A Preliminary Study on the Detection of Glacial Lake Outburst Flood in Norway Using Sentinel 2 Imagery

Over recent decades, glacier retreat has altered dynamics and increased hazards, such as glacial lake outburst floods (GLOFs), leading to severe downstream flooding. While GLOFs are typically detected through on-site monitoring, existing remote sensing methods primarily map glacial lakes, without identifying GLOF events. This paper proposes an unsupervised approach to detect GLOFs using Sentinel-2 image time series. We train a deep network to model non-draining lake time series, preventing them from appearing as anomalies when drained. The model, based on convolutional Long-Short-Term Memory and a 3D convolutional neural network, reconstructs non-draining lake time series. Inference errors reveal draining lakes. Preliminary experiments on Norwegian glacial lakes show promising results for automated GLOF detection.