The sea ice on the oceans in the Arctic and Antarctic is a relatively thin blanket that significantly influences the exchange between the ocean and the atmosphere. The sea ice thickness is a major parameter, which is of great importance for diagnosis and prediction. Determining seasonal and interannual variations in sea ice thickness was the primary objective of ESA's CryoSat Earth Explorer mission. ESA's second Earth Explorer mission, SMOS, provides L-band brightness temperature data that can also be used to infer the thickness of the sea ice, although that was not its primary objective. Both missions complement each other strongly in terms of spatiotemporal sampling and their sensitivity to different ice thickness regimes. In order to further improve the synergistic use of low-frequency radiometric data for sea ice applications, it is imperative to better characterize the uncertainties and covariances associated with the retrieval. A key factor is a thorough understanding of the physical processes that determine the emissivity of sea ice in order to improve the forward model used for retrieval. A thermodynamic model is used to estimate the vertical temperature profile through the snow and sea ice. Therefore, additional meteorological data such as from atmospheric reanalyses and parameterizations of snow and sea ice properties must be taken into account. Natural sea ice is not a homogeneous medium of uniform sea ice and snow thickness, but can only be described by statistical distribution functions on different spatial scales. Thin ice and open water in leads within the compact pack ice also have a significant influence on the brightness temperature measured by SMOS. In order to take all these effects into account, the forward model or the observation operator must be of the appropriate complexity. The inversion to determine the geophysical sea ice parameters can be optimized with a-priori information and parameterizations as well as with information from other satellite sensors. The presentation will focus on a review of the current retrieval method used to generate the AWI-ESA level 3 and level 4 Sea Ice Thickness products and the way forward to improve the emissivity model and to define a common basis metrics validation to assess algorithms evolution considering that in-situ validation data is only sparsely available.
In the wake of climate change, Arctic sea ice has decreased significantly in both extent and thickness in recent decades, with this trend being particularly evident in the summer months. Satellite-derived products in the polar regions - such as sea ice thickness and concentration - are based on data collected at different spatial and temporal scales. Accurate sea ice parameter estimation requires a model framework to effectively extract the sea ice composition information contained in the various satellite observations. However, direct analyses of sea ice are limited to a characteristic length scale similar to the resolution of polar-orbiting satellites acquiring data in the low-microwave spectrum (∼10-50 km). As a result, models rely largely on empirically determined small-scale sea ice properties that are assumed to be captured in space-based observations, while operational sea ice products often depend on manual categorization by experienced analysts.
Data-driven approaches have provided promising results by harnessing the sensitivity ranges of multi-source satellite data for sea ice parameters. However, these approaches are often a purely statistical analysis of consecutive observations where the information is obtained only from a temporal, pixel-based view. Without taking into account the semantic meaning of the data sets during analysis, automated models lack generality to obtain consistent solutions with sufficient stability. Thus, given the dynamic nature and complexity of sea ice, geostatistical analysis is required to integrate both the temporal and the spatial interactions into the model.
In this work, a three dimensional segmentation approach based on Bayesian inference is used to segment Arctic sea ice in an unsupervised manner in a complete spatio-temporal context. The statistical properties are accounted for by a Gaussian Mixture Model, and both the spatial interactions and temporal variability of images are reflected using Hidden Markov Random Fields. Hereby, passive microwave observations from the Soil Moisture and Ocean Salinity (SMOS) and Advanced Microwave Scanning Radiometer 2 (AMSR2) missions are used as input features to examine the joint effect of CIMR-equivalent set of brightness temperature data at 1.4, 6.9, 10.6, 18.7, and 36.5 GHz. The goal was to segment the Arctic region into a number of relevant classes based on the synergy effects of the multiple observations, and to verify the developing class patterns for the common operational sea ice products sea ice thickness and concentration.
Comparison of the spatial patterns with images of existing sea ice products revealed that class shapes are largely consistent with those of developmental stages and thickness ranges. The information contained in the low-frequency channels allows the algorithm to reveal ranges of thin sea ice, and ranges of thicker ice can be determined from the relationship between the high-frequency channels and the changing surface conditions as the sea ice ages and thickens during freezing. Evaluation of class formation over several years indicated the temporal stability of obtained classes, recognizing annually recurring patterns that can be considered significant even over a one-year period. Based on the statistical model parameters, clusters are analyzed to understand the individual and combined sensitivity of input features to the obtained classes and associated sea ice properties. Model uncertainty is quantified by an entropy measure based on the class membership probabilities. Preliminary results have shown that – subsequent to classification - class probabilities can also be related to the distribution of sea ice thickness. In terms of the dependence of sea ice growth on its thickness, accounting for the distribution of sea ice thickness
would have an advantage over estimating only a single value for each grid cell. The approach presented is suitable for combining large data sets and provides appropriate metrics for class analysis and interpretation, allowing informed decisions to be made about integrating data from future missions into sea ice products.
