---
pretty_name: GROC
task_categories:
- object-detection
tags:
- remote-sensing
- geospatial
- aerial-imagery
- benchmark
- object-counting
- object-localization
size_categories:
- 10K<n<100K
---

# GROC

GROC is a geospatial remote-sensing benchmark with paired raster tiles, labels, split files, tile extents, and source map vector data. The data are acquired from PDOK.

## Dataset Structure

```text
GROC/
  rgb/<group>/          RGB image tiles
  cir/<group>/          Color-infrared image tiles
  basemap/<group>/      Basemap image tiles
  dsm/<group>/          Digital surface model tiles
  lc/<group>/           Land-cover tiles
  labels/<group>/       Public annotation files for train and validation samples
  splits/               Train, validation, test, and benchmark split files
  benchmark/            Benchmark RGB and CIR image tiles
  gpkg/                 Source map vector data as GeoPackage
  extent.csv            Per-tile geospatial bounding boxes
```

## Contents

- `rgb`, `cir`, `basemap`, `dsm`, and `lc` each contain 14,499 image tiles.
- `labels` contains 12,427 public annotation files.
- `splits/train.txt` contains 9,651 samples.
- `splits/val.txt` contains 2,070 samples.
- `splits/test.txt` contains 2,072 samples.
- `splits/benchmark.txt` contains 200 benchmark samples.
- `splits/benchmark_low-light.txt` and `splits/benchmark_weather.txt` each contain 100 benchmark samples.
- `benchmark/rgb` and `benchmark/cir` each contain 200 benchmark image tiles.
- `gpkg/top10nl_Compleet.gpkg` contains 31 vector layers in EPSG:28992.

## File Naming

Each split file stores sample stems without file extensions. The hosted dataset shards the large image and label folders by group to satisfy Hugging Face's per-directory file limit. The group is the leading `<category>_group_<id>` prefix of the sample stem.

For a sample stem such as:

```text
airport_group_001_feature_10873_x1_y0
```

the group is:

```text
airport_group_001
```

and the corresponding raster and label paths are:

```text
rgb/airport_group_001/airport_group_001_feature_10873_x1_y0.png
cir/airport_group_001/airport_group_001_feature_10873_x1_y0.png
basemap/airport_group_001/airport_group_001_feature_10873_x1_y0.png
dsm/airport_group_001/airport_group_001_feature_10873_x1_y0.png
lc/airport_group_001/airport_group_001_feature_10873_x1_y0.png
labels/airport_group_001/airport_group_001_feature_10873_x1_y0.txt
```

For test samples, the image products are present but label files are not included.

## Test Labels

This release withholds labels for the samples listed in `splits/test.txt`. The corresponding label files are intentionally excluded from `labels/` to support benchmark-style evaluation, prevent test-set leakage, and keep reported results comparable across submissions.

If you would like to submit testing results or need help using the dataset, please contact jiayi.wang@whu.edu.cn.

## Cloud and Shadow Synthesis

For users who want to synthesize clouds or shadows for satellite images, please visit [SatelliteCloudGenerator](https://github.com/strath-ai/SatelliteCloudGenerator).

## Tile Extents and Vector Features

`extent.csv` provides the geospatial extent of each tile:

```text
id,xmin,ymin,xmax,ymax,crs
airport_group_001_feature_311_x1_y3,176005.706999999,407660.699000001,176261.706999999,407916.699000001,EPSG:28992
```

Use the `id` column to match a tile stem from the raster folders or split files. The `xmin`, `ymin`, `xmax`, and `ymax` columns define the tile bounding box in the coordinate reference system given by `crs`.

The corresponding map vector features can be retrieved directly from `gpkg/top10nl_Compleet.gpkg` by querying layers with the tile bounding box from `extent.csv`. The GeoPackage layers use EPSG:28992 and include point, line, polygon, and multipolygon TOP10NL feature layers such as buildings, roads, water, terrain, rail, relief, places, and functional areas.

Typical workflow:

1. Read a sample stem from `splits/train.txt`, `splits/val.txt`, `splits/test.txt`, or `splits/benchmark.txt`.
2. Look up the same stem in `extent.csv` to get the patch extent.
3. Use that extent as a bounding box query against `gpkg/top10nl_Compleet.gpkg`.
4. Use the returned vector features together with the raster tile products for geospatial analysis or model training.