Processing (optional)

Install:

pip install cdse-client[processing]

Local RGB preview

preview_product() builds a true-color RGB preview (Sentinel-2: B04/B03/B02) from a downloaded product.

from cdse.processing import preview_product

result = preview_product(
    safe_path="S2A_MSIL2A_....zip",
    bbox=[9.10, 45.40, 9.28, 45.52],
    resolution=10,
    display=True,
)

print(result["preview_path"])

Crop and stack

from cdse.processing import crop_and_stack

tiff = crop_and_stack(
    safe_path="S2A_MSIL2A_....zip",
    bbox=[9.10, 45.40, 9.28, 45.52],
    bands=["B04", "B03", "B02", "B08"],
    resolution=10,
)

Bands and resolution

An L2A product ships three resolution folders, and which bands live in each is not simply "everything at or above its native resolution":

Folder	Bands
R10m	B02, B03, B04, B08
R20m	B01, B02, B03, B04, B05, B06, B07, B8A, B11, B12
R60m	the 20m set, plus B09

Two consequences worth knowing. B08 exists only at 10m — B8A is its 20m and 60m counterpart, not a resampled B08 — and B01 and B09 are resampled up, so B01 is available at 20m despite being 60m native. B10 is dropped during L2A processing and exists in L1C only.

Asking for a band the requested folder does not contain raises ValidationError naming the resolutions where it is available, rather than quietly dropping it from the result. The agriculture, vegetation and all_20m entries of BAND_COMBINATIONS need resolution=20.

L1C products have no resolution subfolders at all: every band comes at its native resolution and the resolution argument selects nothing.

Indices: NDVI, NDWI, NDMI, NBR

All four are the same formula, (a - b) / (a + b), on a different band pair. INDEX_BANDS lists the pairs; compute_index() goes from a product to a cropped index GeoTIFF in one call:

from cdse.processing import compute_index

ndvi = compute_index("S2A_MSIL2A_....zip", "ndvi", bbox=[9.10, 45.40, 9.28, 45.52])
nbr = compute_index("S2A_MSIL2A_....zip", "nbr", bbox=[9.10, 45.40, 9.28, 45.52])

Index	Bands	Resolution	Used for
ndvi	B08, B04	10 m	vegetation vigour
ndwi	B03, B08	10 m	open water
ndmi	B8A, B11	20 m	vegetation moisture
nbr	B8A, B12	20 m	burn severity (pre minus post = dNBR)

compute_index picks the native resolution of the pair unless you pass resolution=. For bands you already have on disk, normalized_difference(a, b, out, name="NBR") does the arithmetic alone, and calculate_ndvi(nir, red, out) is the same thing named NDVI.

Cloud mask from SCL

The tile-level cloud cover you filter on at search time says nothing about the clouds over your field. L2A products carry a per-pixel scene classification (SCL, 20 m) that does:

from cdse.processing import cloud_mask_from_scl, extract_bands_from_safe, scl_clear_fraction

scl = extract_bands_from_safe("S2A_MSIL2A_....zip", ["SCL"], resolution=20)["SCL"]

if scl_clear_fraction(scl) < 0.7:
    print("mostly cloud over the AOI, skip this date")

mask = cloud_mask_from_scl(scl)  # uint8 GeoTIFF: 1 = clear, 0 = cloud/shadow/no-data

SCL_CLOUD_CLASSES (no-data, cloud shadow, medium and high probability cloud, thin cirrus) is the default set masked out; pass classes= to change it. SCL_CLASSES names all twelve.

SCL is categorical

Never resample it bilinearly: the average of "vegetation" (4) and "cloud" (8) is not a class. stack_bands uses nearest-neighbour for a band named SCL on its own; for other tools, or for reproject, pass resampling="nearest" yourself.

Note

On some Windows/Python combinations, rasterio wheels may be unavailable. If you hit install issues, try Python 3.11/3.12 or conda-forge.