ssta-hackathon / README.md
sebasmllr's picture
Card: precision, recall, confusion matrices
5bc2a4c verified
|
Raw History Blame Contribute Delete
1.79 kB
---
license: other
license_name: mixed-open-licences
pretty_name: SSTA super-resolution workshop data
---
# SSTA super-resolution workshop data
Small data packages for the notebooks of the workshop "Super-Resolution Satellite Imagery as a
Regional Public Good for Sustainable Development" (ADB, PhilSA, DHSUD; 26–27 November 2026).
The notebooks are in [adb-sr/sr-client](https://github.com/adb-sr/sr-client).
## `demo2/`: Practical demonstration 2
One file per place (2 km squares in the Philippines): Sentinel-2 enlarged to 2.5 m, the
super-resolved images of stage 1 and stage 2 (red, green, blue, near-infrared; reflectance ×
10,000), the map a land-cover model draws from each, rasterised public reference maps
(buildings, roads) and a validity mask. `summary.json` holds the scores (F1, IoU, precision,
recall, pixel-exact and with a tolerance of 2.5 m and 5 m). The package is made by
`tools/demos/build_demo2_package.py` in the repository above.
`oem_reference.json` and `oem_examples.npz`: the same land-cover model on 79 hand-labelled
OpenEarthMap validation tiles (13 cities), from the aerial image, from Sentinel-2 and from the
two super-resolution stages; pooled scores for buildings and roads, pixel-exact 8-class
confusion matrices, and three example tiles
(images as 8-bit display RGB, labels and maps as class numbers, 0 = unlabelled).
Sources and licences:
- Contains modified Copernicus Sentinel data (2026).
- Building footprints: Google–Microsoft–OSM Open Buildings by VIDA (ODbL).
- Roads: © OpenStreetMap contributors (ODbL), via Overture Maps.
- OpenEarthMap tiles and labels: OpenEarthMap (Xia et al., 2023), CC BY-NC-SA 4.0.
- Land-cover maps: made with a SatlasPretrain (Ai2) model fine-tuned on OpenEarthMap
(CC BY-NC-SA 4.0, non-commercial use).