With accelerating climate change, the global ice sheets are melting at ever-increasing rates. Between 1992 and 2018, the Greenland ice sheet (GrIS) lost 3902 ± 342 billion tonnes of ice, equivalent to ~10.8 mm of global sea-level-rise (SLR) [1
]. Similarly, Antarctica lost 2720 ± 1390 billion tonnes of ice during 1992–2017 and contributed ~7.6 mm to global SLR [2
]. While ice mass loss on the GrIS due to glacier dynamics and particularly surface melting has been monitored and discussed extensively (e.g., [3
]), the impact of surface melting on mass changes on the Antarctic ice sheet (AIS) has not been investigated in sufficient detail. As the AIS holds ~91% of the global ice mass and represents the largest uncertainty in current projections of global sea-level-rise, it is yet of essential need to understand the response of Antarctic outlet glaciers to further increasing surface air temperatures resulting in enhanced surface melting and the formation of an increasing number of supraglacial lakes in local surface depressions of the ice sheet. Supraglacial meltwater accumulation can impact ice sheet dynamics and mass balance through three main processes (A-C, Figure 1
]. First, meltwater injection to the glacier bed can alter the basal conditions and enhance sliding ultimately leading to transient ice flow accelerations and increased ice discharge (A, Figure 1
], as observed over the Greenland ice sheet (e.g., [6
]). Second, surface melting and runoff causes ice thinning which directly affects surface mass balance (SMB) (B, Figure 1
]. The third process results from the repeated filling and draining of supraglacial lakes into fractures or crevasses initiating their downward propagation and consequent removal of ice shelves or small icebergs, also referred to as hydrofracturing (C, Figure 1
]. Ice shelf removal causes rapid ice flow accelerations and increased ice discharge through the loss of the efficient buttressing force exerted on the inland ice, as documented after several ice shelf collapses along the Antarctic Peninsula (API) (e.g., [12
]). Another important aspect to consider are positive feedback mechanisms that further trigger surface melting, e.g., through an increased absorption of solar radiation hence a decreasing surface albedo caused by the increasing abundance of supraglacial meltwater (D, Figure 1
]. Similarly, increased surface melting leads to an increasing appearance of rock outcrop which again decreases surface albedo and enhances ice melting and lake ponding (D, Figure 1
). Figure 1
illustrates supraglacial meltwater features at an ocean-terminating outlet glacier as well as the main processes acting on ice sheet dynamics and mass balance.
To enable more detailed analyses on the effects of supraglacial meltwater accumulation on Antarctic ice dynamics, mass balance, or ice shelf stability, a thorough mapping of the Antarctic surface hydrological network is required. For this purpose, spaceborne remote sensing is an ideal tool and enables both, unprecedented spatial coverage and high temporal resolution compared to ground-based mapping efforts. In addition to several local- and regional-scale investigations using manual to semi-automated mapping techniques (e.g., [18
]), only few studies implemented algorithms for either automated or continent-wide mapping of Antarctic supraglacial lakes in optical satellite imagery. Of these, Moussavi et al. [25
] used optical Landast 8 and Sentinel-2 data in combination with fixed band and index thresholding to automatically extract supraglacial lake extents and volumes over four East Antarctic ice shelves. Next, Halberstadt et al. [26
] employed unsupervised k-means clustering on Landsat 8 scenes over two East Antarctic ice shelves to train a suite of supervised classifiers and Dell et al. [24
] advanced the “FAST” (Fully Automated Supraglacial lake area and volume Tracking) algorithm [27
] to automatically track supraglacial lakes in Landsat 8 and Sentinel-2 imagery over Nivlisen Ice Shelf, East Antarctica. Finally, our companion paper [28
] presented an automated Sentinel-2 supraglacial lake classification algorithm developed on basis of a random forest (RF) classifier and was tested on spatially and temporally independent acquisitions distributed across the Antarctic continent. While these studies rely on the use of optical multi-spectral data, synthetic aperture radar (SAR) data are currently used for mainly visual interpretation and evaluation of Antarctic supraglacial lakes (e.g., [22
]). In fact, an automated or spatially transferable supraglacial lake detection algorithm using SAR data is entirely missing. Given the advantage of SAR data, e.g., for detection of subsurface lakes or considering its year-round as well as all-weather imaging capabilities, the development of an automated SAR-based mapping method is overdue. In this context, year-round image acquisitions are particularly suitable for analyses on intra-annual lake dynamics revealing whether lakes refreeze or drain at the onset of Antarctic winter as well as for delivery of complementary mapping products to optical lake extent classifications.
