Coastal Sea Ice Concentration Derived from Marine Radar Images: A Case Study from Utqiaġvik, Alaska

1. Introduction

In recent decades, the Arctic has witnessed significant changes in its climate with key projected impacts on permafrost, biodiversity, sea level and Inuit health [1,2,3,4,5,6,7,8]. Of all the observed changes, the most striking is the decline in the minimum sea ice extent (SIE), at a rate of -12.7 %/decade since 1979 [7,9], and the begining of the transition from a predominantly perennial to a seasonal sea ice cover since the record low SIE of 2007 [10,11]. These changes in the environment affect Arctic indigenous communities directly, in their way of living, hunting capacity, health, transportation, and infrastructures [12,13,14,15].

Output from different sensors – e.g., the Defense Meteorological Satellite Program (DMSP) Special Sensor Microwave Imager (SSM/I, 12.5-25 km) and the Special Sensor Microwave Imager/Sounder (SSMIS, 12.5-25km), the Aqua Project Advanced Microwave Scanning Radiometer 2 (AMSR2, 6.25-12.5-25km), and the Advanced Very High Resolution Radiometer (AVHRR, 2 km) – have been used since the start of the modern satellite era in the late seventies, to monitor changes in the sea ice cover – e.g., the National Snow and Ice Data Center (NSIDC) Climate Data Record (CDR) [16,17]. These sensors have spatial resolutions ranging from 1 km to 50 km [18,19] and allow for a quantitative assessment at pan-Arctic scale, including, for instance, changes in the location of the sea ice edge or detecting the presence of large polynyas [20]. Passive microwave derived sea ice concentration however suffers from land-contamination [21] and the resulting uncertainty reduces their usefulness to coastal communities. Since the mid-nineties, sea ice conditions can also be assessed from synthetic aperture radars (SARs) with spatial resolutions ranging from about 1 meters up to a kilometer (e.g., ice charts from the Canadian Ice Service, [22] ). The SAR images, however, are more difficult to interpret and not consistently available on a daily basis [23,24,25]. Recently, high-resolution images in the visible range from MODIS were combined with sea ice information from AVHRR to produce a pan-Arctic, 1 km resolution, daily product (MODIS-AMSR2, [18]) potentially bridging the spatial scale gap between global satellite product and human scale. This dataset however is not available in real-time making not usable for operational purposes.

To address this issue, marine and coastal radars have rerently been used to monitor the stability of the landfast ice cover, sea ice drift and the presence of coastal flaw-lead polynya [26,27]. In the early 1970s, three C-band coastal radars were installed in the Sea of Okhotsk (Hokkaido, Japan), monitoring the sea ice drift over a total area of 17,500 square km² using cross-correlation technique [28]. More recently, new techniques emerged to derive sea ice drift estimates from radar images in both coastal and marine environment. These include dense and sparse optical flow techniques, feature tracking and phase-correlation algorithms, and normalized cross-correlation methods Mv et al. [29], Karvonen [30,31], Oikkonen et al. [32], Lund et al. [33]. While coastal radars have historically been used to distinguish between first-year ice, multi-year ice, and open-water [34,35,36], recent applications have used high-frequency radar to identify the sea ice edge and near Shirasawa et al. [26], and optical flow techniques to detect the landfast ice edge [29]. One of the earliest and still operating marine coastal radar is the Utqiaġvik (Alaska) Coastal Sea Ice Radar System (CSIRS) installed in 1977 to study ice deformation events (ridge building) along the coast allowing for the establishment of a stable landfast ice cover and sea ice motion of the pack ice on the northward side of the flawlead polynya [36,37]. A detailed analysis of the same radar images has also demonstrated that flickering from radar reflectors can potentially be as an early warning signal for landfast ice break-up events [37]. The latest version of the CSIRS was installed in 2007 and validated against sea ice drift estimates from a moored ice profiling sonar deployed offshore of Utqiaġvik. During its operation, the radar has detected many ice motion events near the coast, including breakout, convergence, and high-speed anomalies [38,39]. The same radar images have also supported search and rescue and marine transportation activities [39]. As of now, however, no sea ice concentration estimates have been derived directly from the CSIRS images, serving as the motivation the present article.

