# Module 5 — Park brief (EuroSAT practice)

**To:** Rangers, ecologists and Landscape Recovery partners  
**From:** PDLL trainee (worked example)  
**Re:** Draft land-cover classifier — **practice dataset only**

## What we mapped

We classified **four land-cover groups** from public **EuroSAT** satellite tiles, chosen as teaching proxies for Peak District habitats:

| EuroSAT class | Habitat group (PDLL mapping) |
| --- | --- |
| Pasture | Improved grassland / grazing |
| HerbaceousVegetation | Rough grass, heather mosaic, moor edge |
| Forest | Woodland |
| AnnualCrop | Valley-bottom arable |

This is **not** an operational map of the Peak District. It is a skills exercise on 64×64 RGB patches.

## How

- **Data:** EuroSAT RGB (Zenodo), filtered to four classes, 85% / 15% train–test split with stratification.  
- **Model:** **MobileNetV2** backbone (ImageNet weights, frozen), new four-class head, trained on scaled 0–255 inputs with Keras `preprocess_input`.  
- **Evaluation:** Test set untouched until final scoring; confusion matrix and sample gallery saved.

## Accuracy

**Test accuracy: 93.4%** (transfer-learning model saved as `models/eurosat_peaks.keras`).

**Main confusion:** **HerbaceousVegetation vs Forest** — 23 herb tiles predicted as forest and 18 forest as herbaceous in the test set. Both are dark and textured from above; encroaching scrub and mixed pixels explain most errors. **Pasture vs herbaceous** swaps were smaller but still present (10 + 8 tiles).

See `outputs/charts/module5_confusion_and_samples.png`.

## Limitations

1. **Geography:** EuroSAT covers Europe, not Derbyshire; colours and field patterns differ from APGB.  
2. **Resolution:** 64×64 Sentinel-2 tiles ≠ 512×512 12.5 cm orthophotos used in the published PDNNP project.  
3. **No ground survey:** Labels come from the dataset, not ranger verification on the ground.  
4. **Season unknown:** Single-date tiles; heather and grass change appearance through the year.  
5. **No OS merge:** Buildings, roads and paths are not integrated as in production.

## Recommendation

**Use as a draft teaching map only.** Before any grazing, planting or subsidy decision, verify **pasture boundaries and herbaceous–forest edges on foot** — especially where the model confuses green herbaceous vegetation with woodland. For Landscape Recovery monitoring, this workflow shows the right *process* (honest split, confusion analysis, written brief) but would need **APGB imagery, Taylor-schema labels and ground truth** before it could support operational work.

---

*This brief follows the PDLL template used before a map would be treated as operational. Data: EuroSAT practice only.*