Inputs and preprocessing

Accepted inputs

Input	Example
GeoTIFF path	model.super_resolve("scene.tif")
numpy array (C, H, W)	model.super_resolve(arr, band_names=[...])
torch tensor (C, H, W) or (1, C, H, W)	model.super_resolve(tensor)
xarray DataArray with a band coordinate	model.super_resolve(cube.isel(time=0))

DN (L2A integer counts) and reflectance (floats up to about 1) are both accepted; reflectance is detected automatically. NaN and infinite values are treated as NoData.

Bands

The model uses ten bands:

Role	Bands	Native resolution
Super-resolved	B04, B03, B02, B08	10 m
Spectral context	B05, B06, B07, B8A, B11, B12	20 m

The bands are identified as follows:

1. From band_names, or from the GeoTIFF band descriptions. Names are case-insensitive, and B2 and B02 are equivalent.
2. Otherwise, from the channel count:
   • 10 channels are read as the order above;
   • 12 channels as the L2A order B01 to B12 without B10;
   • 13 channels as the L1C order.

Any other layout raises a ValueError that names the accepted orders.

The 20 m bands should be on the 10 m grid with nearest-neighbour resampling. This is how Earth Engine, fetch_sentinel2 and gdalwarp -r near deliver them. They inform the network but are not returned at 2 m.

Radiometry

Reflectance is (DN + offset) / 10000. Products from processing baseline 04.00 onward, from January 2022, carry BOA_ADD_OFFSET = -1000:

Source	What to pass
fetch_sentinel2 output	nothing: the offset is stored in the BOA_ADD_OFFSET tag and applied automatically
Earth Engine COPERNICUS/S2_SR_HARMONIZED	nothing: the offset is already removed
SAFE / JP2 from the Copernicus Data Space, baseline 04.00 or later	offset=-1000 (CLI: --offset -1000)

Masking

Source	Masked
NoData	the raster NoData value, NaN / inf, all-zero pixels
Scene classification (SCL)	classes 0 no data, 1 saturated or defective, 3 cloud shadow, 8 and 9 cloud, 10 cirrus

from synapse_sr import Pro
model = Pro.from_pretrained()

result = model.super_resolve("scene.tif", scl="scene_SCL.tif")   # 10 m or 20 m SCL
result = model.super_resolve("scene.tif")                        # uses scene_scl.tif next to the input if present
result = model.super_resolve("scene.tif", scl=None)              # no SCL masking

Masked pixels are written as NaN and receive support class 0. No heuristic cloud detection is performed.

Grid checks

GeoTIFF inputs must be north-up with square pixels of about 10 m. Rotated or sheared transforms, non-square pixels and other resolutions are rejected with an explanatory error. They are not silently resampled.