In this study, we derive sea ice concentration (SIC) at small scale (∼10km) and high resolution (21.5 ± 0.5 m) from the Utqiaġvik (Alaska) Coastal Sea Ice Radar System (CSIRS) [39] using image processing techniques, and compare with two existing satellite-derived products (CDR, 25 km resolution and MODIS-AMSR2, 1km). The new data product can be used to assess uncertainties and limitations in existing satellite product, including land contamination and SIC estimates in the shoulder season at freeze-up and melt onset.

The paper is organized as follows. Section 2 presents the data sets used in this study. Section 3 describes the canny sea ice edge detection algorithm and the method used to find optimal parameters. Section 4 presents the optimization steps results, and the CSIRS SIC products and its comparison to the satellite data sets. Section 5 discusses the findings and the limitations.

2. Data

2.1. (Near-Real-Time) Climate Data Record Sea Ice Concentration

We use the (Near-Real-Time) daily National Oceanic and Atmospheric Administration (NOAA) and National Snow and Ice Data Center (NSIDC) Climate Data Record of sea ice concentration version 4 (2022) and version 2 (2023), [referred to as CDR SIC in the following, 16,17]. Both datasets are a blend from the NASA Team (NT) and the NASA Bootstrap (NB) algorithms applied to Brightness Temperatures from the (NRT) Defense Meteorological Satellite Program (DMSP) Special Sensor Microwave Imager (SSM/I) and the Special Sensor Microwave Imager/Sounder (SSMIS) [17,40,41]. For each day and grid cell, the higher SIC from both algorithms is used [17]. The SICs are stored on the NSIDC polar stereographic grid with a spatial resolution of 25 km [17]. The error is 5% during winter and larger in the summer ( 20%, from a comparison with higher resolution products) due to difficulties in differentiating surface melt and melt ponds from open water [17,19]. In the following, we refer to the (Near-Real-Time)-CDR as the

2.2. Merged MODIS-AMSR2 Sea Ice Concentration

We use the SIC derived from Daily merged Moderated Resolution Imaging Spectroradiometer (MODIS) - Advanced Microwave Scanning Radiometer 2 (AMSR2) Thermal Infrared and Passive Microwave data available for the winter (September-May) 2022-2023 [18]. The merged dataset takes advantage of the continuous spatial coverage, but low spatial resolution (5km) of AMSR2, and high spatial resolution, but limited to cloud-free scene of MODIS [18], and resolve surface features such as leads between 60% and 90% and polynyas that are not seen by coarser Passive Microwave SIC data set [20]. To this end, the MODIS SICs are adjusted to maintain the AMSR2-derived mean SIC at 5km resolution[18]. The SIC are stored on the NSIDC polar stereographic grid with cells of a spatial resolution of 1 km [18]. The error on the merged product are between 5% and 10% from February to April and up to 18% in May [18], when the MODIS errors are larger ( 30%).

2.3. Coastal Sea Ice Radar System Images

We use 4-minute snapshots from the University of Alaska Fairbanks coastal sea ice radar in Utqiaġvik available in near-real time at https://seaice.alaska.edu/gi/observatories/barrow_radar/ [36,37,38,39,42]. The radar is located 22.5 m above sea level on the rooftop of local bank downtown Utqiaġvik [39]. The operating radar is a 25 kW Furuno FAR2117X band system with a 2.44 open array antenna [39]. The radar can detect sea ice features up to 11 km depending on atmospheric conditions in its current configuration [39]. Specifically, fog and precipitation attenuate the radar signals, which makes the identification of sea ice features more difficult [29]. The images are reprojected to a Cartesian plane where each pixel is $21.5\pm 5$ m [38]. We note that coastal sea ice radars are mostly sensitive to ridges and floe edges – due to the small grazing angle–, with little energy reflected from smooth surface inside the pack of ice [38].

3. Methods

3.1. Floe Edge Detection Algorithms

We take advantage of the radar sensitivity to floe edges to identify ice floes and open water using the canny edge detection operator [43,44] and the contour detection algorithm [45] from the open-source computer vision library OpenCV.

