability of annotated data for training supervised models. Few-Shot Semantic Segmentation (FSSS) provides a promising alternative by enabling segmentation from a small number of annotated examples, but its application to seismic data remains underexplored and sensitive to the spatial properties of seismic volumes. In this work, we investigate the combined impact of spatial data partitioning and support sample locality on FSSS for seismic facies interpretation.
We evaluate four representative few-shot segmentation models, PANet, PFENet, HSNet, and ASNet, under a controlled 1-way 1-shot setting using two 3D seismic datasets, Parihaka and Penobscot. Two partitioning strategies are considered: dense sub volumes (Large Rectangular Prisms) and sparse global sampling (Equally Distant Slices), together with two support selection modes: unrestricted sampling and locality-constrained sampling from the same seismic slice as the query. Experimental results demonstrate that enforcing support locality consistently improves segmentation performance across models and datasets. Moreover, combining sparse global partitioning with locality-aware support selection achieves competitive or superior performance while substantially reducing annotation effort. Our analysis further reveals notable differences in model robustness under severe class imbalance, highlighting the importance of domain-aware episodic design for reliable few-shot seismic facies segmentation.