Seismic Segmentation Al Eval

seismic-segmentation-al-eval

Effective Data Selection for Seismic Interpretation through Disagreement — Benkert et al. (2024) (arXiv:2406.05149, 2024)

What this evaluates

This evaluation probes a model's ability to perform semantic segmentation on 3D seismic data under an active learning regime. It measures how effectively a model generalizes to unseen geological volumes when trained on a sequentially selected subset of annotated sections, rather than a fixed passive dataset.

Datasets

• F3 benchmark — total ?; splits: train (-1), test (-1)
• Parihaka — total ?; splits: train (-1), test (-1)

Metrics

• mIoU (primary) — range: [0, 1]
  • Mean Intersection-over-Union across all 6 facies classes. Calculated as the average of IoU (intersection over union) for each class.
• class accuracy — range: [0, 1]
  • Per-class pixel accuracy averaged across the 6 facies categories.

Input / output format

Input: 3D seismic data sections/volumes (clipped to [-1500, 1500], depth sub-sampled by factor 3). Each instance is a spatial slice or volume section fed to a DeepLabV3-ResNet18 model.

Output: Pixel-wise segmentation mask assigning one of six facies categories to each voxel/pixel in the input section.

Scoring recipe

def compute_miou(predictions, ground_truth, num_classes=6):
    ious = []
    for c in range(num_classes):
        pred_c = (predictions == c)
        gt_c = (ground_truth == c)
        intersection = np.logical_and(pred_c, gt_c).sum()
        union = np.logical_or(pred_c, gt_c).sum()
        if union == 0:
            ious.append(0.0)
        else:
            ious.append(intersection / union)
    return np.mean(ious)

Common pitfalls

• The Parihaka test set is not the original held-out test volume; it is a synthetic split of the original training volume because original test labels are unavailable.
• Training is halted dynamically when training mIoU reaches 0.9, rather than using a fixed number of epochs, which may bias convergence comparisons across acquisition functions.
• Active learning batch size is fixed at 2 sections per round, which is unusually small and may not reflect standard active learning protocols.

Evidence (verbatim from paper)

During each active learning round, we train the model until it achieves an at least an mean-intersection-over-union (mIoU) performance of 0.9 and log performance on the test volume in overall mIoU, as well as class accuracy.

Citation

@misc{benkert2024atlas,
  title={Effective Data Selection for Seismic Interpretation through Disagreement},
  author={Benkert et al. (2024)},
  year={2024},
  note={arXiv:2406.05149}
}

• arXiv: 2406.05149