description: Super-resolve a Sentinel-2 scene from 10 m to 2 m resolution in three lines of Python with SYNAPSE-SR: GeoTIFF in, georeferenced 2 m GeoTIFF out.¶

Quick start¶

No installation needed to try it

runs the whole tour below on a free GPU. See Colab and Kaggle.

1. Install¶

pip install "synapse-sr[stac]"      # [stac] adds the Sentinel-2 downloader used below
synapse-sr --env                    # what hardware and kernels will be used

2. Get a scene¶

import synapse_sr

scene = synapse_sr.fetch_sentinel2(lat=12.9237, lon=77.4987,          # RV University, Bengaluru
                                   start="2025-01-01", end="2025-03-15", size_m=1280)

scene is a 10-band GeoTIFF on the 10 m grid, with its cloud mask saved next to it and applied automatically. Any Sentinel-2 L2A GeoTIFF of your own works the same way (see Inputs).

3. Super-resolve¶

r = synapse_sr.super_resolve(scene)

In a terminal or notebook you see a live progress bar and then a one-line summary:

✓ synapse-flash-v1: 640×640 px at 2 m in 0.9 s · consistency ≤ 9.72τ · 27% observation-determined

That line was measured on a laptop CPU (Intel i9, no GPU used). Pro (model="pro") processes the same scene in about 5 s on an A100 slice with the Triton kernel, the kind of GPU runtime Colab and Kaggle give you.

synapse_sr.super_resolve loads the default model (Flash) once and caches it. To choose the model and the device:

from synapse_sr import Pro, Flash

model = Flash.from_pretrained(device="cpu")         # default model: fast on any CPU, laptop or ARM machine
model = Pro.from_pretrained(device="cuda")          # most detail; GPU recommended
r = model.super_resolve(scene)
r = synapse_sr.super_resolve(scene, model="pro")    # or in one call

4. Look at it¶

r.summary()                                          # table: size, backend, consistency, support, time
r.show(["image", "x_base", "prior", "support"])      # needs matplotlib

Panel	Meaning
`image`	the 2 m result
`x_base`	what the 10 m observation alone determines
`prior`	what the network added; invisible to the sensor, so the observation neither confirms nor contradicts it
`support`	green: observation-determined · amber: medium · red: prior-dominated or masked

5. Save it¶

r.save("rvu_2m.tif")     # GeoTIFF, same CRS and bounds: B04 B03 B02 B08, ERRSCALE x4, SUPPORT

It opens directly in QGIS or ArcGIS, or with rasterio.

6. Use it¶

idx = r.indices()                                    # ndvi savi evi gndvi ndwi, + ndre ndbi nbr mndwi
fields = synapse_sr.boundaries(r, "field")           # crop-parcel edges
urban = synapse_sr.boundaries(r, "urban")            # building and road edges
err = r.uncertainty()                                # calibrated expected error per pixel

before = synapse_sr.super_resolve(synapse_sr.fetch_sentinel2(12.9237, 77.4987, "2024-01-01", "2024-03-15", size_m=1280))
flood = synapse_sr.change(before, r, "ndwi")         # two dates of the same place -> changed area in km²
print(f"{flood.area_km2:.3f} km2 changed")

See Applications for crop monitoring, urban analysis, water and disaster assessment.

Command line¶

synapse-sr scene.tif scene_2m.tif                    # Pro, progress bar, summary table
synapse-sr scene.tif scene_2m.tif --model flash --device cpu
synapse-sr scene.tif scene_2m.tif --json             # machine-readable summary on stdout
synapse-sr --models                                  # what is published

Next: Choosing a model and a device · Inputs and preprocessing.