In order to address the lack of a SAR-based mapping technique, the main aim of this study is to develop the first automated supraglacial lake classification algorithm using Sentinel-1 SAR data. The Sentinel-1 constellation consists of two polar-orbiting satellites operating at C-band (5.405 GHz center frequency) and a revisit frequency of six days at the equator and up to daily in polar regions [30
]. Over Antarctica, Sentinel-1 operates in either extra wide (EW) swath or interferometric wide (IW) swath acquisition mode delivering data products at 40 m and 10 m pixel spacing and in HH/HV and HH polarizations, respectively. However, the detection of supraglacial lake features in radar imagery over Antarctica can be a difficult task using traditional threshold- or segmentation-based techniques even if high spatial resolution Sentinel-1 IW products are available. In fact, previous experiments with machine learning classifiers including random forest or support vector machines and single-polarized Sentinel-1 imagery at 10 m pixel spacing revealed inaccurate classification results even considering temporal image metrics. As can be seen in Figure 2
, Antarctic supraglacial lakes do not always appear with sharp round boundaries and homogeneous low backscatter (Figure 2
a) but instead with fuzzy lake edges (Figure 2
b,f), low visual contrast (Figure 2
c), strongly varying shapes and sizes (Figure 2
a) or strongly varying backscattering values ranging from approximately −20 dB to −40 dB, e.g., in the case of shallow slushy lakes (Figure 2
b,f,g) or when lakes are covered with a thin layer of ice or small floating icebergs (Figure 2
c–g,i,j). In addition, speckle noise as well as wind may further roughen the appearance of lake surfaces in radar imagery. Due to the inhomogeneous appearance of supraglacial lakes, surface features with similar shapes and sizes as well as low backscatter signatures such as open cracks and fractures on ice tongues (Figure 2
l), small blue ice or wet snow patches (Figure 2
h,k,m), topographically induced shadowing (Figure 2
o) or wet/dry snow on crevasse fields and fractured ice tongues (Figure 2
n,p) can be difficult to differentiate from supraglacial lakes or streams if no additional image information (e.g., different radar polarizations, temporal information, elevation data) is available. Another difficult surface feature are drained lakes leaving round wet patches in the snow.
To overcome most of the described issues, we propose the use of a convolutional neural network (CNN) allowing to consider the spatial image context through convolutional extraction of feature maps from a given input image. CNNs are particularly suitable where traditional classification methods fail and were commonly used in computer vision and particularly for semantic segmentation and object detection tasks in remote sensing [31
], e.g., for automated calving front detection using SAR and optical satellite imagery [33
], for sea-land classification in optical satellite imagery [36
], for multi-temporal crop type classification in optical Landsat imagery [37
], for sea ice concentration estimation in dual-polarized RADARSAT-2 imagery [38
], for automated mapping of ice-wedge polygons in high-resolution satellite and unmanned aerial vehicle images [39
], or for building and road extraction in optical and aerial satellite imagery [41
], to only name a few. In this study, we extract supraglacial lake extents from single-polarized Sentinel-1 imagery using a modified version of a CNN originally developed for semantic segmentation of biomedical images (U-Net) [43
]. In Earth Observation, encoder-decoder designs and particularly variants of U-Net are most commonly used for semantic image segmentation due to their better performance compared to other approaches including naïve-decoder models [31
]. With the main aim of this study being the derivation of lake areas by means of pixel-wise classifications, object detection approaches including multi-task instance segmentation (e.g., [39
]) were not considered or tested. In detail, we aim at the segmentation of open water lakes as well as lakes that are roughened at their surface, e.g., by wind or return higher backscatter values due to a thin ice cover and appear with fuzzy edges, low contrast and speckle noise. Lake regions and subsurface lakes that are entirely covered with thick ice or lakes that are subject to very severe wind roughening are not considered in this analysis. Particular focus during method development was the spatio-temporal transferability of the segmentation algorithm. Another aim of this study is to demonstrate the importance of fusing Sentinel-1 and Sentinel-2 supraglacial lake classifications for retrieval of more complete lake extent mapping products for a given time period. In fact, the exploitation of the advantages of both sensor types conjointly is particularly crucial in Antarctica where polar darkness during austral winter as well as a frequent cloud coverage oftentimes restricts the availability of optical data and where SAR data may be limited by the described effects in order to capture the full picture of lake occurrences.
