---
license: mit
language:
- en
tags:
- remote-sensing
- archaeology
- satellite-imagery
- Sentinel-2
- FABDEM
- computer-vision
task_categories:
- image-classification
- object-detection
size_categories:
- 1K<n<10K
---

# Archaeological Site Dataset ([CAA UK 2025](https://uk.caa-international.org/caa-uk-2025/))

## Dataset Summary

This dataset provides a comprehensive multi-channel remote sensing dataset for training machine learning models to detect archaeological sites. The dataset combines Sentinel-2 satellite imagery, FABDEM elevation data, and derived spectral indices to create 11-channel representations of 1×1 km grid cells at 10m resolution.

**Key Features:**
- **Multi-modal data**: 6 spectral bands + 3 spectral indices + 2 terrain features
- **Balanced dataset**: Positives, integrated negatives, landcover negatives, and unlabeled samples
- **Extensive augmentation**: Geometric (rotation) and radiometric augmentations
- **High resolution**: 100×100 pixels per grid cell (10m/pixel)
- **Geographic context**: Integrated negatives from same regions as archaeological sites

## Dataset Structure

### Data Instances

Each sample consists of:
- **11 channels** stored as separate `.npy` files (float32, 100×100 pixels each)
- **Binary label**: 1 (archaeological site), 0 (non-site), or -1 (unlabeled)
- **Metadata**: Geographic coordinates, rotation angle, augmentation type, site information

Example directory structure:
```
grid_000001_rot000/
├── channels/
│   ├── B2.npy          # Sentinel-2 Blue
│   ├── B3.npy          # Sentinel-2 Green
│   ├── B4.npy          # Sentinel-2 Red
│   ├── B8.npy          # Sentinel-2 NIR
│   ├── B11.npy         # Sentinel-2 SWIR1
│   ├── B12.npy         # Sentinel-2 SWIR2
│   ├── NDVI.npy        # Normalized Difference Vegetation Index
│   ├── NDWI.npy        # Normalized Difference Water Index
│   ├── BSI.npy         # Bare Soil Index
│   ├── DEM.npy         # Elevation (FABDEM)
│   └── Slope.npy       # Terrain slope
├── labels/
│   ├── binary_label.npy
│   ├── pos_type.txt
│   └── neg_type.txt
└── info.json
```

### Data Fields

#### Channel Schema (11 channels per grid)

| Index | Channel | Source | Resolution | Wavelength/Description |
|-------|---------|--------|------------|------------------------|
| 0 | B2 | Sentinel-2 | 10m | Blue (490nm) |
| 1 | B3 | Sentinel-2 | 10m | Green (560nm) |
| 2 | B4 | Sentinel-2 | 10m | Red (665nm) |
| 3 | B8 | Sentinel-2 | 10m | NIR (842nm) |
| 4 | B11 | Sentinel-2 | 20m→10m | SWIR1 (1610nm) |
| 5 | B12 | Sentinel-2 | 20m→10m | SWIR2 (2190nm) |
| 6 | NDVI | Calculated | 10m | (B8-B4)/(B8+B4) - Vegetation |
| 7 | NDWI | Calculated | 10m | (B3-B8)/(B3+B8) - Water |
| 8 | BSI | Calculated | 10m | Bare Soil Index |
| 9 | DEM | FABDEM | 30m→10m | Elevation (meters) |
| 10 | Slope | Derived | 10m | Terrain slope (degrees) |

#### Metadata Fields (`grid_metadata.parquet`)

| Column | Type | Description |
|--------|------|-------------|
| `grid_id` | string | Unique grid identifier (e.g., "grid_000001_rot000") |
| `centroid_lon` | float | Grid center longitude (WGS84) |
| `centroid_lat` | float | Grid center latitude (WGS84) |
| `label` | int | 1 = site, 0 = non-site, -1 = unlabeled |
| `label_source` | string | Data source origin |
| `image_path` | string | Path to grid directory |

### Data Splits

**CRITICAL: Prevent Data Leakage**

Do **NOT** split randomly! Rotations and augmentations of the same site must stay in the same split.

**Recommended approach:**
1. Group samples by original site index (extracted from `grid_id`)
2. Split sites (not samples) into train/val/test
3. All rotations/augmentations of a site go to the same split

**Suggested ratios:**
- Train: 70% of sites
- Validation: 15% of sites
- Test: 15% of sites

## Dataset Composition

### Sample Types

Given **N** known sites and rotation step of 120° (R=3 rotations):

| Data Type | Count | Label | Description |
|-----------|-------|-------|-------------|
| **Positives (base)** | 3×N | 1 | Original + 2 rotations per site |
| **Positives (augmented)** | 9×N | 1 | 3 radiometric variants × 3 rotations |
| **Total Positives** | **12×N** | 1 | All positive samples |
| **Integrated Negatives (base)** | 3×N | 0 | From same areas as sites |
| **Integrated Negatives (aug)** | 9×N | 0 | 3 variants × 3 rotations |
| **Total Integrated Neg.** | **12×N** | 0 | Surrounding landscape context |
| **Landcover Negatives** | 3×N | 0 | Urban/water/cropland |
| **Unlabeled** | ~1.5×N | -1 | Background samples |
| **TOTAL** | **~28.5×N** | | Complete dataset |

