Landsat Collection 2 Level-2 from Microsoft Planetary Computer

v.landsat_mpc is a source verb for the MPC landsat-c2-l2 collection. It discovers STAC observations, signs asset URLs and uses stackstac to stream remote cloud-optimized GeoTIFF windows lazily into a Dask-backed xarray DataArray with dimensions (time, band, y, x). Full scenes are not downloaded up front. Discovery retrieves metadata; .compute() triggers imagery reads.

from cubedynamics import pipe, verbs as v

cube = (
    pipe(None)
    | v.landsat_mpc(
        bbox=[-105.35, 39.90, -105.15, 40.10],
        start="2024-07-01",
        end="2024-09-01",
        bands=("red", "nir08"),
        max_cloud_cover=20,
        epsg=32613,
        resolution=30,
        chunksize=1024,
    )
)
da = cube.unwrap()
assert da.chunks is not None
red = da.sel(band="red")
nir = da.sel(band="nir08")
ndvi = (nir - red) / (nir + red)  # still lazy
small_ndvi = ndvi.isel(time=0, y=slice(300, 332), x=slice(250, 282)).compute()

Optical values are physical surface reflectance. Stackstac applies STAC raster:bands scale/offset exactly once: do not apply * 0.0000275 - 0.2 again. Valid reflectance can fall outside 0–1. No NDVI or pixel-level cloud mask is baked into the source. max_cloud_cover is a strict scene-level percentage filter, not pixel-level cloud masking. For a pipe-based NDVI stage, the existing configurable index helper works with these names:

ndvi_pipe = pipe(da) | v.ndvi_from_s2(nir_band="nir08", red_band="red")

Use MPC asset keys: blue, green, red, nir08, swir16, swir22. Landsat 8/9 also have coastal. qa_pixel and qa_radsat contain encoded quality bits; qa_aerosol is specific to the Landsat 8/9 processing path. Select QA assets explicitly to construct a scientifically appropriate mask. They share the floating array dtype (with missing values); convert finite values to integers before bit operations. Stackstac honors asset nodata metadata. Do not treat QA values as reflectance. See the USGS QA definitions.

The collection includes Landsat 4–9. Common optical asset names avoid mission-specific numbered-band mapping; this does not harmonize sensors or QA semantics. Request platforms=("landsat-8", "landsat-9") to restrict missions. Missing requested assets raise an error, including mission-specific assets absent from older scenes. See the MPC collection history.

The bbox is WGS84 lon/lat and is transformed using bounds_latlon into the output CRS. Resolution is in output CRS units; 30 means meters for EPSG:32613. That CRS is appropriate for Boulder, not a global default. With epsg=None, the adapter uses only complete, unanimous item/asset proj:epsg or EPSG proj:code metadata. Missing or mixed projection metadata requires explicit epsg=. Output bounds snap outward to the resolution grid; coordinates label the upper-left pixel corners. The output grid is rectangular, but scene edges, missing observations and nodata can still leave empty pixels.

Direct invocation is v.landsat_mpc(None, **options); calling without the input creates a pipe stage. Core installation declares pystac-client, planetary-computer and stackstac. Signed URLs expire: rebuild the source stage for later remote computations. Chunk size controls Dask tasks, not a total IO budget; STAC search materializes the item catalog.

v.landsat8_mpc and the existing visualization helpers remain compatibility interfaces. The Landsat 8 stage retains band_aliases=("red", "nir"), chunks_xy={"x": 1024, "y": 1024}, float32 output and the nir coordinate, with a verified platform="landsat-8" filter. It now correctly returns scaled reflectance and accepts epsg= and resolution=; provide EPSG if metadata cannot infer it. No new deprecation warning is imposed. Visualization helpers materialize data intentionally and also accept epsg=. Prefer the new source stage for analysis and separate NDVI from visualization.

Run the bounded live Boulder check explicitly:

CUBEDYNAMICS_TEST_LANDSAT_MPC=1 python -m pytest tests/verbs/test_landsat_mpc.py -k boulder_live -q

The live test is marked online, integration, and streaming; normal offline validation excludes it. It checks the projected bbox, lazy bands and NDVI, small remote reads and plausible values without imposing 0–1 on every pixel.