Sentinel-2 Super-Resolution GUI

A simple, user-friendly desktop application to increase the resolution of Sentinel-2 satellite images. Based on the L1BSR project https://github.com/centreborelli/L1BSR.

This tool takes the standard 10-meter Blue, Green, Red, and Near-Infrared (NIR) bands and uses a pre-trained AI model to create a higher-resolution, 2x upscaled image. It's designed for users who want to easily enhance their satellite imagery without complex command-line tools.

News

• Windows binary for ONNX version created with PyInstaller in Windows 10 here

Screenshot

---

TL;DR (For the Impatient)

1. Install stuff: pip install pyqt5 torch safetensors rasterio pillow numpy
2. Get the Model: Download the REC_Real_L1B.safetensors file from Releases.
3. Run the App: python L1BSR-GUI.py
4. Load Data: Select the model file and your four Sentinel-2 band files (B02, B03, B04, B08). The app is meant to be used for Sentinel-2 data downloaded from Copernicus Browser. The band files requirements are:

• Image format: TIFF (16-bit)
• Image resolution: HIGH (original code asked for band files with less than 500px x 500px resolution but larger resolutions work, albeit slowly)
• Coordinate system: UTM (not tested on other systems available in Copernicus Browser but they might work)
• Layers (Raw): B02, B03, B04, B08.
• Layers (Visualised): True color (optional).

5. Process: Click the "Process" button.
6. Save: Save your new high-resolution image as a GeoTIFF or a JPG.

---

Features

• 2x Super-Resolution: Upscales 10m Sentinel-2 bands (B02, B03, B04, B08) to a higher resolution.
• Graphical User Interface: No command line needed! A simple point-and-click interface.
• Interactive Preview:
  • See a large preview of your final, upscaled image.
  • Click and hold on the preview to instantly see the "before" image for a quick comparison.
  • Use the scrollbars to pan across the entire image if it's large.
• Visual Sharpening: Optionally apply an Unsharp Mask filter to the preview and JPG output to make details pop.
• Flexible Output Formats:
  • GeoTIFF: Save the full 4-band (B,G,R,NIR) output with all geographic data preserved. This is the best option for scientific use.
  • JPG + JGW: Save a visually-appealing 3-band (RGB) image with an accompanying world file (.jgw) for use in GIS software.

Requirements

You will need Python 3 and the following libraries, which can be installed easily using pip:

pip install pyqt5 torch safetensors rasterio pillow numpy

For GPU Acceleration (Optional but Recommended): If you have an NVIDIA graphics card, make sure you install a version of PyTorch that supports CUDA. This will make the processing significantly faster.

Onnx inference

Included in this repository is an app variant that uses OnnxRuntime for inference (L1-BSR-GUI-ONNX.py). The user interface is the same but the model file used in this model is a Onnx file that can be found in the Releases. Onnx inference requires a similar but reduced set of python libraries:

pip install onnxruntime pyqt5 numpy pillow rasterio

How to Use

1. Place the Model File: For convenience, create a folder named trained_models in the same directory as the script and place the REC_Real_L1B.safetensors file inside it. The app will find it automatically on startup.

   - your_project_folder/
     |- L1BSR-GUI.py  (this script)
     |- trained_models/
        |- REC_Real_L1B.safetensors
   

   If you place the model elsewhere, you can select it manually using the "Browse..." button.

2. Launch the Application: Run the script from your terminal:

   python L1BSR-GUI.py

3. Select Input Bands: Click the "Select B02...", "Select B03...", etc., buttons to load your four input GeoTIFF files. A small preview of each band will appear. The files must have the same dimensions and geographic projection.

4. Process the Image: Click the big "Process (2x Super-Resolution)" button. Wait a few moments (it will be much faster with a GPU). When it's done, the upscaled RGB preview will appear in the output window.

5. Preview and Sharpen:

   • Click and hold the mouse on the output image to compare it with the original.
   • To make the image look sharper, check the "Sharpen" box and adjust the slider. This only affects the preview and the final JPG, not the GeoTIFF file.

6. Save Your Output:

   • Click "Save as GeoTIFF..." to save a georeferenced TIFF (4-band).
   • Click "Save as JPG + JGW..." to save a visually-ready 3-band image for reports or simple GIS viewing.

Attribution and License

This application's user interface and workflow are built upon the core model and concepts from the L1BSR project. The AI model architecture (RCAN) and processing logic are derived from their original work.

• Original Repository: https://github.com/centreborelli/L1BSR

This work, like the original it is based on, is licensed under the GNU General Public License v3.0.