### Data Augmentation

**1. Geometric Augmentation (Rotation)**
- 3 rotations per site: 0°, 120°, 240°
- Extracted at 1.5× size, rotated, then center-cropped
- Applied to positives and integrated negatives

**2. Radiometric Augmentation**
Three variants per rotated sample:
- **aug1**: +8% brightness, +5% contrast, noise σ=0.015
- **aug2**: -8% brightness, -5% contrast, noise σ=0.015
- **aug3**: No brightness/contrast, noise σ=0.025

## Dataset Creation

### Source Data

**Satellite Imagery:**
- **Sentinel-2**: Multi-spectral optical imagery (2023-2024)
- **FABDEM**: Forest And Buildings removed Copernicus DEM

**Archaeological Sites:**
- Known archaeological site locations (latitude/longitude)
- Site types may include geoglyphs, mounds, settlements, etc.

**Negative Samples:**
- **Integrated negatives**: 4 corners of rotated grids (same geographic areas)
- **Landcover negatives**: Urban (40%), water (30%), cropland (30%)
- **Unlabeled**: Random background samples with exclusion buffer

### Data Collection Pipeline

1. **Known site extraction**: Multi-channel data centered on archaeological sites
2. **Rotation generation**: Geometric augmentation (0°, 120°, 240°)
3. **Integrated negatives**: Corner sampling from same regions
4. **Landcover negatives**: Sampling from urban/water/crop areas
5. **Unlabeled sampling**: Random background with site exclusion
6. **Radiometric augmentation**: Brightness/contrast/noise variations

## Usage

### Loading the Dataset

```python
import numpy as np
import pandas as pd
from pathlib import Path

# Load metadata
metadata = pd.read_parquet('grid_metadata.parquet')

# Load a single sample
def load_sample(grid_path):
    channels = {}
    channel_names = ['B2', 'B3', 'B4', 'B8', 'B11', 'B12', 
                     'NDVI', 'NDWI', 'BSI', 'DEM', 'Slope']
    
    for ch in channel_names:
        channels[ch] = np.load(f'{grid_path}/channels/{ch}.npy')
    
    # Stack into (11, 100, 100) tensor
    data = np.stack([channels[ch] for ch in channel_names], axis=0)
    
    # Load label
    label = np.load(f'{grid_path}/labels/binary_label.npy')
    
    return data, label

# Example
sample_data, sample_label = load_sample('grid_images/grid_000001_rot000')
print(f"Data shape: {sample_data.shape}")  # (11, 100, 100)
print(f"Label: {sample_label}")  # [1] or [0] or [-1]
```

### PyTorch DataLoader with Proper Splitting

```python
import torch
from torch.utils.data import Dataset, DataLoader, WeightedRandomSampler
import pandas as pd
import numpy as np

class ArchaeologicalDataset(Dataset):
    def __init__(self, metadata_df, base_path):
        self.metadata = metadata_df
        self.base_path = base_path
        
    def __len__(self):
        return len(self.metadata)
    
    def __getitem__(self, idx):
        row = self.metadata.iloc[idx]
        grid_path = f"{self.base_path}/{row['image_path']}"
        
        # Load channels
        channel_names = ['B2', 'B3', 'B4', 'B8', 'B11', 'B12', 
                        'NDVI', 'NDWI', 'BSI', 'DEM', 'Slope']
        channels = [np.load(f'{grid_path}/channels/{ch}.npy') 
                   for ch in channel_names]
        data = torch.FloatTensor(np.stack(channels, axis=0))
        
        # Load label
        label = torch.FloatTensor(np.load(f'{grid_path}/labels/binary_label.npy'))
        
        return data, label, row['grid_id']

# Load metadata and create splits
df = pd.read_parquet('grid_metadata.parquet')

# Extract site index (CRITICAL: group by original site!)
df['site_index'] = df['grid_id'].str.extract(r'(grid|ineg)_(\d+)')[1]

# Split by sites, not samples
unique_sites = df[df['grid_id'].str.startswith('grid_')]['site_index'].unique()
np.random.seed(42)
np.random.shuffle(unique_sites)

n_train = int(0.7 * len(unique_sites))
n_val = int(0.15 * len(unique_sites))

train_sites = unique_sites[:n_train]
val_sites = unique_sites[n_train:n_train+n_val]
test_sites = unique_sites[n_train+n_val:]

# Assign splits
df['split'] = 'test'
df.loc[df['site_index'].isin(train_sites), 'split'] = 'train'
df.loc[df['site_index'].isin(val_sites), 'split'] = 'val'

# Create datasets
train_dataset = ArchaeologicalDataset(
    df[df['split'] == 'train'], 
    base_path='grid_images'
)

# Create balanced sampler for training
train_df = df[df['split'] == 'train']
weights = torch.zeros(len(train_df))
weights[train_df['label'] == 1] = 0.50 / (train_df['label'] == 1).sum()
weights[train_df['label'] == 0] = 0.40 / (train_df['label'] == 0).sum()
weights[train_df['label'] == -1] = 0.10 / (train_df['label'] == -1).sum()

sampler = WeightedRandomSampler(weights, len(train_df), replacement=True)
train_loader = DataLoader(train_dataset, batch_size=32, sampler=sampler)
```

