A deep neural network for lake ice detection with Sentinel-1 data

Abstract

Ice cover of lakes is an indicator of climate conditions and possible changes thereof. It is therefore identified as an essential climate variable, and tracking its worldwide timing, duration and extent is important. Due to the vast number of lakes on Earth however, it can be difficult to find efficient ways to continuously monitor the formation, duration and break-up of lake ice. Remote sensing can be a useful tool in that regard, but optical passive remote sensing can be hindered by the presence of clouds or night-time. In this study, the use of synthetic aperture radar (SAR) imagery is therefore proposed, an active system that can penetrate clouds and works both day and night. Because ice conditions can vary strongly through space and time, a fully convolutional network (FCN) is constructed. This deep learning network is specifically designed for semantic image segmentation: learning patterns from large amounts of data and assigning labels to each pixel in the imagery. The model is trained on four study areas from different parts of the world, and overall results show a mean accuracy of >80%. Predictions are better for non-frozen conditions ( ̴90%) compared to frozen conditions ( ̴72%). Slight overfitting of the data indicates that the use of additional study areas may be required to optimize model performance, but the overall results are promising and demonstrate the usefulness of its application in worldwide lake ice monitoring.