🚀 Fast multi-GPU super-resolution for Sentinel-2 imagery 🌍 — seamless weighted blending ✨ removes patch artifacts, while flexible input support (.SAFE 📂, S2GM 🛰️, GeoTIFF 🗺️) makes generating high-quality SR products effortless.

Info

• This Repo is currently udnergoing major revision
• It is functional for RGB+NIR images currently.
• Functionalities coming in the future:
  • Input many tiles at once
  • Pass more file types
  • Run SR for a selectable amount of bands

Description

This package performs super-resolution with any PyTorch or PyTorch lighning model for the Sentinel-2 10m bands (R-G-B-NIR).

Functionalities:

• The Input can be either:
  • a ".SAFE"
  • a ".SAFE.ZIP" zipped .SAFE folder, as they come for batched downloads from the OA-Hub
  • a "S2GM" folder as available from the S2 Global Mosaic Hub - WARNING: opensr-model does not work w/ mosiacs
  • any ".tif" file or similar that can be loaded by rasterio (4-band RGB-NIR)
• The following is performed automatically:
  • Patching of input images by selectable size (eg 128x128)
  • Super-Resolution of individual patches with provided model
  • writing of georeferenced output raster
  • overlapping and sigmoid weightning of patches by selectable pixel amount to reduce patching artifacts - See image
  • CPU, GPU and multi-GPU inference is supported via PL-Lightning
  • logging including example patches and a georef .tif excerpt example form the full SR
• Supported Models:
  • 'LightningModule': Any PL Lightning model with a .predict() or .forward() function. If this model type is passed, multi-GPU and multi-batch processing is activated, which leads to a significant inference speed increase.
  • 'torch.nn.Module': Any SR model with a .forward() function can be passed. The drawback is that for this model type, multi-GPU and multi-batch processing is not supported. This is therefore considerably slower.

🚀 Command-Line Usage

Once installed (pip install opensr-utils), you can run the large-file super-resolution pipeline directly from the command line.

Basic Example

opensr-run /data/sentinel2_tile.SAFE LDSRS2

This command will:

• Take the Sentinel-2 .SAFE folder as input
• Run the LDSR-S2 model for 4× super-resolution
• Save outputs in the same directory (with a placeholder .tif and temp folder for patches)

---

Full Example with Options

opensr-run /data/flood_scene.tif LDSRS2   --window_size 128 128   --factor 4   --overlap 12   --eliminate_border_px 2   --device cuda   --gpus 0 1   --save_preview

Arguments explained:

• root → Path to the input file or folder (.tif, .SAFE, or S2GM).
• model → Must be LDSRS2 for CLI usage (any other value falls back to a no-model pipeline).
• --window_size → Patch size in LR pixels (default: 128 128).
• --factor → Upscaling factor (default: 4).
• --overlap → Overlap in pixels to avoid patch seams (default: 8).
• --eliminate_border_px → Pixels cut at patch edges to reduce artifacts.
• --device → cpu or cuda (GPU).
• --gpus → List of GPU IDs (e.g. 0 1).
• --save_preview → Saves LR/SR preview images every 10% of progress.
• --debug → Process only ~100 windows for testing.

---

Run as Python Module

You can also execute via Python’s -m flag:

python -m opensr_utils /data/flood_scene.tif LDSRS2 --device cpu

🚀 Code usage

First, download a .SAFE tile from the Copernicus Browser, a Mosaic folder from S2G, or get your RGB-NIR .tif file ready.

1. install libraries

pip install opensr-utils
pip install opensr-model

2. Create Model - in this case our LDSR-S2 Model

# Instanciate Model
import opensr_model # import pachage
model = opensr_model.SRLatentDiffusion(config, device=device) # create model
model.load_pretrained(config.ckpt_version) # load checkpoint

3. Run large-scale Inference

#
import opensr_utils
sr_object = opensr_utils.large_file_processing(
			root=path,                 # File or Folder path
			model=model,               # your SR model
			window_size=(128, 128),    # LR window size for model input
			factor=4,                  # SR factor (10m → 2.5m)
			overlap=12,                # overlapping pixels for mosaic stitching
			eliminate_border_px=2,     # No of discarded border pixels per prediction
			device=device,             # "cuda" for GPU-accelerated inference
			gpus=[1,2,3],                   # pass GPU ID (int) or list of GPUs
			save_preview=True,        # save a low-res preview of the output, and a tif georef
			debug=False,
		)

sr_object.start_super_resolution()

Overlapping Strategy

In order to avoid patching artifacts that are present in many SR products, we perform a weighting based on the distance to the edge of the patches. It can be a linear weighting, but what works best is a sigmoid weight curve which puts leads to a more even edge. Additionally, in order to eliminate edge-artifacts that are present in many SR models, it is also possible to discard a fixed number of pixels along the edges of all patches.

Example of Patching Artifacts in other SR models:

Stats