---
language:
- en
license: other
tags:
- remote-sensing
- earth-observation
- mineral-exploration
datasets:
- Thomaschtl/miningexploration
---

# Mining Exploration Datapoints

## Dataset Summary

This dataset contains per-pixel feature vectors engineered for mineral exploration around the Bou Azzer district (Morocco). Each sample aggregates multi-resolution satellite observations into a single record with georeferenced coordinates and a rich spectral/radar feature stack derived from Sentinel-1, Sentinel-2, ASTER, Landsat 8, and a Digital Elevation Model (DEM).

> The physical cubes (Zarr) were used to generate this dataset.  
> Due to size constraints, only derived parquet tables are hosted here.

| Field | Description |
| --- | --- |
| `x`, `y` | Projected coordinates in EPSG:32629 (UTM Zone 29N) defining the pixel centroid. |
| Sentinel-2 fields | VNIR/SWIR reflectances (`s2_b02`, `s2_b03`, `s2_b04`, `s2_b08`, `s2_b11`, `s2_b12`) plus cloud/validity masks (`s2_vnir_mask`, `s2_swir_mask`). |
| Sentinel-1 fields | Radar backscatter (`s1_vv`, `s1_vh`), ratio (`s1_vv_div_vh`), and validity mask (`s1_mask`). |
| ASTER fields | VNIR/SWIR reflectances (`aster_vnir_b01`, `aster_swir_b04`‒`aster_swir_b09`) with corresponding masks. |
| Landsat fields | Multispectral reflectances (`landsat_b02`‒`landsat_b07`) and validity mask (`landsat_mask`). |
| DEM fields | Elevation (`dem_elevation`) and validity mask (`dem_mask`). |
| `label` | Optional target label for downstream tasks (currently unset). |

## Data structure

Two export formats are available:

- `mining_datapoints.parquet` — columnar, compressed Parquet table (recommended for analytics and ML pipelines).
- `mining_datapoints.csv` — wide CSV table for compatibility with spreadsheet tools (≈1.6 GB).

Both files share the same schema described above.

## Source data & processing

1. Sentinel-2 L2A mosaics (10 m) and Sentinel-1 median composites (10 m) are aligned on a 10 m grid and validated with sensor-specific masks.
2. ASTER VNIR/SWIR, Landsat 8 OLI, and DEM inputs (30 m) are resampled and co-registered to the same spatial footprint.
3. The 10 m cube is downscaled and merged with the 30 m cube band-wise.
4. Each pixel is transformed into a `MiningDataPoint` record with flattened spectral and ancillary features.

Processing scripts live in the `cube/` and `datapoint/` modules of the Mirage Metrics repository.

## Usage

```python
import pandas as pd

df = pd.read_parquet("mining_datapoints.parquet")
print(df.columns)
```

For geospatial workflows, convert `x`/`y` into geometries using `geopandas`:

```python
import geopandas as gpd
gdf = gpd.GeoDataFrame(df, geometry=gpd.points_from_xy(df.x, df.y), crs="EPSG:32629")
```

## Recommended citation

Please cite the Mirage Metrics project or provide attribution to the dataset maintainers when reusing these samples. Replace this section with a full citation once the project has a formal reference.

## License

The license is currently marked as `other`. Update this section with the appropriate license text before publishing.