Antarctica, the coldest, windiest, and most remote continent on our planet, plays a crucial role in the global climate system. Its ice mass loss is a major driver of rising sea levels, with projections indicating a potential rise of several meters in the coming centuries. However, there remains considerable uncertainty about the future trajectory of Antarctic mass loss. A major area of uncertainty is the fate of ice shelves—floating extensions of land ice that surround much of Antarctica and act as barriers, slowing the flow of glaciers into the ocean. Ice shelves are affected by warm water from below, which thins them and increases their vulnerability to cracking, as well as by warm air from above, which melts the surface and forms ponds of meltwater.

This research focuses on surface melt, a phenomenon where meltwater forms and either refreezes or accumulates on the ice shelf surface. When the water accumulates, it can seep into cracks, causing them to deepen and widen, which can weaken the ice shelves. In today’s era of abundant satellite imagery and advanced deep learning techniques, we can efficiently process large volumes of data, enabling more comprehensive research on surface melt dynamics. The aim of this dissertation is to enhance the mapping and understanding of surface meltwater on Antarctic ice shelves using remote sensing and deep learning methods.

The introductory chapter provides an overview of the Antarctic Ice Sheet, emphasizing the continent's immense scale and importance. Written in an accessible style, it presents key concepts about Antarctica and explores how ice shelves and surface melt influence the continent. The chapter also describes the use of satellite data to map surface melt and discusses advancements in computational resources and deep learning, which have significantly improved our ability to analyze the expanding catalog of satellite data. It concludes with an overview of the research questions addressed in the thesis.

In the second chapter, various remote sensing datasets are compared to illustrate how and why satellite observations of surface melt differ. Using state-of-the-art melt detection algorithms, we analyze surface melt patterns and observe large differences, especially in icy areas, regions with subsurface melt, and during winter. These differences arise from factors such as satellite overpass times, spatial resolution, signal penetration, cloud cover, and detection methods. Despite these challenges, the variations create opportunities to combine data from multiple satellites, enhancing the overall accuracy of surface melt detection across Antarctica.

The third chapter builds on the previous findings and addresses the challenge of balancing spatial and temporal resolution in satellite observations. Surface melt in Antarctica is highly dynamic and varies regionally, making high-resolution mapping essential. To tackle this, we develop UMelt, a surface melt dataset for all Antarctic ice shelves with high spatial (500 m) and temporal (12 h) resolution, covering the period from 2016 to 2021. Our deep learning model integrates data from multiple satellites, allowing for detailed detection of surface melt while maintaining high temporal resolution. UMelt offers the potential for new insights into how ice shelves respond to changing atmospheric conditions.

In the fourth chapter, we shift from mapping the presence of surface melt to estimating its volume. Since surface melt is mainly driven by local processes, high-resolution regional climate models (RCMs) are necessary. However, current RCMs have a coarse resolution (25--30 km) that is insufficient for capturing small-scale melt processes. To address this, we introduce SUPREME, a deep learning method that downscales surface melt to 5.5 km resolution using a physically-informed super-resolution model. This model combines remote sensing data on albedo and elevation with a 27 km resolution Regional Atmospheric Climate Model (RACMO), accounting for the diverse drivers of surface melt across Antarctica. SUPREME demonstrates the potential of super-resolution techniques with physical constraints for high-resolution surface melt mapping, providing valuable insights into localized melting patterns.

The fifth chapter examines the hydrology of surface meltwater lakes on Antarctica, investigating whether they refreeze or drain into fractures at the end of the melt season, potentially destabilizing ice shelves. Monitoring these lakes with optical satellite imagery is often limited by cloud cover, complicating the tracking of their changes over time. To overcome this, we develop a spatiotemporal deep learning model using radar imagery from Sentinel-1, which allows us to classify the evolution of meltwater lakes regardless of cloud conditions. Our findings reveal no clear connections between lake evolution and ice shelf parameters, highlighting the need for further research and model refinement. The study is an initial step in using deep learning and Sentinel-1 data to monitor the evolution of supraglacial lakes on Antarctic ice shelves.

