From a STAC catalogue

When a collection is catalogued with the projection and raster STAC extensions, the catalogue already says everything the VRT needs. from_stac() uses it directly and opens no file:

VRT element	catalogue field
grid (dims, x/y coordinates)	proj:shape, proj:transform
<SRS>	proj:epsg / proj:code / proj:wkt2
time coordinate	item datetime
one array per asset	asset keys with raster:bands
dtype, nodata, scale, offset, unit	raster:bands[0]
sources	asset href, placed at the item’s time position

from xvrt import from_stac
from _fixtures import geotiffs, WKT_4326

tifs = geotiffs(n=3)
items = [
    {
        "id": f"day{i}",
        "properties": {
            "datetime": f"2026-01-0{i + 1}T00:00:00Z",
            "proj:shape": [3, 4],
            "proj:transform": [1.0, 0, 145.0, 0, -1.0, -42.0, 0, 0, 1],
            "proj:wkt2": WKT_4326,
        },
        "assets": {
            "analysed_sst": {
                "href": str(p),
                "raster:bands": [{"data_type": "int16", "nodata": -32768,
                                  "scale": 0.001, "offset": 25.0, "unit": "celsius"}],
            }
        },
    }
    for i, p in enumerate(tifs)
]
out = from_stac(items, tifs[0].parent / "catalogue.vrt")

from osgeo import gdal
gdal.UseExceptions()
a = gdal.OpenEx(str(out), gdal.OF_MULTIDIM_RASTER).GetRootGroup().OpenMDArray("analysed_sst")
print([d.GetSize() for d in a.GetDimensions()], a.GetScale(), a.GetOffset(), a.GetUnit())
a.GetUnscaled().ReadAsArray()[:, 0, :]

[3, 3, 4] 0.001 25.0 celsius

array([[12.  , 12.05, 12.1 , 12.15],
       [12.1 , 12.15, 12.2 , 12.25],
       [12.2 , 12.25, 12.3 , 12.35]])

items may be a list of STAC item dicts (nested properties or flat stac-geoparquet rows; pystac.Item objects work too) or the path or URL of a stac-geoparquet file.

What it checks

The catalogue is trusted, but not blindly:

• One grid. Items on different grids are refused (no resampling).
• One item per time. Duplicate datetimes are refused.
• One storage description per asset. If dtype, nodata, scale or offset change between items the asset is refused, because one array needs one decoding. A unit missing from some items is filled from the others; conflicting units warn and use the most common.
• Missing assets are gaps. An item without an asset leaves nodata at its time position, so a variable that starts later than the others needs no special handling.

The template Dataset it builds uses zero-memory placeholders (numpy.broadcast_to), so a catalogue of thousands of full-resolution days costs nothing to describe. See the GHRSST MUR example for 8913 days of five variables.