In the following, Section 2
describes the selected study sites and datasets as well as the implemented methodology for data pre-processing, deep learning model training, post-processing, and the accuracy assessment. Following this, Section 3
presents mapping results as well as the outcome of the accuracy assessment and Section 4
discusses the results as well as remaining limitations of our algorithm. Finally, Section 5
summarizes the findings of this paper.
This study for the first time performed an automated mapping of Antarctic supraglacial lake extents in single-polarized Sentinel-1 SAR imagery using a convolutional neural network. In detail, we modified a U-Net with atrous convolutions and residual connections for semantic segmentation of supraglacial lake features. The main aim during algorithm development was the spatio-temporal transferability of the classification method in order to enable automated circum-Antarctic mapping efforts in the future as well as to provide complementary mapping products for optical supraglacial lake extent classifications. The deep learning network was trained on 57 Sentinel-1 acquisitions covering 13 training regions and evaluated by means of ten spatially and temporally independent testing regions as well as one additional region over the Greenland ice sheet. Post-classification mainly involved the masking of false lake classifications using a Sentinel-1 derived coastline as well as TanDEM-X topographic data.
The mapping results reveal the good functionality of our workflow with reliable lake classifications for all test data where supraglacial lakes were found to vary strongly with respect to sizes, shapes and appearances. Additionally, testing our supraglacial lake detection method in a study region in Southwest Greenland highlighted its potential for spatio-temporal transferability. An intra-annual analysis of supraglacial lake dynamics over the course of the 2019/2020 melting season at northern George VI Ice Shelf, Antarctic Peninsula, moreover revealed that supraglacial lake coverage peaked in mid-January and fluctuated on lower levels in the month before and thereafter. In addition, we present the first supraglacial lake extent mapping product for January 2020 over northern George VI Ice Shelf. The mapping product was computed by fusing automatically derived Sentinel-1 lake classification maps with the corresponding Sentinel-2 maximum lake extent classification, generated as part of our companion paper, and highlights the importance of considering both, SAR and optical data in order to capture the most complete picture of supraglacial lake formation. Overall, the fused classification product revealed a total supraglacial meltwater coverage of ~770 km2 for January 2020. Finally, the accuracy assessment returned an average Kappa coefficient of 0.925 as well as a of 93.0% for the water class across all Antarctic test sites. In this context, the main remaining limitations of our workflow were identified to be (1) the lack of high-resolution topographic as well as up-to-date coastline data, e.g., resulting in few misclassifications over open ocean, (2) misclassifications over radar shadow and small blue ice and wet snow patches, and (3) the lack of ground truth data for improved evaluation of classification products.
Overall, the results of this analysis highlight the suitability of the developed modified U-Net for supraglacial lake classification in single-polarized Sentinel-1 SAR imagery. Future developments mainly involve the integration of more training data covering open ocean as well as blue ice and wet snow patches, the circum-Antarctic and intra-annual mapping of supraglacial lake extents using the proposed method as well as an integrated analysis with data on ice dynamics and climate parameters. This will be crucial in order to assess the impact of the Antarctic surface hydrological network on ice dynamics and mass balance; thus, global sea-level-rise as well as to analyze effects of albedo changes that may trigger further surface melting. Lake extent mapping products will be generated for intra-annual analyses at both, sub-monthly and monthly resolution. For derivation of monthly supraglacial lake extents, the fusion of optical Sentinel-2 and Sentinel-1 SAR supraglacial lake extent mapping products will aid in obtaining a more complete picture of meltwater ponding. On the other hand, analyses at sub-monthly resolution will provide important insight into the spatio-temporal behavior of supraglacial lakes and enable conclusions about their refreezing or draining, e.g., at the onset of Antarctic winter.