Geological Mapping Segmentation Eval

---
name: geological-mapping-segmentation-eval
description: Evaluates a CNN's ability to perform semantic segmentation on airborne magnetic data to identify three major lithological groups (dykes, plutons, greywackes). It tests transfer learning from synthetic geostatistical data to real-world geological contexts. Use when the user wants to benchmark on Malartic geological model & synthetic augmentations, or asks about evaluating this task. Reports IOU.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2110.14440
  bibtex_key: [REDACTED]
  confidence: medium
---

# geological-mapping-segmentation-eval

> Predictive Geological Mapping with Convolution Neural Network Using Statistical Data Augmentation on a 3D Model — Cedou et al. (2021) (arXiv:2110.14440, 2021)

## What this evaluates

Evaluates a CNN's ability to perform semantic segmentation on airborne magnetic data to identify three major lithological groups (dykes, plutons, greywackes). It tests transfer learning from synthetic geostatistical data to real-world geological contexts.

## Datasets

- **Malartic geological model & synthetic augmentations** — total ?; splits: train (-1), test (-1); repo https://github.com/MatthieuCed/GSCNN-apply-to-airborne-magnetic

## Metrics

- `IOU` **(primary)** — range: [0, 1]
  - Intersection over Union: the area of overlap between the predicted segmentation mask and the ground truth mask divided by the area of their union. Calculated across the three merged lithological classes.

## Input / output format

**Input**: 2D grids of residual magnetic anomaly values (images) representing airborne magnetic survey data.

**Output**: 3-class semantic segmentation map (dykes, plutons, greywackes), a binary contact detection mask, hierarchical clustering labels, and multi-scale attention maps.

## Scoring recipe

```python
def calculate_iou(pred_mask, gt_mask, num_classes=3):
    ious = []
    for c in range(num_classes):
        pred_c = (pred_mask == c)
        gt_c = (gt_mask == c)
        intersection = np.logical_and(pred_c, gt_c).sum()
        union = np.logical_or(pred_c, gt_c).sum()
        iou = intersection / union if union > 0 else 0.0
        ious.append(iou)
    return np.mean(ious)
```

## Common pitfalls

- The original 10 lithology classes were merged into only 3 classes (greywackes, plutons, dykes) for training and evaluation, which simplifies the task.
- Synthetic training data exhibits sharper contacts and higher variability than real data, causing a performance drop when transferring to the original Malartic model.
- Some geological objects are inherently undetectable due to small size or similar magnetic susceptibility to surroundings, capping achievable IOU regardless of model architecture.

## Evidence (verbatim from paper)

> The quantitative result for the segmentation validation is 0.687 IOU. It was calculated from synthetic models which were not used in the training process. The result on the original model presents a 0.543 IOU.

## Citation

```bibtex
@misc{cedou2021predictive,
  title={Predictive Geological Mapping with Convolution Neural Network Using Statistical Data Augmentation on a 3D Model},
  author={Cedou et al. (2021)},
  year={2021},
  note={arXiv:2110.14440}
}
```

- arXiv: 2110.14440