SYNAPSE-SR: Sentinel-2 super-resolution to 2 m resolution

Weights for synapse-sr, an open-source package that super-resolves Sentinel-2 L2A imagery from 10 m to 2 m resolution (B04 B03 B02 B08). A physics model of the Sentinel-2 sensor keeps every output consistent with the measurement; every pixel carries a support label (measured vs inferred) and a calibrated uncertainty. Docs: https://sharadhnaidu.github.io/synapse-sr/

Use

pip install synapse-sr

import synapse_sr

r = synapse_sr.super_resolve("sentinel2_l2a.tif")                  # Flash; model="pro" for Pro
r.save("sentinel2_2m.tif")                                         # georeferenced GeoTIFF
r.uncertainty()                                                    # calibrated expected error per pixel

synapse-sr --fetch 12.92,77.50 --dates 2025-01-01:2025-03-15 out.tif --preview preview.png

Weights download once and are SHA-256 verified. Offline: Flash.from_pretrained(weights="Flash/synapse-flash-v1.safetensors").

Benchmark

Official opensr-test protocol, mean over NAIP, SPOT, Spain urban, Spain crops and VENuS (178 scenes); synapse-sr 0.4.1 defaults.

SYNAPSE is best on five of the seven columns: Pro on improvement, spectral and reflectance error, Flash on detail correlation and (tied) RMSE.

How it works

x_hat = x_base + P_N(delta): x_base is a regularised inversion of the exact Sentinel-2 sensor model (per-band point-spread function); delta is predicted by the network; P_N projects it onto the part of the image the sensor cannot see, so the network cannot change what the satellite measured.

• Pro: Mamba state-space backbone (6 x 8 visual state-space blocks), 20 m spectral-context stem, frequency mixer, direct x5 PixelShuffle head. Default physics fit: tight (0.5 noise units), for the most recovered detail.
• Flash: re-parameterised convolutional network on all ten bands (trains as three-branch convolutions, folds to one 3x3 convolution per layer), trained by knowledge distillation from Pro. After the projection, each 10 m pixel's measured mean reflectance is restored exactly with a smooth correction.

Training data

• Sentinel-2 L2A paired with 0.6 m NAIP aerial imagery (951 real training pairs, United States).
• A streamed corpus of 3,044 native 1 m NAIP patches across 15 land-cover classes, observed through the exact Sentinel-2 sensor model, with composed structures (roads, buildings, field and water edges).
• ISRO Cartosat-2E and Cartosat-3 imagery (sample data from NRSC's Bhoonidhi portal): 1,001 Indian training tiles from three scenes, plus a separate held-out Cartosat-3 scene, observed through the Sentinel-2 sensor model.

Calibrated uncertainty

Measured coverage on held-out development pixels at the 80 / 90 / 95 % levels: Flash 82 / 91 / 95 %, Pro 82 / 91 / 96 %.

Files

Examples

Licence and acknowledgements

CC0-1.0. Third-party components and their licences are listed in THIRD_PARTY_NOTICES in the package. Sentinel-2 data: Copernicus programme, European Space Agency. Cartosat data: ISRO / NRSC (Bhoonidhi).