Matthieu Denisselle (Magellium, France); Thomas Vaujour (Magellium, France); Noémie Lalau (Magellium, France); Michaël Ablain (Magellium, France)

The Surface Water Ocean Topography (SWOT) mission launched in December 2022 started a new era of spatial altimetry and hydrology. Its objectives are to characterise ocean mesoscale and submesoscale circulation, and to characterise spatial and temporal variations in surface waters. Two types of topography data are generated: Low Rate (LR) data over the oceans, with a spatial resolution from 250 m to 2 km, and High Rate (HR) data over inland waters, with a spatial resolution from 10 to 60 m. The study of glaciated regions, whether located in the open ocean or inland, still represents major scientific and technical challenges. However, the groundbreaking performance of the SWOT mission could allow us to study them in detail. Indeed, already available SWOT data show the potential for detecting ice over continental areas as well as sea ice.

This study develops an ice detection algorithm for lakes and rivers based on SWOT HR data. A first study showed the ability to discriminate between ice and water over lakes at the Swedish/Norwegian border using the Level 2 Pixel Cloud Product: by combining σ0, height and coherence information, water-filled cracks in ice layers (called leads) were detected. Based on these findings, segmentation algorithms were tested on a scene featuring lake Athabasca in Canada in May 2023, when the ice-cover started to break up in pieces. Unspervised machine learning algorithms were implemented, taking as input 2D images of σ0, height and coherence values. After some preprocessing steps, a Principal Component Analysis (PCA) followed by a clustering algorithm separates the points into several groups. Based on their σ0 and coherence values, each cluster in the image is classified as either ”ice” or ”water.” The output is a 2-D water and ice mask matching the sampling of the input products. In order to quantify the performance of each algorithm, a small dataset of hand-labelled Sentinel-2 optical images was created.

The results of this study demonstrate the potential of SWOT data to detect ice. The existing algorithm should be refined in the future by adapting it to SWOT version D products and making it more robust to different ice and water conditions. The updated algorithm could then be used to train a Machine Learning model able to detect sea ice. Being able to detect ice both on the ocean surfaces and on inland waters is a key issue for numerous scientific and socio-economic topics.