The sixth and final chapter reflects on the research and outlines future directions. It begins by summarizing the state of Antarctic surface melt research at the start of my PhD. The chapter then highlights the key contributions of this thesis and concludes with three proposed research ideas aimed at advancing our understanding of surface melt processes in Antarctica.@en

The melting of the Antarctic ice sheet is anticipated to play a significant role in sea level rise over the upcoming decades. Long-term mass and volume changes of the Antarctic ice sheet are predominantly caused by changes in the movement of the ice layer, referred to as ice dynamics. Mass and volume changes of the Antarctic ice sheet are monitored by gravimetry and altimetry satellites. Their data are corrected for glacial isostatic adjustment and changes in the surface climate, namely cumulated surface mass balance anomalies and firn thickness changes, to obtain ice-dynamical mass and volume time series. These time series are modelled using dynamic state-space models in this thesis.

State-space models decompose the data into several components. These components consist of parameters that are constant with time and states that vary with time. This thesis considers two modelling approaches to estimate the parameters and states, referred to as the the Frequentist and Bayesian approaches. The Frequentist approach entails estimating model parameters using maximum likelihood estimation and subsequently determining the states at each epoch using the Kalman Filter and Smoother. The Frequentist parameter estimates are deterministic. As a result, potential stochasticity of parameters is not accounted for when determining the states. This may lead to overconfident small uncertainties in Frequentist models. The Bayesian approach remedies this by considering model parameters to be stochastic. Following the Bayesian approach, the marginal posterior distributions of the parameters are estimated using Markov Chain Monte Carlo methods. Samples from these distributions are used to estimate the conditional distributions of the states at each epoch using the simulation smoother. Because parameter samples are used to estimate the states, stochasticity of the parameters is accounted for when sampling the states following the Bayesian approach.

Three distinct Antarctic ice drainage basins are investigated in this thesis. For their ice-dynamical ice mass data, significant differences between the uncertainties of Frequentist and Bayesian models are found. The Bayesian uncertainties are consistently larger than the Frequentist uncertainties. The largest differences between Frequentist and Bayesian uncertainties are found for the slope components of state-space models. Depending on the complexity of the data that are being modelled, the Bayesian uncertainty of a slope component can be up to 4 times larger than its Frequentist uncertainty.

Frequentist and Bayesian methods to combine the trends of gravimetry-based and altimetry-based state-space models of ice-dynamical ice mass in Antarctic ice basins are also investigated in this thesis. Epoch-wise weighted averaging, with weights based on the uncertainties in the gravimetry-based and altimetry-based models, is done to combine the trends. It is found that the gravimetry data is of significantly higher quality than the altimetry data, having much smaller uncertainties. As a result, the weighted average of the trends aligns closely to the gravimetry-based trend. Finally, several choices that have to be made when working with Bayesian Markov Chain Monte Carlo methods for state-space modelling and their impact on the results are discussed.

Satellite radar altimetry is often considered to be the most succesful spaceborne remote sensing technique ever. Satellite radar altimeters were designed for static geodetic and ocean dynamics applications. The goal of the geodetic mission phases, which have a dense ground-track spacing, is primarily to acquire information about the marine gravity field. This enables the estimation of mean dynamic topography (geographical sea surface height patterns due to ocean currents) and deep-ocean bathymetry. The primary goal of the oceanographic mission phases is to gain information about time-varying currents and ocean dynamics. TOPEX/Poseidon is the first altimetry mission to reveal sea surface height variations related to ocean dynamics as the El Niño Southern Oscillation (ENSO). During the mission it became clear that secular changes in sea level could also be monitored. Already in 1995, Nerem (1995) computed a Global Mean Sea Level (GMSL) time series from the TOPEX/Poseidon data. Currently, the GMSL record spans 26 years, in which TOPEX/Poseidon time series is extended with the Jason-1&2&3 observations. The estimated secular trend of GMSL over the altimetry era is approximately 3 mm yr−1. The succes of the TOPEX/Poseidon mission spawned the Argo project with the deployment of the first floats in the year 2000. One argued that Argo would support the future Jason missions in separating changes into the two components (density and mass) of sea level. The Argo project aims to estimate temperature and salinity over a depth of 2000 meter using floats, which enable the estimation of density or steric sea level changes. By subtracting the steric signal from the absolute sea level measured by Jason (steric-corrected altimetry), the second component of sea level changes, mass, is estimated. The launch of the Gravity Recovery And Climate Experiment (GRACE) satellites in 2002 made it possible to independently validate oceanic mass variations. If the sum of the mass and steric components equals total sea level within the uncertainties, the sea level is said to be closed. Besides these two oceanic components, ocean bottom deformation or Vertical Land Motion (VLM) also affects the sea level observed by altimeters. Over the open ocean VLM signals are generally small after a correction for Glacial Isostatic Adjustment (GIA), but near large mass variations they might become significant. Additionally, tide-gauge records are affected by VLM changes, because they are connected to land. Therefore they measure sea level relative to the sea floor, while the satellite altimeters observe the absolute variations. To bring tide gauges in the same reference frame as the altimeters, corrections for VLM have to be applied, which is usually done with nearby Global Navigation Satellite System (GNSS) data... @en

