Used in research¶
Independent research has already used Umbra's open archive as an ISR-style training source. This page points at that work and at the umbra-py path that gets you from catalog filters to trainable chips — discovery → arrays, not the model.
ProSR¶
ProSR (Kim & Kim, KAIST-VICLab) curated about 502 Umbra SLC acquisitions into roughly 132k patches at 0.25 m as an ISR super-resolution benchmark. Code and data notes live at KAIST-VICLab/ProSR.
umbra-py does not reimplement ProSR or any diffusion / SR model. It covers
the data path those papers need: filtered search, size-aware download, and
chips. ProSR tiled slant-range SLC amplitude (not map geometry) —
sicd_to_amplitude_geotiff is the matching convert; sicd_to_geocoded_cog
warps into ground-range and is a different product. GEC chips are the fast
smoke stand-in, not a ProSR-equivalent training tensor.
How umbra-py helps¶
- Index search with polarization and incidence filters
(
CatalogIndex.from_release()/index.search(...)) so multi-filter corpus assembly does not crawl S3. - Download (
download_item) after confirming asset size (e.g. HTTPHEAD/Content-Lengthon the assethref). - Convert (
sicd_to_amplitude_geotiff/umbra convert --slant-plane) for slant-plane SLC amplitude (the ProSR-like tensor);sicd_to_geocoded_cog/umbra convertwhen you need that amplitude on a map instead. - Chips (
write_chips/umbra chips) for georeferenced training tiles from open GEC — a fast smoke stand-in, not ProSR's slant-range SLC patches.
Walkthrough:
examples/09_isr_training_set.ipynb
(search → size-check → chip). Chip mechanics alone:
examples/05_detection_chips.ipynb.
License & affiliation¶
Umbra open imagery is CC BY 4.0 — attribute "Contains Umbra open data, licensed under CC BY 4.0." when you publish derived products. This is an independent, unofficial toolkit and is not affiliated with Umbra Lab, Inc.