---
pretty_name: hyperComp Satellite Training Scenes
task_categories:
- image-to-image
size_categories:
- 10K<n<100K
---

# hyperComp satellite training scenes

This dataset contains 10,000 training scenes and 1,000 validation scenes for
training the hyperComp compression autoencoder. Every scene is an unpadded,
unresized 1024x1024 RGB provider response. The training loader slices each scene
into sixteen non-overlapping 256x256 patches at indexing time.

The imagery comes from seven open provider families: NASA Worldview
MODIS/VIIRS, USDA NAIP, Queensland aerial orthophotos, Netherlands PDOK, Spain
PNOA, France BD ORTHO, and Switzerland SWISSIMAGE. Exact decoded-pixel,
perceptual, and pretrained ResNet-18 semantic deduplication are applied before
the leakage-safe train/validation split.

`manifests/dataset.csv` records the split, source family, provider, product,
location, capture date, request URL, license URL, attribution, image hashes,
ground resolution, leakage group, and diversity measurements for every scene.
`manifests/dedup_report.json` and `manifests/diversity_report.json` document the
build filters and five diversity axes. Consult each row's license and
attribution fields when redistributing or using an image.