## Train/Val/Test Split Guidelines

### CRITICAL: Prevent Data Leakage

**NEVER split randomly!** Rotations and augmentations of the same site must stay together.

### Step-by-Step Guide

```python
import pandas as pd
import numpy as np

# Load metadata
df = pd.read_parquet('grid_metadata.parquet')

# Extract base site index from grid_id
# Examples:
#   grid_000001_rot000_aug1 -> 000001
#   ineg_000045_rot120 -> 000045
df['site_index'] = df['grid_id'].str.extract(r'(grid|ineg)_(\d+)')[1]

# Get unique sites (positives only for stratification)
unique_sites = df[df['grid_id'].str.startswith('grid_')]['site_index'].unique()

# Shuffle and split SITES (not samples!)
np.random.seed(42)
np.random.shuffle(unique_sites)

n_train = int(0.7 * len(unique_sites))
n_val = int(0.15 * len(unique_sites))

train_sites = unique_sites[:n_train]
val_sites = unique_sites[n_train:n_train+n_val]
test_sites = unique_sites[n_train+n_val:]

# Assign splits based on site membership
df['split'] = 'test'
df.loc[df['site_index'].isin(train_sites), 'split'] = 'train'
df.loc[df['site_index'].isin(val_sites), 'split'] = 'val'

# Distribute landcover negatives and unlabeled randomly
mask = df['grid_id'].str.startswith(('lneg_', 'unla_'))
df.loc[mask, 'split'] = np.random.choice(
    ['train', 'val', 'test'],
    size=mask.sum(),
    p=[0.7, 0.15, 0.15]
)

# Verify no leakage
train_sites_set = set(df[df['split'] == 'train']['site_index'])
val_sites_set = set(df[df['split'] == 'val']['site_index'])
test_sites_set = set(df[df['split'] == 'test']['site_index'])

assert len(train_sites_set & val_sites_set) == 0, "Train-Val leakage!"
assert len(train_sites_set & test_sites_set) == 0, "Train-Test leakage!"
assert len(val_sites_set & test_sites_set) == 0, "Val-Test leakage!"

print(f"Train: {len(df[df['split']=='train'])} samples from {len(train_sites)} sites")
print(f"Val: {len(df[df['split']=='val'])} samples from {len(val_sites)} sites")
print(f"Test: {len(df[df['split']=='test'])} samples from {len(test_sites)} sites")
```

## Considerations for Use

### Data Characteristics

1. **Integrated Negatives**: Sampled from corners of the same geographic areas as positives (after rotation), representing surrounding landscape context. Sharp boundaries between corners with no blending.

2. **Unlabeled Data** (label = -1): Random background samples with exclusion buffer around known sites. May contain undiscovered archaeological sites. Suitable for semi-supervised learning or active learning scenarios.

3. **Landcover Negatives**: Explicitly sampled from urban areas (40%), water bodies (30%), and cropland (30%) to ensure the model learns to reject obvious non-archaeological features.

4. **Cloud Cover**: Maximum 20% cloud cover per image.

## Citation

If you use this dataset, please cite:

```bibtex
@inproceedings{li2025fusing,
  title={{Fusing Text and Terrain}: {An LLM}-Powered Pipeline for Preparing Archaeological Datasets from Literature and Remote Sensing Imagery},
  author={Li, Linduo and Wu, Yifan and Wang, Zifeng},
  booktitle={{CAA UK 2025}: Computer Applications and Quantitative Methods in Archaeology},
  year={2025},
  month={December},
  address={University of Cambridge, UK},
  organization={CAA UK},
  note={Conference held 9--10 December 2025}
}
```

**Presentation Links:**
- [Abstract](https://drive.google.com/file/d/1zQ3-LrlsDmZiI1D3QkJo1lQ43e6MwYq1/view?usp=sharing)
- [Slides](https://drive.google.com/file/d/1n79eqJ7XM3h3ftmJakhP5JOrqjONAnRE/view?usp=sharing)

## License

This dataset is released under the MIT License.

## Acknowledgments

- **Sentinel-2 satellite imagery** was provided by the European Space Agency (ESA) through the Copernicus Programme.
- **FABDEM elevation data** were provided by the University of Bristol.
- **Google Earth Engine** was used as the primary data processing and analysis platform.
- **Geoglyph location data** were derived from publicly available archaeological compilations curated by James Q. Jacobs (2025), *JQ Jacobs Archaeology*, last modified July 31, 2025: https://jqjacobs.net/archaeology/geoglyph.html

---

**Contact**: linduo.li@ip-paris.fr

**Last Updated**: December 2025