The Spire GNSS radio occultation constellation of over 40 operational satellites collects GNSS signals and is relied upon to support critical numerical weather prediction. In 2019, these satellites were reprogrammed to also receive signals reflected coherently from the Earth’s surface at grazing angles (less than 30 degrees of elevation), adding GNSS reflectometry (GNSS-R) measurements to Spire’s operational capability. We present these novel measurements, available for collaboration and research, with demonstration of our application to sea ice detection and ice type classification algorithms in addition to detection and characterisation of the Antarctic Marginal Ice Zone.
These measurements consist of both the phase and amplitude data of the reflected and direct components of the signal at dual frequency (L1 and L2/E5 band). This allows the calculation of a phase-based coherence metric in addition to the surface reflectivity, based upon the relative received power of the reflected and direct signals. These variables are then applied to the identification of surface types in the Northern and Southern sea ice zones to estimate sea ice extent and sea ice type through a simple thresholding algorithm. Two years of sea ice detection and classification products are presented, with detection in agreement with external sources to 98% in the Antarctic and 94% in the Arctic, and binary sea ice classification (First-Year vs. Multi-Year) to 73% using these basic thresholds. Use of decision tree machine learning techniques through gradient-boosting library XGBoost to these variables increases this agreement to 84.7% for comparisons with OSI SAF operational sea ice type product, and to 70.3% for comparisons with aggregated stage of development data from the US National Ice Center weekly ice charts. Multi-class classification and variable exploration using XGBoost shows Young Ice to give the lowest coherence of reflections, First-Year Ice the highest, and Multi-Year ice medium reflection coherence.
An additional variable, removing the amplitude component of the phase-dependent coherence metric shows strong geographic patterns linked to expected ice type distributions and a clear sensitivity to the Marginal Ice Zone in the Antarctic. Areas of current product development and research are presented (floe size and type, as well as in-ice wave penetration) as well as opportunities for data availability and collaboration for research into these unique datasets.
Further to these grazing angle GNSS-R products, two new conventional geometry, near-nadir Spire GNSS-R satellites are collecting polar data, and first sea ice results from these satellites will be presented. Work is ongoing to leverage these two geometries of GNSS-R to create a sensor fusion product, harnessing the differences in viewing geometry over variable surfaces. Together, grazing angle GNSS-R data from Spire’s GNSS-RO satellites and data from conventional GNSS-R satellites is affording new insights into the cryosphere.
Robust and reliable mapping of sea ice types is required for both operational and environmental applications in the Arctic. Operational sea ice charts are distributed on a daily basis by national ice services, using synthetic aperture radar (SAR) images as the main data source. Besides an increasing amount of available images and considerable research effort in the field of (semi-)automated sea ice mapping, the operational charting process is presently still carried out manually. Furthermore, while it has been demonstrated in the literature that SAR imagery acquired at C- and L-band contain complementary information, there has not been much focus on adding L-band data to the operational workflow of the ice services, which is based on C-band imagery. This is mostly due to the fact that – in contrast to C-band data – L-band data is not yet routinely available for operational purposes. This is expected to change with ESA’s planned mission for an operational L-band SAR satellite (ROSE-L) [1].
In this study, we therefore investigate the benefits of adding L-band data to the (semi-)automated operational ice charting process. We use images acquired by ALOS-2 for L-band and Sentinel-1 for C-band, respectively. From a total set of more than 1000 ALOS-2 images acquired at different test sites in both Wide Beam (WB) and Fine Beam (FB) mode, we identify those images with sufficient spatial overlap and a short enough difference in acquisition time to Sentinel-1 extra wide (EW) and interferometric wide (IW) swath mode images. Test sites with the most frequent image pairs that fulfill these criteria are located in the Lincoln Sea (North Greenland), Fram Strait, and Belgica Bank. The selected image pairs are corrected for ice drift using an algorithm developed at Chalmers University of Technology, and the resulting aligned image pairs can be stacked and used for both single- and multi-frequency classification.
We analyze the image pairs by applying incident-angle aware supervised classification [2,3] and unsupervised segmentation [4], as well as semi-supervised, graph-based label propagation [5]. During the analysis, it must be considered that, besides frequency, other data and image properties such as spatial resolution, local incident angle, noise level, and the number of looks will influence visual interpretation and thus the initial definition of ice classes, as well as class separability and classification accuracy.
First results clearly demonstrate distinct advantages of combining aligned C- and L-band data for ice type classification, in particular for identifying areas of deformed and thin ice.
