Skip to content

Deep-learning super-resolution

Free satellite imagery, four times finer.

S2DSR uses deep neural networks to turn 10 m Sentinel-2 images into 2.5 m imagery. Its models learn from real satellite images paired with Swiss aerial imagery, and are tested on ground they have never seen.

The same view of Murten, super-resolved to 2.5 mMurten and Lake Murten as Sentinel-2 records them, at 10 mSentinel-2 · 10 mS2DSR · 2.5 m

Why S2DSR

Free imagery is everywhere. Fine detail isn’t.

Sentinel-2 images Earth’s land surface every few days, free of charge, but at 10 m a family house covers a pixel or two. Aerial imagery shows far more, yet it is costly and acquired only every few years. S2DSR brings the free imagery closer to aerial detail.

How it works

From 10 m to 2.5 m, checked against reality.

The same ground near Effretikon, three ways.

  1. A motorway junction and housing near Effretikon, as Sentinel-2 records themSentinel-2 · 10 m

    01

    The free satellite image

    Sentinel-2 records red, green, blue and near-infrared light at 10 m, and Copernicus makes every image free to use.

  2. The same junction, super-resolved to 2.5 mS2DSR · 2.5 m

    02

    Super-resolved by a neural network

    A deep network trained on thousands of Swiss image pairs predicts the 2.5 m image.

  3. The same junction in Swiss aerial imageryAerial image · reference

    03

    Compared with aerial imagery

    swisstopo’s SWISSIMAGE of the same ground, reduced to 2.5 m, is the reference every result is scored against.

Explore

14 areas across Switzerland.

Every area is held-out test data: the models never saw it while learning. Open one to compare it in the viewer.

See the difference yourself.

Swipe between Sentinel-2, three deep-learning models and the aerial image, or run a model on your own device.

Open the viewer