Inside the Sentinel‑2 Super‑Resolution Model of the EOMasters Toolbox
- Marco Peters
- Jun 22
- 3 min read
Hello EOMasters,
A while ago, I wrote a post about super‑resolving Sentinel‑2 imagery — what it is, why it matters. If you missed it, here’s the link: 👉 https://www.eomasters.org/post/super-resolution-of-sentinel-2-imagery
Since then, quite a few people have asked specifically about the super‑resolution model integrated in the EOMasters Toolbox, my plugin for ESA SNAP. So in this post, I want to walk through what the model actually does, where it comes from, and what happens behind the scenes when you click Run.
This isn’t a formal manual — more like a friendly tour through the internals.
Comparison of the original Sentinel-2 image and the super-resolved one
Where the model comes from
The Sentinel‑2 Super‑Resolution Model included in the EOMasters Toolbox is based on work developed in the EVOLAND Horizon Europe project. The model was created by Julien Michel.
The idea behind the model is simple and elegant:
Use real 5 m VENµS observations to teach a neural network how Sentinel‑2 pixels should look at 5 m resolution.
This means the model isn’t guessing or artificially sharpening — it’s reconstructing fine‑scale structure based on real paired data.
What actually happens when the operator runs
Even though the user interface looks simple, the model performs a surprisingly sophisticated sequence of steps. The most important part is the three‑stage inference process, which determines how the model generates the 5 m pixels.
To make this easier to understand, I’ve included the diagram below - it visualizes the three steps clearly:
Here’s what the model does:
1. Feature Extraction
The model first analyses the spatial and spectral patterns in the Sentinel‑2 input. This includes:
edges
gradients
textures
vegetation structure
This step transforms the raw reflectance values into a rich internal representation — essentially a learned description of what is happening inside each pixel.
2. Detail Reconstruction
This is the heart of the model. Using what it learned from thousands of Sentinel‑2 ↔ VENµS training pairs, the network reconstructs the fine‑scale structures that typically exist inside each 10 m or 20 m pixel.
This is not interpolation.
It’s not sharpening.
It’s not pan‑sharpening.
It’s predictive reconstruction based on real 5 m observations.
The model has learned how typical sub‑pixel patterns look in different landscapes — agriculture, forests, urban areas, water bodies, etc. So when it sees a Sentinel‑2 patch, it can infer the most likely 5 m structure that would have been observed by VENµS.
This is also where B11 and B12 come into play. The EVOLAND team extended the training dataset with SWIR patches so the model could learn realistic 5 m structures for these bands as well.
3. Learned Upsampling to 5 m
Once the internal representation is built and the fine‑scale structure is reconstructed, the model converts everything into a full 5 m reflectance image.
This step uses learned upsampling layers — not simple resampling. The output pixels are predicted, not averaged.
In practice:
each 10 m pixel becomes four 5 m pixels
each 20 m pixel becomes sixteen 5 m pixels
The result is a 5 m product that is sharper, more detailed, and still radiometrically meaningful.
L1C or L2A?
Short answer: L2A. The model was trained on surface reflectance, so L2A matches the training domain. L1C works, but you’ll see more variability.
Note
I’m not the creator of the model, and some parts of this explanation may simplify or slightly reinterpret the technical details. This post reflects my understanding of the Sentinel‑2 Super‑Resolution Model based on the available papers and documentation. For the full technical description and official references, please check the sources below or reach out to Julien Michel, who developed the model within the EVOLAND Horizon Europe project:
Wrapping up
The Sentinel‑2 Super‑Resolution Model in the EOMasters Toolbox is a great example of how research from EVOLAND Horizon Europe can flow directly into practical tools. Thanks to the work of Julien Michel, there is now a robust, physically meaningful way to enhance Sentinel‑2 imagery to 5 m — right inside SNAP,
Tschüss & Goodbye
Marco
Comments