­Updating Norway’s glacial lake inventory - an automated workflow using Sentinel-1 & 2 data and machine learning

Glacier lakes have been expanding globally in quantity and size due to accelerated glacier melt. In Norway, a growth in lakes has been seen in recent decades. The expansion and formation of glacier lakes pose significant risks, including the increased frequency of Glacier Lake Outburst Floods (GLOFs), which can impact downstream communities. Given the difficulty in accessing mountainous regions, the application of remote sensing is fundamental to monitoring glacier lakes for understanding the impacts of climate change and assessing the risks associated with GLOFs. However, there is no universally accepted definition of a glacier lake, which complicates regional comparisons. Standard mapping techniques for glacier lakes, such as thresholding the Normalized Difference Water Index (NDWI) from remote sensing data, face challenges due to cloud cover, terrain shadows, and ice cover variability. This has led to the use of manual or semi-automatic methods, often requiring labour-intensive post-processing to improve accuracy. Recent advancements in machine learning offer promising alternatives, enabling more efficient and accurate mapping by integrating multiple input data sources. However, existing methods still rely on digital elevation models (DEMs), which may not accurately reflect recent glacier retreat and the formation of new lakes. This study aims to address these limitations by developing an automated, reproducible workflow to update the glacier lake inventory of Norway using Sentinel-1 and Sentinel-2 imagery from 2023/24. We employ a random forest classifier trained on a 10th percentile Sentinel-2 summer composite without relying on DEMs. To mitigate misclassification, particularly due to mountain shadows, we propose a novel post-processing step that uses differences between ascending and descending Sentinel-1 images. Our fully automated workflow, implemented in Google Earth Engine and Python, is expected to improve the efficiency and reproducibility of glacier lake mapping. A comparison of the results with Norway's most recent glacier lake inventory from 2018/19 shows further glacier retreat with associated lake expansion and formation of new lakes. The method performs best in flat and low-lying glacier environments, whereas some manual editing is still needed in steep, high-alpine regions due to shadowing and year-round lake ice cover.