FAQ¶
Is the 2 m output "real" 2 m detail?¶
Partly, and synapse-sr tells you which part. A 10 m pixel cannot uniquely determine the 25 pixels of 2 m inside it. The output therefore has two components:
x_baseis fixed by the measurement;prioris inferred by the network and is invisible to the sensor.
support labels each pixel by which component dominates, and uncertainty() gives a calibrated expected error.
The output grid is 2 m. How much of the detail in any pixel is resolved from the measurement rather than inferred
is reported per pixel by support and uncertainty().
How is this different from bicubic, other super-resolution models or a GAN?¶
Bicubic adds no information. Other learned models, including GAN-based methods, add plausible detail but do
not tell you where it came from, and most can alter what the satellite measured. synapse-sr constrains the
network so that re-observing the output reproduces the measurement (consistency ≈ 1 noise unit). It also returns
the observed / inferred split with every result. You can check all of this yourself:
Verify it yourself.
Which bands come out at 2 m?¶
B04, B03, B02 and B08 (red, green, blue, near-infrared). The six 20 m bands (B05, B06, B07, B8A, B11 and B12)
help the network as spectral context. They are available on the output grid via r.band("B11") and in the
red-edge and SWIR indices, but they are not super-resolved.
Pro or Flash?¶
Flash (the default) runs anywhere, about 1 s per 1.28 km scene even on a laptop CPU. Pro adds the most detail and is best on a GPU (Colab, Kaggle, a workstation): super_resolve(src, model="pro").
The physics guarantees are identical. See Choosing a model and a device.
How long does it take?¶
Measured on a 1.28 km x 1.28 km scene (128 x 128 input pixels, 640 x 640 output):
| Hardware | Pro |
|---|---|
| A100 slice, Triton kernel | 5.2 s (first call 22 s, including the one-time compile and weight download) |
| Laptop RTX 4070, Windows, PyTorch scan | 80 s |
| Flash, laptop CPU (Intel i9) | about 1 s; a 10 km x 10 km scene in 14 s |
Time grows with area: a 10 km x 10 km scene has about 61x as many pixels.
Does it need the internet?¶
Only to download the weights once (~58 MB, SHA-256 checked) and, if you use it, for fetch_sentinel2. After that
everything runs offline, or from a local file with from_pretrained(weights=...). See
Offline use.
What input do I need?¶
A Sentinel-2 L2A (surface reflectance) scene on the 10 m grid that contains the 10 model bands, as a GeoTIFF, a numpy or torch array, or xarray. DN or reflectance are both fine, and band order is detected from names. L1C top-of-atmosphere data is accepted as input, but the model was trained on L2A. See Inputs.
Why are some pixels NaN or red in the support map?¶
NaN means the input was invalid there (NoData, cloud, cloud shadow, cirrus or saturation, from the SCL layer).
Red, or support 0, means the prior dominates, which is often the case at sharp edges and along the scene border,
where tiles have no context on one side. Use support == 2 for decisions that must rest on the measurement.
My result looks darker or brighter than expected.¶
This is almost always a radiometric offset problem. Products of processing baseline 04.00 and later store DN
with +1000. synapse-sr reads the BOA_ADD_OFFSET tag. If your file lacks it, pass offset=-1000 for raw DN, or
offset=0 for data where the offset was already removed (for example GEE S2_SR_HARMONIZED and Earth Search).
See also Troubleshooting.
Can I use it commercially? How do I cite it?¶
The package is CC0-1.0. Third-party components and their licences are listed in THIRD_PARTY_NOTICES. The citation is in the
README.