The canny edge and Suzuki algorithms extract floe edges from the masked gray-scale images in six separate steps [43,45]. First, the land, Elson lagoon and the boundary at the edge of the field of view are masked from all frames to limit the analysis to the ocean field of view that is resolved by the radar (Figure 1 a-b). Second, a 5x5 Gaussian filter is applied to remove noise, followed by a two-dimensional 3x3 Sobel kernel operator to identify edge gradients. Third, the edge gradients are refined using a non-maximum suppression step keeping only the local maximum in gradients. Fourth, the floe edges are formed by applying a two-threshold hysteresis procedure where pixels with a gradient larger than an upper threshold ($T_{up}$) are considered as an edge, and pixels with a gradient in between the lower and upper thresholds ($T_{low}<$ gradient $<T_{up}$) are considered as an edge only if they are connected to another edge pixel [43,46]. A threshold below the lower threshold is always rejected. Fifth, a binary mask is used to remove the radar image boundary where large gradients are present. Sixth, the Suzuki algorithm is used to link the points identified as edges by the canny edge algorithm [45]. Specifically, the edges are dilated by convoluting the image with a square kernel and with the pixel below the anchor point equal to the maximum value within the kernel square, resulting in well-defined (sharp gradient) edge segments that are all connected together [46, Figure 1 d]. The optimization procedure to determine the upper and lower thresholds and kernels is discussed below in subsection 3.3.

3.2. Post-Processing

We first delete (rare events of) closed contours within larger contours, assuming that they are primarily shadows behind ridges, rather than open water. This is confirmed visually from inspection of the animated radar images for 10 scenes also analyzed by human analysts (see Section 3.3 for details) where detection of ice and open water is easier. In turn, the areas of low-radar return – in the shadow of ridges or covered by smooth ice inside edges – are marked as sea ice despite having a similar signature as open water [38, Figure 1 e]. Finally, we calculate the (sea ice) area within each closed contours. The sea ice concentration is the ratio of the sea ice area to the total radar ocean field of view.

3.3. Parameters Optimization

Following [47], we use the sea ice edge analysis of fourteen frames by 10 different analysts from our research group as truth for the optimization of the two thresholds and kernel. The fourteen frames, selected in the 2022-2023 winter season, are representative of the full range of sea ice conditions seen by the radar, i.e., 100% sea ice, 100% open water, and a mixture of open-water, landfast, and drifting ice.

The floe detection algorithm was run over the full parameter space for the lower and upper thresholds 8-bits (0 – 255) and kernel dilation window size, (3, 5, 7, 9, 11, 13) to find the set of parameters that minimizes the root mean square error (RMSE) and slope of the best line fit of sea ice concentration for 14 different frames produced by the algorithm from 10 analysts. We choose the RMSE for the loss function instead of the mean bias error (MBE) because we want an algorithm that performs well for a wide range of SIC.

3.4. CSIRS Daily SIC Product

The CSIRS daily mean SIC is calculated from the output of the floe detection algorithm run with optimal parameters on each 4 min frame (360 in total) for each day of 2022-2023. Missing frames are omitted from the calculation.

3.5. Comparison of SIC Products

We perform a qualitative comparison of the NASA WorldView MODIS Terra (https://worldview.earthdata.nasa.gov) and the marine radar images, and a point-by point comparison between the merged MODIS-AMSR2 and an interpolated coarse-grained CSIRS SIC of similar spatial resolution (1 km), for April 14, 2022, a day when no clouds were present. We give an estimate of the spread in the merged MODIS-ASMR2 from the mean, minimum and maximum SIC for all grid cells that are within the field of view of the radar for comparison with the reconstructed sea ice concentration from the marine radar (Figure 2). For the CDR dataset, only one grid cell is in the field of view of the marine radar and no spread estimate is possible. We use the standard deviation as a proxy for the error (associated with poor radar return and land contamination in both datasets. For the correlation between CDR, MODIS-AMSR2 and CSIRS dataset, the anomaly time series are calculated by removing a 30-day running mean.

4. Results

4.1. Optimization/Validation

The best line fit between the canny edge algorithm and the analysis of 14 images from 10 analysts is obtained with the 11x11 kernel, and thresholds in pixel intensity ($T_{low}$ and $T_{high}$) equal to 4 and 82, respectively (Figure 3 (a)). The resulting slope and intercept of the linear regression are 0.94 and 0.036, respectively, while the RMSE is 0.035. An optimization based on the RMSE instead results in a small reduction in error (RMSE = 0.035) but a large maximum bias (= 0.1 for SIC=1) and is therefore not the preferred option (Figure 3 (a)). (Figure 3 (b)). The inter-analyst error (= 0.026) – a measure of the error on the "truth" – is small compared with the RMSE indicating that most of the difference between the kernels can be attributed to the algorithm (Figure 3 b). Finally, the algorithm can incorrectly identify open water as sea ice when there is fog or noise in the image, which can explain the radar SIC’s overestimation compared to the analyst’s.

