Tirauxcloud Eval

tirauxcloud-eval

TIRAuxCloud: A Thermal Infrared Dataset for Day and Night Cloud Detection — Apostolakis et al. (2026) (arXiv:2602.21905, 2026)

What this evaluates

Evaluates semantic segmentation models for day-and-night cloud detection using thermal infrared imagery, specifically testing how auxiliary environmental features improve segmentation accuracy and how well models transfer across different satellite sensors and resolutions.

Datasets

• Landsat Main — total ?; splits: train (-1), val (-1), test (-1)

Metrics

• Accuracy — range: [0, 1]
  • Fraction of correctly classified pixels divided by total pixels.
• mIoU (primary) — range: [0, 1]
  • Mean Intersection over Union across all classes. IoU for a class is the intersection of predicted and ground truth pixels divided by their union.

Input / output format

Input: Multi-channel image patches containing thermal infrared bands and optionally auxiliary features (clear-sky reference, DEM, meteorological variables). Patches are grouped by scene ID.

Output: Per-pixel segmentation mask with class labels: 2-class (0: clear, 1: cloud) or 3-class (0: clear, 1: thin cloud, 2: cloud).

Scoring recipe

def compute_metrics(pred_mask, gt_mask, num_classes=2):
    correct = (pred_mask == gt_mask).sum()
    total = gt_mask.numel()
    accuracy = correct / total
    ious = []
    for c in range(num_classes):
        pred_c = (pred_mask == c)
        gt_c = (gt_mask == c)
        intersection = (pred_c & gt_c).sum()
        union = (pred_c | gt_c).sum()
        iou = intersection / union if union > 0 else 0.0
        ious.append(iou)
    miou = sum(ious) / num_classes
    return accuracy, miou

Common pitfalls

• Spatial leakage: Must enforce a strict group-split policy by scene ID so that no patch from the same scene appears in different splits.
• Thin cloud ambiguity: The 3-class setup includes 'thin cloud' which has inherent aleatoric label uncertainty due to semi-transparency, causing consistently lower mIoU scores that reflect label noise rather than model deficiency.
• Domain shift magnitude: VIIRS has 750m resolution and different spectral wavelengths compared to Landsat's 100m, requiring substantial fine-tuning for transfer rather than simple inference.

Evidence (verbatim from paper)

We present the Accuracy and mean Intersection over Union (mIoU) in [Tables II] and[III] for the two-class (0: clear, 1: cloud) and three-class (0: clear, 1: thin cloud, 2: cloud) targets, respectively.

Citation

@misc{apostolakis2026tirauxcloud,
  title={TIRAuxCloud: A Thermal Infrared Dataset for Day and Night Cloud Detection},
  author={Apostolakis et al. (2026)},
  year={2026},
  note={arXiv:2602.21905}
}

• arXiv: 2602.21905