Borehole Segmentation Eval

---
name: borehole-segmentation-eval
description: This evaluation probes a model's ability to perform weakly supervised multimodal segmentation of acoustic borehole images by refining threshold-guided pseudo-labels using depth-aligned well logs. It measures how well the predicted segmentation aligns with a provisional target map, testing spatial coherence and multimodal feature fusion rather than absolute geological accuracy. Use when the user wants to benchmark on Antilope25 & Botorosa47 borehole intervals, or asks about evaluating this task. Reports permutation-invariant agreement.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2603.20729
  bibtex_key: silva2026weakly
  confidence: high
---

# borehole-segmentation-eval

> Weakly supervised multimodal segmentation of acoustic borehole images with depth-aware cross-attention — Silva (2026) (arXiv:2603.20729, 2026)

## What this evaluates

This evaluation probes a model's ability to perform weakly supervised multimodal segmentation of acoustic borehole images by refining threshold-guided pseudo-labels using depth-aligned well logs. It measures how well the predicted segmentation aligns with a provisional target map, testing spatial coherence and multimodal feature fusion rather than absolute geological accuracy.

## Datasets

- **Antilope25 & Botorosa47 borehole intervals** — total ?; splits: test (-1)

## Metrics

- `permutation-invariant agreement` **(primary)** — range: [0, 1]
  - Fraction of pixels assigned to the same class after optimal matching of class identities between the predicted segmentation and the pseudo-label reference. Computed as the maximum agreement over all possible class label permutations.

## Input / output format

**Input**: 2D acoustic borehole image (amplitude) concatenated with 1D well log channels (CAL, GR, DEN, NEU, DTC, RES90) interpolated to the image depth grid and laterally replicated to match image dimensions.

**Output**: 4-class pixel-wise segmentation map assigning each pixel to one of four structural classes.

## Scoring recipe

```python
def compute_agreement(pred, ref, num_classes=4):
    best = 0.0
    for perm in itertools.permutations(range(num_classes)):
        mapped = [perm[c] for c in pred.flatten()]
        correct = sum(p == r for p, r in zip(mapped, ref.flatten()))
        score = correct / len(ref.flatten())
        if score > best: best = score
    return best
```

## Common pitfalls

- Class labels are arbitrary; failing to compute optimal permutation matching will yield artificially low scores.
- The reference is a pseudo-label map, not expert ground truth; high agreement measures internal consistency, not geological correctness.
- Multimodal log fusion does not universally improve performance; it can dilute image features in certain geological intervals.

## Evidence (verbatim from paper)

> The raw global thresholding in Figure 2e remains highly fragmented, reaching a permutation-invariant agreement (see Methods section) of 0.6002 with the pseudo-label reference. This metric measures the fraction of pixels that agree after optimal matching of class identities between the two segmentations.

## Citation

```bibtex
@misc{silva2026weakly,
  title={Weakly supervised multimodal segmentation of acoustic borehole images with depth-aware cross-attention},
  author={Silva (2026)},
  year={2026},
  note={arXiv:2603.20729}
}
```

- arXiv: 2603.20729