Satellite data provided by the European Space Agency
Schindler’s research group has plenty of experience with satellite images. It uses them to predict population density in crisis areas, to determine war damage to buildings in Ukraine and to measure the height of forests around the world. But how does an artificial intelligence read snow depth off satellite images?
First, it needs millions of examples: for their technology, the researchers used optical and infrared images taken by Sentinel-2 satellites operated by the European Space Agency (ESA). These satellites capture every location on Earth every five days with a resolution of up to 10 by 10 metres per pixel, making them the most detailed images currently available free of charge and accessible by everyone. By examining these images, the AI can tell when and where Switzerland has snow on the ground and how the snow line changes from week to week.
But that alone is not enough, Schindler says: “Just looking at the white bits on the satellite images doesn’t immediately tell us how deep the snow is. For that, we need more data.”
Learning by reality comparisons
In addition to the satellite images, the researchers therefore also fed their AI data on Switzerland’s terrain. This is because more snow will melt on a steep south-facing hillside than in a shady hollow. Detailed terrain data of this kind is readily available from the public data published by swisstopo, the Swiss Federal Office of Topography.
The researchers trained their AI system to determine the snow depth based on this combination of satellite and terrain data. This involved asking the system to estimate the depth of the snow and then comparing the results with real measurements. “For each grid point, we note how far off the AI estimate was and gradually tweake the system so that the errors became smaller,” Schindler explains. The technical term for this is supervised learning.
For the first round of training, the ETH researchers used snow maps provided by ExoLabs, which matched up very well with the snow data from the Swiss meteorological stations. These maps use images not only from Sentinel-2 satellites but also from other satellite missions – which offer lower spatial accuracy but do provide daily images. Using the snow maps from ExoLabs, the AI was able to learn the detailed snow distribution patterns that cannot be captured by the rather loose-knit network of meteorological stations.
Fine-tuning using data from Dischma
The AI was then fine-tuned using extremely detailed snow data, which the Swiss Federal Institute for Forest, Snow and Landscape Research WSL collects only in the Dischma valley in eastern Switzerland. From this data, the AI learned how snow depths can change within just a few metres depending on the terrain. It can now apply these spatial relationships across Switzerland and produce accurate snow depth estimates even for those places where no detailed measurement data is available.
Another benefit of the new technology is that it also provides users with an indication of how certain they can be about the estimate. For example, if the weather has been overcast for a while and new satellite images provide no usable data, the uncertainty of the estimate increases.
