Gabriela Siles (Universite Laval, Canada); Samuel Foucher (Universite de Sherbrooke, Canada); Sylvain Biancamaria (LEGOS, France); Anthony Mandeville (Universite de Sherbrooke, Canada); Toumia Ghribi (Universite de Sherbrooke, Canada)

Lake and river ice play an essential role in northern latitudes, contributing to various human activities (such as transportation and fishing), weather regulation, aquatic ecosystems and as an indicator of climate change. Remote sensing data are increasingly used to monitor lake and river ice due to the scarcity of in situ observations, but significant challenges remain due to some limitations of optical and C, L, and X-band radar sensors. More accurate and frequent measurements of ice dynamics are also required. The SWOT satellite, with its Ka-band radar interferometer and high spatial and temporal coverage, offers an unprecedented opportunity to monitor lake and river ice. With SWOT data available since March 2023, we now have two whole winter seasons to study the signal on lake and river ice.

As backscatter (sig0) decreases but interferometric coherence remains high (or phase noise standard deviation remains low), ice-covered areas can be detected. However, the signal also varies with the incident angle, requiring the development of an algorithm that takes sig0, coherence, phase noise standard deviation,and incident angle into account. Pixel-by-pixel classification with a random forest allows the monitoring of the spatial variability of ice cover. Pixels are classified into four categories: Water (original SWOT classes 3 and 4), Darkwater (class 5), Low-Coherence (classes 6 and 7) and Ice. The figure below shows an example for Richardson Lake in Canada in spring 2025.

The pixel-by-pixel classification is then used in a second random forest to develop a dynamic ice flag for the HR_LakeSP product. The lake-scale classification was developed on lakes in Canada (15 lakes) and Norway (9 lakes) and tested on independent lakes (45 in Canada and 30 in Norway) during the 2024-2025 winter season, achieving an precision of more than 92% when compared to the visual interpretation of Sentinel-2 images, regardless of incident angle, lake size, or region. This classification will be applied to lakes in other regions (e.g. Alaska and Tibetan Plateau) in the next months to validate its global application. A similar classification for rivers is currently being developed at HR_RiverSP node level, again using pixel-by-pixel classification and random forest.