The Greenland Ice Sheet has a total volume of 2900000 km3. In recent decades, the ice-sheet has been losing mass rapidly and has nearly doubled its contribution to sea-level rise. One main contributing factor has been the recent widespread acceleration of the tidewater glaciers that terminate in deep and narrow glacial fjords. However, our understanding of the subsurface water properties causing this acceleration has been extremely limited owing to the lack of observations around these locations. We hypothesize that while the ice-sheet is not coupled to the ocean in the current General Circulation Models (GCMs), they can still be used to improve our understanding of the subsurface warming around these regions. In this study, we first evaluate two GCMs (HADGEM2-ES and GISS-E2-R) around three tidewater glacier locations, by interpolating the high resolution CTD data to the GCM grid. Our results substantiate that GISS-E2-R performs better than the HADGEM2-ES model along the Western Margin. However, along the Eastern Margin, we find that the HADGEM2-ES model is more consistent with the observations. The modelled spatio-temporal variability of water masses was investigated in a Pre-Industrial and Historical simulation, later used to quantify the evolution of subsurface warming under a future warming (RCP 8.5) scenario. Our results show that the Kangerlussuaq glacier (shelf) is more sensitive to greenhouse gas (GHG) forcings than Helheim and Jakobshavn. With respect to the Pre-Industrial climate, the mean warming of the Kangerlussuaq shelf under the RCP 8.5 scenario is 3.76°C, which is considerably greater than Helheim (2.38°C) and Jakobshavn (1.92°C). This is predominantly driven by the subsurface Return Waters of the Arctic Atlantic which are seen to warm the most under a future warming (RCP 8.5) scenario. For each of the three simulation, we decompose the Ocean Heat Content (OHC) time-series into intrinsic oscillatory modes so as to discern and quantify the high frequency variability from the multi-year/decadal variability and trends in the OHC. Where the contribution from the high-frequency modes to the total energy contained in the OHC signal is considerable in a Pre-Industrial scenario; the Historical and RCP 8.5 derived OHC are seen to contain a very dominant low-frequency response to GHG forcings, which contains (almost) all of the energy contained in the signal. For the RCP 8.5 scenario, we infer that at Kangerlussuaq, such a response is considerably greater than other locations. Furthermore, a consistent long-term increasing trend is seen in the upper ocean heat content for Kangerlussuaq, throughout the 21st century (1.01 × 10^8 J/m^2/month), unlike our inferences from other locations. This trend is found to be an order of magnitude higher than Helheim (2.36 × 10^7 J/m^2/month) and Jakobshavn (1.27 × 10^7 J/m^2/month). Our results also indicate that sea-ice free months at the Nares Strait (A.D. 2055 onwards) allows for a relatively greater mixing of the Lincoln Sea and Baffin Bay waters, which is consistent with the decline and stagnation of the rising OHC at Jakobshavn.

The main goal of this thesis involves the development of a refined methodology to
separate the mass change signals associated with glacial isostatic adjustment (GIA)
from those of surface ice/firn by exploiting the strengths of independent data sets,
such as those from gravimetry, altimetry, climate data, and others. To achieve this,
various research efforts were conducted addressing specific aspects of the methodology and subsequent data processing. This led to a number of new contributions to the topic,@en