About S2DSR

Satellite imagery forces a choice. The Copernicus Sentinel-2 satellites image Earth’s land surface every few days and their images are free, but each pixel covers 10 × 10 m of ground. Aerial and commercial imagery shows far more, but it is costly and taken far less often.

In Switzerland, both exist for the same ground. S2DSR learns from the places where both exist, and uses what it learns to bring satellite images to 2.5 m, sixteen pixels for every one.

S2DSR is learned super-resolution. A deep neural network sees a 10 m Sentinel-2 image and predicts the 2.5 m image of the same ground. It learns from many examples where the answer is known: Sentinel-2 images paired with swisstopo aerial imagery of the same places.

The areas you can explore were kept out of training, so what you see is how the models do on ground that is new to them.

There are two datasets, each with its own models.

• Red, green and blue

  Sentinel-2’s visible bands, paired with swisstopo’s SWISSIMAGE orthophotos.

  8 areas to explore

• Red, green, blue and near-infrared

  The visible bands plus near-infrared, paired with SWISSIMAGE RS aerial images, which record near-infrared too.

  6 areas to explore

Every result is compared with the aerial image of the same ground. The viewer shows the comparison as PSNR, in decibels: how close the result is to the aerial image, where higher is closer. Plain enlargement, which adds no detail, is shown beside it as the baseline to beat.

When you run a model in the viewer, your device computes the result again and checks that it matches the published one.

Super-resolution predicts detail; it does not observe it. Small objects can be softened, missed or drawn slightly differently from reality, which is why every result sits beside the aerial image it is measured against.

The models were trained and tested on Switzerland. How well they do elsewhere has not yet been measured.

S2DSR began as research into deep-learning super-resolution of satellite imagery at ETH Zurich, and is developed by an ETH Zurich student.

S2DSR is an independent project, not an official product of ETH Zurich. Contact details are in the imprint.

Contains modified Copernicus Sentinel data · Aerial imagery © swisstopo.