Ice floes breaking off with a loud roar, deserts advancing, devastating forest fires — these are often the images citizens see when media cover climate change. Some people get frightened, others get tired or feel powerless. Master’s student in industrial design Frederik Ueberschär from TU Delft wanted to draw attention to climate change in a completely different way: by making use of the stunning beauty of the earth’s landscape as seen from high altitudes.

Ueberschär built a video installation called LANDSHAPES, in which he let artificial intelligence generate non-existent satellite images of the earth’s surface. He strung them together into short movies. In the movies you can see how green vegetation rapidly overgrows a desert, how a river meanders through a tropical rainforest like a snake, shifting its bed little by little, and how a land mass seems to eat into an ocean like a living organism. The images seem true to life, but are an artistic creation of artificially intelligent software. 

“In my research, I discovered that there is a need to experience what climate change is doing to the earth, without immediately imposing a political agenda”, Ueberschär says. “I wanted to stay away from causing fear of climate change, but instead create engagement through beautiful images.” 

Artificial intelligence invents new landscapes

To create the artificial earth observation images, Ueberschär used a special piece of artificial intelligence: a so-called Generative Adversarial Network (GAN), a technology first developed in 2014 to create artificial images. A GAN consists of two neural networks that play a match against each other. Each neural network mimics in a highly simplified way how human brains recognise and create images. 

One network in a GAN can be seen as a forger, the other as a detective. The forger learns to create an artificial image based on the training data. The detective determines whether the image is realistic or fake, that means whether it fits the training data or not. If the detective thinks it is fake, the forger creates a new image. This competition continues until the detective can no longer distinguish the artificial image from a realistic image from the training data. 

Ueberschär collected some 4,500 freely available high-resolution earth observation images from the ESA satellite Sentinel-2. The images came from all over the world and were very diverse: from deserts to sea ice, from oceans to rainforests, and from urban environments to river deltas. He then trained an existing piece of artificially intelligent software — StyleGAN2 — for about a hundred hours to recognise patterns in those images. After the training phase, Ueberschär had the StyleGAN2 network produce over thirty thousand completely new, non-existent satellite images. From these new images, he made short videos in which landscapes organically flow into one another. He projected these images in his LANDSHAPES installation.

“Due to the corona measures, I was limited in buying materials for my installation and had to limit the number of subjects for the final test to five”, Ueberschär explains. “The final installation was therefore somewhat simpler than I would like to make for a museum in the future. But the prototype was just right for carrying out the planned scientific experiments.” 

In the prototype, Ueberschär projected his images onto a wall and the test subjects were given a kind of joystick with which they could playfully change the landscapes themselves. “In the LANDSHAPES design that I conceptualized for museums, the images are projected in a circle on the ground and a kind of upright railing runs around the projection. Attached to that railing is a wooden handle that allows visitors to change the landscapes by moving around the installation, and discover its workings in a playful manner through trial and error.”