Supraglacial lake monitoring using remote sensing data over Northeast Greenland

The Greenland Ice Sheet has already undergone significant mass loss in the last few decades and models predict these losses to increase over the coming century. This mass loss not only contributes to sea level rise, but also influences ocean dynamics, chemistry and biodiversity, while the additional glacial melt leads to further glacier velocity increase and destabilization of the ice sheet. The Northeast Greenland Ice Stream has experienced some of the most significant effects of climate change in recent decades. Having undergone an air temperature increase at a faster rate than the global average, Northeast Greenland has seen an increase in surface melt, basal melt and ice velocity. These changes have already led to the complete disintegration of the floating tongue of Zachariæ Isstrøm, one of the major marine-terminating glaciers in the region. While currently not undergoing retreat, there are indications that Nioghalvfjerdsbræ or 79N Glacier, which currently has the largest floating tongue in Greenland, will soon meet a similar fate. To better understand the current and future status of these glaciers, various observation techniques can be employed to more accurate quantify and monitor different glaciological parameters. Specifically, the development of supraglacial lakes, which store and drain surface meltwater, can be observed using remote sensing techniques, specifically with multispectral satellites. The amount of the meltwater stored in and lost from supraglacial lakes is not well estimated, specifically for regions such as Northeast Greenland, resulting in inaccurate estimates of surface meltwater contribution to glacial mass loss and meltwater influx to the ocean. In this thesis, various dimensions of supraglacial lake behavior are explored. Firstly, estimates of supraglacial lake area over Northeast Greenland are derived over the 2016 to 2022 summer melt seasons. To achieve this, the efficiency and adaptability of deep learning is utilized through the implementation of a U-Net convolutional neural network applied to Sentinel-2 imagery. Through this, large interannual variability in supraglacial lake area was found, with annual lake area peaks ranging from 67 km2 in 2018 to 380 km2 in 2019. Furthermore, the importance of a high quality cloud segmentation model in the creation of supraglacial lake area time series was highlighted through the incorporation of a deep learning-based model specifically trained to identify clouds over polar regions. Comparing the results of supraglacial lake outlines derived from deep learning methods to traditional thresholding methods, similar trends were found, while the deep learning method produced smoother results and eliminated a substantial portion of falsely classified shadows. This thesis additionally presents an analysis of different supraglacial lake depth estimation techniques. The effectiveness of using empirical relationships between depth and reflectance was evaluated through the development of two novel depth estimation methods. Based on in situ data gathered in Northeast Greenland, the first empirical method relates sonar-based depths to Sentinel-2 reflectance values. The second empirical method derives depths from 19 ICESat-2 crossings in Northeast and Southwest Greenland, which were also correlated to Sentinel-2 reflectance values. Through an evaluation of the red, green, and blue bands on the ICESat-2-derived data, a better suitability of the green band over the traditionally used red band was found. Furthermore, the influence of sediment present in supraglacial lakes on depth estimation was distinctly found through comparison of data gathered in Northeast versus Southwest Greenland. To evaluate the effectiveness of the two novel empirical methods, a comparison against a standardly used radiative transfer model as well as bathymetry derived from digital elevation models was conducted. The empirical methods were found to be effective up to depths of roughly 8 m, allowing the estimation of supraglacial lake depth to be simplified while retaining sufficient accuracy. Finally, in this thesis, the phenomenon of rapidly draining supraglacial lakes is investigated, as the influencing factors triggering a rapid drainage are not very well understood. Through this research, the spatial and temporal patterns of 152 rapidly draining lakes in Northeast Greenland were mapped and characterized. While supraglacial lakes at higher elevations tended to drain later in the season, other large-scale spatial and temporal patterns were not apparent. Various environmental and geophysical parameters were additionally compared with drainage timing. Investigation of the first and second principal strain rate of the ice yielded similar behavior for both rapidly and non-rapidly draining lakes. A theory of drainage upon reaching a water volume threshold was also dispelled through this research. Furthermore, while average surface temperature aligned with the number of rapid drainages in some years, an overall trend could not be identified. The further investigation of glacier ice dynamics and localized stress regimes should be conducted in combination with the previously mentioned influencing factors in order to more thoroughly understand rapid drainage patterns. Through this research, advances have been made in the study of supraglacial lakes by investigating supraglacial lake segmentation, depth estimation, and rapid drainage occurrence. The applicability of deep learning methods to supraglacial lake segmentation was demonstrated, highlighting its adaptability while acknowledging certain drawbacks. Furthermore, the specific use cases and limitations of four supraglacial lake depth estimation methods were discussed, while emphasizing the strong influence of lake bed sediment on depth estimation in multispectral imagery. Finally, by determining which factors do not singlehandedly influence the rapid drainage of supraglacial lakes, the direction of future research can be more productively focused.