Datasets:

conquer997
/

ControlEarth-4K

license: cc-by-nc-sa-4.0
task_categories:
  - text-to-image
  - image-to-image
language:
  - en
tags:
  - remote-sensing
  - image-generation
  - semantic-segmentation
  - 4K
  - high-resolution
size_categories:
  - 10K<n<100K

ControlEarth-4K

ControlEarth-4K is a large-scale, high-resolution remote sensing dataset for controllable 4K (4096×4096) image generation. Each sample is a triplet of a real aerial/satellite image, a grayscale semantic label map, and an RGB colorized label map, supporting layout-controllable generation, semantic synthesis, and related evaluation protocols.

Dataset Structure

ControlEarth-4K/
├── core/          # Core release (recommended starting point)
│   ├── train/     # 22,266 triplets (images / labels / labels_RGB), ~665 GB
│   └── test/      # 100 triplets, ~2.4 GB
└── extended/      # Extended release (full version, additional splits)
    ├── train/     # ~1.6 TB
    ├── val/       # ~156 GB
    └── test/      # ~1.3 TB

Each split contains three aligned folders:

images/ — real remote sensing images (4096×4096, RGB)
labels/ — grayscale semantic label maps (4096×4096, uint8)
labels_RGB/ — colorized label maps for visualization (4096×4096, RGB)

How to Extract Tar Shards

Large splits (core/train and all extended/*) are stored as tar shards (~12–29 GB each). Each shard internally contains the aligned images/, labels/, and labels_RGB/ folders. After downloading the shards of a split, extract them in place:

for f in *.tar; do tar -xf "$f"; done

This yields the three folders with all samples merged. (core/test is provided as plain files for direct preview.)

Semantic Categories

13 foreground classes plus an unlabeled category (pixel value 0, ignored in training and evaluation):

ID	Class
0	unlabeled (ignore)
1	water
2	road
3	ship
4	car
5	aircraft
6	vegetation
7	bareland
8	baseball diamond
9	basketball court
10	tennis court
11	storage tank
12	building
13	background

Sources

Curated and harmonized from multiple public remote sensing benchmarks, including AIRS, FAIR1M, GID, GLH-Bridge, GLH-water, HRSCD, Inria Aerial Image Labeling, LandCover.ai, MiniFrance, Potsdam, SpaceNet 1/2, STAR, WHU-Building, and iSAID. All images are re-annotated/unified into the 13-class scheme at native 4096×4096 resolution.

Intended Use

Controllable remote sensing image generation (label-to-image)
High-resolution (4K) generative model training and benchmarking
Semantic layout fidelity and texture-detail evaluation (e.g., FID, KID, CLIP-I, PSNR/SSIM, and texture/detail fidelity metrics)

Citation

If you use this dataset, please cite:

@dataset{controlearth4k,
  title  = {ControlEarth-4K: A Large-scale High-resolution Dataset for Controllable Remote Sensing Image Generation},
  author = {Ye, Ziqi},
  year   = {2026},
  url    = {https://huggingface.co/datasets/conquer997/ControlEarth-4K}
}

Paper (coming soon)

GitHub: https://github.com/conquer997/ControlEarth (code coming soon)