[1] W. Dierking. “Synergistic Use of L- and C-Band SAR Satellites for Sea Ice Monitoring”. Proceedings 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS. 2021. DOI: 10.1109/IGARSS47720.2021.9554359
[2] J. Lohse, A.P. Doulgeris, W. Dierking. “Mapping sea-ice types from Sentinel-1 considering the surface-type dependent effect of incidence angle”. Annals of Glaciology, 2020. DOI: 10.1017/aog.2020.45
[3] J. Lohse, A.P. Doulgeris, W. Dierking. “Incident Angle Dependence of Sentinel-1 Texture Features for Sea Ice Classification”. Remote Sensing, 2021. DOI: 10.3390/rs13040552
[4] A. Cristea, J. Van Houtte, A.P. Doulgeris. “Integrating incidence angle dependencies into the clustering-based segmentation of SAR images”. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 13, 2020. DOI: 10.1109/JSTARS.2020.2993067
[5] C. Taelman, S. Chlaily, E. Khachatrian, F. van der Sommen, A. Marinoni. “On the Exploitation of Heterophily in Graph-based Multimodal Remote Sensing Data Analysis“. 30th ACM international conference on information and knowledge management (CIKM): Workshop on Complex Data Challenges in Earth Observation, 2021. Accepted for publication
The Copernicus Imaging Microwave Radiometer (CIMR) [1] is a wide-swath conically-scanning multi-
frequency microwave radiometer from 1.4 to 36 GHz. It will to provide a wide range of sea-ice information,
including sea ice concentration, thin sea ice thickness and snow depth over sea ice. The Copernicus Polar
Ice and Snow Topography Altimeter (CRISTAL) [2] will carry a multi-frequency radar altimeter and
microwave radiometer to monitor sea ice thickness and overlying snow depth. Both missions are Coper-
nicus high priority to respond to the Integrated European Union Policy for the Arctic. At the same time,
MetOp-SG will carry the ASCAT instrument, that shows sensitivity to sea ice properties, especially the
ice type. Here, we propose to analyze the potential synergies of these instruments, using existing obser-
vations with similar characteristics (although less optimal).
The combination of AMSR2 and SMAP can mimic CIMR, SARAL and Sentinel-3 are proxies for
CRISTAL, and ASCAT is already available on MetOp-A and -B. A data set of coincident AMSR2,
SMAP, SARAL, Sentinel-3 and ASCAT observations is constructed, over the Poles, over a year. It in-
cludes both the Level 1 and Level 2 products. We concentrate first on the study of the complementarity
between the observations, at Level 1. It has been previously shown that the exploitation of the obser-
vation synergy at Level 1 is more efficient than a posteriori combinations of products, independently
estimated from different instruments [3]. Then, in order to analyze results of this database, the Snow
Microwave Radiometric Transfer (SMRT) [4] model is used. It is an up-to-date radiative transfer model
that is tailored to handle snow and sea ice in a plane-parallel configuration, and it can simulate both
passive and active microwave responses.
A first study [5] has shown that the use of CIMR-like data with the SMRT model can explain tem-
poral and spatial distribution of microwave signatures over the whole North Pole during all year long.
From this interpretation, a realistic characterization of the sea ice and its snow cover has been provided.
Correlation and causalities, between microwave signatures and geophysical properties (such as sea ice
type, sea ice thickness, snow depth or snow microstructure), have been classified.
Here, we extend this study to the Austral Ocean and to altimetric data, southern sea ice being more
covered by current altimeters than northern sea ice. Recent developments in SMRT have made it able to
simulate altimetric signal [6], and are used to interpret CRISTAL-like data. Synergies between CIMR-like
and CRISTAL-like data are highlighted for an improved sea ice and snow characterization.
Sea ice floe edges become indistinguishable inside the winter pack ice of the central Arctic. They freeze together into clusters or groups of floes that move relative to each other only during sea ice deformation events driven by weather patterns. Consequently, the boundaries and shapes of these clusters change on a synoptic scale. These boundaries are typically several kilometers long and nearly straight – commonly named linear kinematic features (LKFs). Geometric characteristics of these ‘active floes’ between the LKFs and their spatio-temporal development can be therefore tracked by studying sea ice deformation concentrated along LKFs. In this study we use sea ice drift retrievals from displacements between SAR image pairs. We derive a deformation-rate thresholding method that enables relatively high spatial resolution of ~200m and bellow. We use the sequences of deformation fields to 1) track sea ice cover than underwent sea ice deformation (damaged ice), and 2) derive LKF area, intersection angle, and active floe area and shape. The method is applied to the freeze-up and early winter of a collection SAR imagery of Sentinel-1 and Radarsat-2 available over the broader area of the MOSAiC (Multidisciplinary drifting Observatory for the Study of Arctic Climate) campaign between October and December 2019. Our preliminary results show continuously active winter sea ice cover and its fracturization at strongest weather events. We also observe regularly distributed LKFs (damaged sea ice) that is seemingly independent of sea ice thickness and age. Some damaged areas are reactivated in temporally distant deformation events.
Once validated against the airborne data collected at MOSAiC our data set can be used for SAR sea ice classification validation and for statistical evaluation of numerical models of sea ice.
Figure caption: Left: example of sea ice damage tracked based on sea ice deformation derived from ESA Sentinel-1 image pairs over one week. Right: example of ‘active floes’ derived from a single ESA Sentinel-1 image pair. Linear kinematic features (LKF) are dark green, while each floe has a different color. The development of area, shape parameters of the floes as well as LKF area and intersection angle is extracted from such maps and tracked in time.