Flip N Slide Eval

---
name: flip-n-slide-eval
description: Evaluates a novel tiling and augmentation strategy for Earth observation imagery against conventional tiling. It measures the method's ability to preserve spatial context and improve semantic segmentation performance on highly imbalanced geospatial data. Use when the user wants to benchmark on Land Cover of Canada (LCC), or asks about evaluating this task. Reports precision.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2404.10927
  bibtex_key: abrahams2024tiling
  confidence: medium
---

# flip-n-slide-eval

> A Concise Tiling Strategy for Preserving Spatial Context in Earth Observation Imagery — Abrahams et al. (2024) (arXiv:2404.10927, 2024)

## What this evaluates

Evaluates a novel tiling and augmentation strategy for Earth observation imagery against conventional tiling. It measures the method's ability to preserve spatial context and improve semantic segmentation performance on highly imbalanced geospatial data.

## Datasets

- **Land Cover of Canada (LCC)** — total 12800; splits: train (-1), test (-1)

## Metrics

- `precision` **(primary)** — range: [0, 1]
  - Standard pixel-wise precision calculated as the number of correctly predicted positive pixels divided by the total number of predicted positive pixels for each class, typically averaged across classes or the full dataset.

## Input / output format

**Input**: 256x256 pixel image tiles derived from Landsat 8 satellite imagery.

**Output**: Pixel-wise semantic segmentation masks assigning one of seven land cover classes to each pixel.

## Scoring recipe

```python
def compute_precision(pred_mask, gt_mask, num_classes=7):
    correct = 0
    predicted_positive = 0
    for c in range(num_classes):
        pred_c = (pred_mask == c)
        gt_c = (gt_mask == c)
        predicted_positive += pred_c.sum()
        correct += (pred_c & gt_c).sum()
    return correct / predicted_positive if predicted_positive > 0 else 0.0
```

## Common pitfalls

- The dataset exhibits extreme class imbalance (e.g., Urban Development is 0.0003%), so global precision may mask poor performance on minority classes.
- The paper explicitly avoids test-time augmentation and label averaging, which differs from standard segmentation benchmarks that often use these techniques.
- Tiling overlap conventions (50% vs. Flip-n-Slide permutations) directly impact context availability and must be strictly controlled during comparison.

## Evidence (verbatim from paper)

> Although both algorithms perform well for the over-represented class case, Flip-n-Slide is more precise, by up to 15.8%, than the conventional strategy (50% tile overlap).

## Citation

```bibtex
@misc{abrahams2024tiling,
  title={A Concise Tiling Strategy for Preserving Spatial Context in Earth Observation Imagery},
  author={Abrahams et al. (2024)},
  year={2024},
  note={arXiv:2404.10927}
}
```

- arXiv: 2404.10927