Conservationists want to know which species live where, so that sensors in remote areas can identify what they detect. Sightings crowd around roads and cities, leaving the places that matter most poorly known, and a model trained on them is only reliable where people have already looked.
This work trains a network on two billion species records, pairing each location with a numeric fingerprint of the satellite imagery covering it, and giving sparsely recorded regions more weight during training.
On regions held out up to hundreds of kilometres away, the satellite fingerprints improved prediction by 51% over location alone.