AdaSemSeg: Trained Model Checkpoints for Adaptive Few-Shot Semantic Segmentation of Seismic Facies

Trained model checkpoints accompanying the paper "AdaSemSeg: An Adaptive Few-Shot Semantic Segmentation of Seismic Facies" (Saha & Whitaker, IEEE Transactions on
Geoscience and Remote Sensing, 2025, DOI: 10.1109/TGRS.2025.3595010).

Includes weights for all three seismic facies datasets studied in the paper (F3,Parihaka, Penobscot) under both 1-shot and 5-shot settings, following a leave-one-out meta-training protocol: each checkpoint is meta-trained on the two source datasets and evaluated on the held-out target dataset.

Contents:
- adasemseg_checkpoints.zip — AdaSemSeg best-model checkpoints, organized as <dataset>/<shots>-shot/bestmodel.pth.tar (6 files: F3, Parihaka, Penobscot x
  1-shot, 5-shot)
- protosemseg_checkpoints.zip — ProtoSemSeg (competing prototype-based few-shot baseline) checkpoints, same <dataset>/<shots>-shot layout (6 files)
- simclr_checkpoint.zip — SimCLR ResNet-50 checkpoint used to initialize the shared image encoder via self-supervised pretraining on unlabeled seismic data

Usage instructions and the corresponding evaluation/reproduction code are at the companion GitHub repository: https://github.com/Surojit-Utah/AdaSemSeg. See
checkpoints/scenarios.json in the repository for the exact mapping between each checkpoint and the paper's reported results.