portolan-pipeline
A metadata-only mirror of 924 National Agriculture Imagery Program scenes over a Colorado study area, flown August to October 2023 at 0.3 m in red, green, blue and near infrared. The Cloud Optimized GeoTIFFs stay on the Microsoft Planetary Computer and are referenced by URL, so what this Collection publishes is the STAC Items and a stac-geoparquet item mirror of them.
The National Agriculture Imagery Program flies the continental United States at least once every three years, during the growing season, and releases the result into the public domain. It is administered by the Aerial Field Photography Office within the USDA Farm Service Agency. These scenes are from the 2023 Colorado flight, reaching this catalog through the Microsoft Planetary Computer, which holds NAIP from 2010 onward. Pixels are in UTM zone 13N, EPSG:26913.
The scene bytes are not here. They total 1.86 TB, around 2 GB each, and stay on the host that publishes them. Reading all 924 footprints from the item mirror is one request. Doing the same by fetching the Items would be 924.
That is the shape this example exists to demonstrate, a Portolan Collection that describes data it does not host. Every other Collection in these examples serves its own bytes.
Agents, AGENTS.md covers the band order, the public domain reading, and why no scene carries a checksum.
This Collection describes scenes it does not host. Each Item carries its Cloud Optimized GeoTIFF as an item-level asset, and the asset href points at the upstream host. Read the whole item index in one query with DuckDB.
Then read any single scene straight from the upstream host, over range requests, without downloading it.
The point of the item mirror is that questions about the collection do not cost a scene read. This one touches no imagery at all.
Footprints are real geometries, so spatial predicates work directly. Colorado Springs falls inside exactly one scene.
A minimal.json sits beside the mirror carrying only bounding boxes
and hrefs, for a browser client that wants the mosaic index without a
Parquet reader.
Land cover and vegetation work at parcel scale, change detection against earlier NAIP flights, and training data for segmentation, the near infrared band making vegetation separable without an index. It is also the worked example for reading a large raster collection you do not host, which is the pattern most public imagery arrives in.
Aerial imagery, not satellite, so acquisition is opportunistic rather
than on a repeat cycle and a scene's date is the day a plane flew it.
Colours are display-oriented 8-bit, not calibrated reflectance, so
cross-scene radiometric comparison is unsound even within one flight.
Band statistics published here are read from each scene's coarsest
overview and are flagged approximate in the metadata for that reason.
Coverage is one study area, not the state.
License, other. Attribution, NAIP imagery provided by USDA Farm Service Agency. Providers, USDA Farm Service Agency (producer, licensor), Esri (processor), Microsoft (processor), Portolan SDI (host). Original source, https://planetarycomputer.microsoft.com/api/stac/v1/search?collections=naip&bbox=-106.6059,38.7455,-104.5917,40.4223&datetime=2023-01-01T00:00:00Z/2023-12-31T23:59:59Z&filter-lang=cql2-json&filter={"op":"=","args":[{"property":"naip:state"},"co"]}&limit=1000 . Scenes, 924. Cloud-native asset, items.parquet (stac-geoparquet item mirror). Scene bytes are hosted upstream and referenced by URL, not copied here. Each scene asset carries file:size read from the host and no file:checksum, which core.md asks for as a SHOULD and which this catalog cannot compute for bytes it never reads. The upstream source is a live endpoint, so it is referenced by URL only and not archived as a source asset.