The detection algorithm with the optimal parameter successfully identifies most of the sea ice edge features observed by MODIS Terra on April 14, 2022, under cloud-free conditions (Figure 4 (a-b)). The open water and the narrow band of open water south-west and directly in front of the building on which the CSIRS is mounted, where the CSIRS is located, are well seen. However, the open water to the north and the water between the broken floes to the north of the radar image however are missed by the algorithm. (Figure 4 (a-b)). Additionally, the rugged features of the sea ice edge are consistently present in both images. The coarse-grain CSIRS SIC estimate also correlates well with the daily merged MODIS-AMSR2 SIC, with the correlation coefficient exceeding 0.6 for most of the day. The drop in correlation during the middle of the day is presumably due to the swaths composing the merged MODIS-AMSR2 SIC were not taken at this time. Overall, the detection algorithm can identify most sea ice edges present in the MODIS image, considering that the exact acquisition time of the merged MODIS-AMSR2 SIC passing over the marine radar field of view is not known.

4.2. CSIRS Reconstructed Sea Ice Concentration

4.2.1. Daily SIC

The CDR data and the merged MODIS-AMSR2 overestimate SIC with respect to the CSIRS with an MBE equal to -0.10 and $-9.0·{10}^{-3}$, respectively (Figure 5 a, Table 1). The correlation coefficients between the daily CDR, MODIS-AMSR2 and the reconstructed SIC from CSIRS are low (r=0.18) and moderate (r=0.54), both significant at the 95% level, and RMSE errors of 0.34 and 0.29, respectively (Table 1). At weekly and monthly time scale, the correlation coefficients for the CDR and MODIS-AMSR2 increase to 0.276-0.71 and 0.68-0.59, respectively (results not shown). In winter, the CDR generally underestimates short-lived low SIC (breakout) events – most notably in early winter 2023 (Jan-Feb 2023) – while MODIS-AMSR2 SICs are in general agreement with observed values (Figure 5 a). In the break-up period (Jun-Jul 2022), CDR is in general agreement with observations, while during the freeze-up (Dec 2022 - Jan 2023) both satellite leads the observed ice onset by approximately 10 days, perhaps due to land contamination with an earlier onset of snow presence on land [48]. In the summer, the algorithm incorrectly identifies clouds as sea ice with reconstructed SICs ranging between 20-25%. The CDR SIC on the other hand correctly identified open water during this period (Fig 4-5).

5. Discussion and Conclusions

To first order, both the CDR and the merged MODIS-AMSR2 datasets captures the same seasonality of SIC at Utqiagvik as the CSIRS. In mid- to late-winter, spring and early summer, the CDR SIC is in good agreement with the CSIRS SIC, while the merged MODIS-AMSR2 is in better agreement in the fall freeze-up and early winter. In early winter of 2023, the CDR underestimates break-up events while the MODIS-AMSR2 under-estimates the CSIRS SIC in mid- to late-winter, presumably due to the AMSR2 mask not capturing the landfast ice area near the coast (see Figure 2). In the summer, the canny edge algorithm interpret the presence of fog and clouds as sea ice near the coast leading to a positive bias in the reconstruction.

In summary, the floe edge detection algorithm is successful at identifying sea ice and open water, including break-up events under all conditions, but misinterpret fog and cloud as sea ice particularly in the summer [37,49,50]. Marine radar images along with floe detection algorithm can be use jointly to assess uncertainties associated with land contamination in low and high resolution satellite products. Raising the coastal radar height above ground (if at all possible) and therefore radar range would provide a useful data set to evaluate changes in land contamination with distance from the coast and would reduce shadow effect from nearby ridges. Two new marine radars were installed recently in Nome and Gambell (AK). Future work include testing the proposed algorithm on the new radar images from locations with very different sea ice regimes as well as implementing an edge tracking feature in time to even better estimate the sea ice concentration.