[LUM#22] In Search of Lost Glaciers
To study glaciers, geologists rely on satellite imagery. This is a long and tedious process, but Isabelle Rocamora aims to simplify it by developing the very first deep learning model designed specifically for mapping glacial moraines.
Most of the fresh water on the surface of our planet is found neither in rivers nor in lakes, but in glaciers. While there are still nearly 200,000 of them, many have disappeared, and those that remain are melting away like snow in the sun. During the 20th century, glacier coverage decreased by 30 to 40% in the Alps, by 75% in the Pyrenees, and by more than 80% in certain tropical mountain ranges such as the Andes and Kilimanjaro.
As glaciers retreat, they still leave a trace of their passage: geological formations called moraines. “It’s a pile of rock debris carried by the glacier’s sliding motion that accumulates at the bottom and on the sides,” explains Isabelle Rocamora. “It’s like when you push sand with your hand and then pull it back—it leaves ridges of sand in front and along the edges,” illustrates the doctoral student, who is conducting interdisciplinary research between the Géosciences Montpellier1 and Tetis, under the supervision of Matthieu Ferry and Dino Ienco. Mapping these moraines therefore provides valuable information, particularly regarding glacial retreat.
Automate tasks
A necessary but particularly time-consuming task: “To create this map, we’ll take measurements in the field and then look for glacial moraines in satellite images, which provide a broader view, ” explains the geologist. To save time, she proposes automating this task using artificial intelligence. “I’ve developed an algorithm that analyzes satellite images; it’s the first deep learning model dedicated to mapping glacial moraines.”
To teach her software, called MorNet, to distinguish a moraine from other rock formations, Isabelle Rocamora began by creating a “traditional” map using satellite images of the Himalayas acquired by the CNES’s Pléiades satellite. “I then fed these maps into the software, telling it ‘this is a moraine’ or ‘this is not a moraine.’ ” The model then identified the common characteristics among the samples in order to learn on its own how to identify a moraine in satellite images. “A computer doesn’t ‘see’ things the way a human does, so rather than giving it human-provided answers, we’ll let it discover its own identification rules, ” explains Isabelle Rocamora.
Background information
To assess MorNet’s performance, the geologist then compared her own mapping with that produced by the artificial intelligence. While MorNet proved effective at recognizing the ridges of moraines, it is less accurate at identifying their flanks; in other words, the model knows where these geomorphological features are located but struggles to delineate them clearly. And for good reason: a moraine is defined not only by its shape but also by a formation pattern, and MorNet cannot see that; to improve it, it would therefore need to be provided with contextual information. “While deep learning saves time, mapping still requires the eye of a geomorphologist to confirm and refine the machine’s conclusions, ” concludes Isabelle Rocamora.
UM podcasts are now available on your favorite platform (Spotify, Deezer, Apple Podcasts, Amazon Music, etc.).
- Principal Investigator (UM, CNRS, University of the Antilles)
↩︎