Spot the craters
The Moon’s surface is riddled with craters, creating hazardous terrain for lunar landings. To identify safe touchdown zones, we need to map these craters, from the largest to the smallest. NASA’s Lunar Reconnaissance Orbiter has been scanning the Moon since 2009, producing millions of images – but having humans looking for craters manually would take far too long.
This image shows a 1.25 x 1.27 km patch of the Moon near 44.12°N, 307.80°E. The red circles mark 3192 potential craters, ranging from approximately 6 to 180 metres across, detected by an artificial intelligence (AI) model developed by ESA.
To recognise craters, AI first needs trustworthy, labelled examples to learn from – but there is no publicly available, high-quality database of small lunar craters. This is where players of ESA’s IMPACT game enter the picture. While building and managing their own lunar base, players gather resources through a minigame in which they mark craters on real satellite images.
“Each crater marked by an IMPACT player is a small but important contribution to mapping the Moon,” says Patrick Bauer, a PhD student working on AI-based lunar crater detection using data from IMPACT. “Together, the players have created the data we needed to teach our AI model what to look for.”
These player contributions were used to train and evaluate ESA’s crater-detection model. The training dataset included 672 tiles, each measuring 512 x 512 pixels and containing craters labelled by the gamers.
Degraded craters with eroded rims are particularly challenging for the model to detect. Although the results are promising, the model is still being developed and its detections are not perfect. Future research will explore whether digital elevation models – 3D representations of the terrain – could improve their detection.
This 1024 x 1024-pixel tile shows a patch on the Moon covering an area of 1.25 x 1.27 km. The red overlays were produced by ESA’s crater-detection model. The tile is an extract from image M1219279265LC, captured by the narrow angle camera of NASA’s Lunar Reconnaissance Orbiter. The full image measures 52,224 x 5064 pixels and covers an area of around 400 km2.
