CoolFace
Modelpublic

WEO-SAS/sen2sr

sourceHugging Facecc0-1.0updated 2mo agoView on Hugging Face
1likes68downloads
Model Card

<p align="center"> <img src="https://cdn-uploads.huggingface.co/production/uploads/6402474cfa1acad600659e92/G1o2oiRwJaqw4ZP9nG0NO.webp" width="100%"> </p>

<p align="center"> <em>Sentinel-2 super-resolution up to 2.5 m — WEO-SAS packaging of <a href="https://huggingface.co/tacofoundation/SEN2SR">tacofoundation/SEN2SR</a></em> </p>

This repository re-packages the original tacofoundation/SEN2SR models with the WEO-SAS standard interface (model.py, predictor.py, config.json) so they can be loaded and used identically to all other WEO-SAS models.

Original work: ESAOpenSR/sen2sr — license CC0-1.0.

Model Variants

Six variants are available as HuggingFace branches, each with a different architecture, input bands, and upscaling factor.

BranchArchitectureInput bandsOutput bandsScaleDescription
main (default)CNN4 (RGBN)4 (RGBN)4×SEN2SRLite — RGBN 10 m → 2.5 m
lite-rswir-x2CNN10 (all S2)6 (RSWIR)2×SEN2SRLite — 20 m bands → 10 m
lite-mainCNN10 (all S2)10 (all S2)4×SEN2SRLite — full 10-band pipeline 10 m → 2.5 m
mamba-rgbn-x4Mamba4 (RGBN)4 (RGBN)4×SEN2SR — RGBN 10 m → 2.5 m (higher accuracy)
mamba-rswir-x2Swin2SR10 (all S2)6 (RSWIR)2×SEN2SR — 20 m bands → 10 m (higher accuracy)
mamba-mainMamba + Swin2SR10 (all S2)10 (all S2)4×SEN2SR — full 10-band pipeline (highest accuracy)

Band order expected as input:

VariantBands
RGBN (main, mamba-rgbn-x4)B04, B03, B02, B08
All others (10 bands)B04, B03, B02, B08, B05, B06, B07, B8A, B11, B12

Installation

bash
# For CNN variants (main, lite-rswir-x2, lite-main)
pip install sen2sr safetensors huggingface_hub rasterio

# For Mamba/Swin variants (mamba-*)
pip install mamba-ssm --no-build-isolation
pip install sen2sr safetensors huggingface_hub rasterio

Usage

All variants share the same interface. Only the revision argument changes.

Load any variant

python
from huggingface_hub import snapshot_download
import sys

# Choose your variant:
local_dir = snapshot_download("WEO-SAS/sen2sr")                            # RGBN 4x (CNN) — default
local_dir = snapshot_download("WEO-SAS/sen2sr", revision="lite-rswir-x2") # RSWIR 2x (CNN)
local_dir = snapshot_download("WEO-SAS/sen2sr", revision="lite-main")      # Full 10-band 4x (CNN)
local_dir = snapshot_download("WEO-SAS/sen2sr", revision="mamba-rgbn-x4") # RGBN 4x (Mamba)
local_dir = snapshot_download("WEO-SAS/sen2sr", revision="mamba-rswir-x2")# RSWIR 2x (Swin2SR)
local_dir = snapshot_download("WEO-SAS/sen2sr", revision="mamba-main")     # Full 10-band 4x (Mamba+Swin)

sys.path.insert(0, local_dir)
from model import Model

model = Model(local_dir=local_dir)
print(model.description)

Array inference

python
import numpy as np

# image: (C, H, W) float32, values in [0, 1]  (C=4 for RGBN, C=10 for full-band)
image = np.random.rand(4, 128, 128).astype("float32")

sr = model.predict(image)   # (C, H*4, W*4) float32
print(sr.shape)             # (4, 512, 512)

GeoTIFF pipeline

Reads Sentinel-2 DN values directly (auto-normalises by /10000), writes a super-resolved GeoTIFF with the correct pixel size.

python
model.predict_tif(
    input_path  = "s2_scene_10m.tif",
    output_path = "s2_scene_2p5m.tif",
    bands       = [0, 1, 2, 3],   # 0-based band indices (default: first C bands)
)

Override config at load time

python
model = Model(local_dir=local_dir, patch_size=256, overlap=64)

RGBN 10 m → 2.5 m (main, mamba-rgbn-x4)

Super-resolves the four 10 m Sentinel-2 bands (Red, Green, Blue, NIR) by 4×.

<p align="center"> <img src="https://huggingface.co/tacofoundation/SEN2SR/resolve/main/assets/srimg02.png" width="100%"> </p>

Full 10-band 10 m → 2.5 m (lite-main, mamba-main)

Multi-stage pipeline: RGBN bands are super-resolved at 4×, while the 20 m bands (B05, B06, B07, B8A, B11, B12) are first sharpened to 10 m then to 2.5 m. All 10 bands are returned at 2.5 m.

<p align="center"> <img src="https://huggingface.co/tacofoundation/SEN2SR/resolve/main/assets/srimg01.png" width="100%"> </p>

RSWIR 20 m → 10 m (lite-rswir-x2, mamba-rswir-x2)

Sharpens the six 20 m Sentinel-2 bands (B05, B06, B07, B8A, B11, B12) to 10 m resolution using all 10 bands as context input.

<p align="center"> <img src="https://huggingface.co/tacofoundation/SEN2SR/resolve/main/assets/srimg03.png" width="100%"> </p>

Large image inference

For images larger than the 128×128 training patch size, predict_tif and predict automatically tile the input with overlapping patches and blend them seamlessly.

<p align="center"> <img src="https://huggingface.co/tacofoundation/SEN2SR/resolve/main/assets/srimg05.png" width="95%"> </p>

Repository structure

Each branch contains a flat directory with the same set of files:

config.json               # Variant-specific inference parameters
model.py                  # Public entry point (WEO-SAS standard)
predictor.py              # Tiled inference logic
sen2sr_pt.py              # HF-aware model loader (handles CNN / Mamba / Swin)
base.py                   # Abstract base class
model.safetensor          # Primary model weights
hard_constraint.safetensor# Hard-constraint weights
load.py                   # Original tacofoundation loading script
mlm.json                  # Original MLSTAC metadata
# multi-stage branches also include:
sr_model.safetensor / sr_hard_constraint.safetensor   (RGBN stage)
f2_model.safetensor / f2_hard_constraint.safetensor   (RSWIR 2x stage)

Citation

If you use these models please cite the original work:

bibtex
@software{sen2sr2024,
  author  = {Aybar, Cesar and others},
  title   = {SEN2SR: Sentinel-2 Super-Resolution},
  url     = {https://github.com/ESAOpenSR/sen2sr},
  year    = {2024}
}