Semi-automated tracking of iceberg B43 using Sentinel-1 SAR images via Google Earth Engine

. Sentinel-1 C-band synthetic aperture radar (SAR) images can be used to observe the drift of icebergs over the Southern Ocean with around 1-3 days of temporal resolution and 10-40 m of spatial resolution. The Google Earth Engine (GEE) cloud-based platform allows processing of a large quantity of Sentinel-1 images, saving time and computational resources. In this study, we process Sentinel-1 data via GEE to detect and track the drift of iceberg B43 during its lifespan of 3 years (2017-2020) in the 15 Southern Ocean. First, to detect all candidate icebergs in Sentinel-1 images, we employ an object-based image segmentation (simple non-iterative clustering – SNIC) and a traditional backscatter threshold method. Next, we automatically choose and trace the location of the target iceberg by comparing the centroid distance histograms (CDHs) of all detected icebergs in subsequent days with the CDH of the reference target iceberg. Using this approach, we successfully track the iceberg B43 from the Amundsen Sea to the Ross Sea, and examine its changes in area, speed, and direction. Three periods with sudden losses of area (i.e. split-offs) 20 coincide with periods of low sea ice concentration, warm air temperature, and high waves. This implies that these variables may be related to mechanisms causing the split-off of the iceberg. Since the iceberg is generally surrounded by compacted sea ice, its drift correlates in part with sea ice motion and wind velocity. Given that the bulk of the iceberg is under water (~30-60 m freeboard and ~150-400


Introduction
When a large ice mass breaks off from an ice shelf or glacier into the ocean, it forms an iceberg. An iceberg has a lifespan of several years or longer and its area ranges from a few square kilometers to thousands of square kilometers. Since Considering the majority of an iceberg is under water, iceberg drift is a good indicator of ocean circulation (Collares et al., 2018). Since the trajectories and speeds of icebergs also depend on multiple and complex environmental variables (e.g. ocean, atmosphere, sea ice, 35 etc.), icebergs provide important insights for the interaction of these variables (Schodlok et al., 2006). In addition, the formation and melting of icebergs influences global climate (Romanov et al., 2008;Mackie et al., 2020), ocean flux (Silva et al., 2006;Rackow et al., 2017;Starr et al., 2021), sea ice production (Martin et al., 2007;Merino et al., 2016), dissolved iron concentration (Lin et al., 2011;De Jong et al., 2015), and ecosystems and biology (Wilson et al., 2016;Schwarz and Schodlok, 2009;Biddle et al., 2015). Furthermore, icebergs can threaten ship navigation (Lasserre, 2015). Therefore, detecting and tracking icebergs is 40 extremely important to understand the changing sea ice, ocean, and atmosphere in the polar regions.
Radar remote sensing, including scatterometer and synthetic aperture radar (SAR), is an efficient tool for monitoring both movements and area changes of icebergs. While multispectral images can be useful for observing icebergs, they cannot be used during polar night or under cloudy conditions. In contrast, synthetic aperture radar (SAR) SAR images can be used for analysis regardless of the weather conditions or time of year (Han et al., 2019;Mazur et al., 2017;Wesche and Dierking, 2012). In particular, 45 although scatterometers facilitate daily position and motion observations of large icebergs with a coarse spatial resolution (Budge and Long, 2018;Stuart and Long, 2011b, a), SAR instruments have a more significant advantage in precise observations of iceberg area changes due to their relatively fine spatial resolution. Indeed, various SAR instruments have been used for detecting or tracking icebergs in the polar oceans including ERS-1 (Young et al., 1998;Willis et al., 1996), ENVISAT (Li et al., 2018;Howell et al., 2004;Mazur et al., 2017;Barbat et al., 2019Barbat et al., , 2021, Radarsat-1 (Wesche and Dierking, 2015;Lane et al., 2002;Power et al., 2001), 50 Radarsat-2 (Scheuchl et al., 2004;Denbina and Collins, 2014), TerraSAR-X (Frost et al., 2016), and Sentinel-1 (Lopez-Lopez et al., 2021;Moctezuma-Flores and Parmiggiani, 2017;Heiselberg, 2020;Han et al., 2019).
Most of these SAR studies used the higher backscatter contrast of icebergs to distinguish icebergs from the surrounding sea ice or water (Mazur et al., 2017). In addition, recently, instead of traditional pixel-based approach, image segmentation methods have become popular to reduce the speckle degradation of SAR images and efficiently detect icebergs (Lopez-Lopez et al., 2021;55 Mazur et al., 2017;Barbat et al., 2019). Although the iceberg detection is commonly conducted automatically, the tracking is still usually conducted by manually finding the images containing the target iceberg (Moctezuma-Flores and Parmiggiani, 2017;Parmiggiani et al., 2018;Li et al., 2018;Mazur et al., 2019), which requires a lot of time and labor so restricts the number of processed images. Moreover, even if both detection and tracking processes are automatized, overall process requires downloading a large number of satellite images (Barbat et al., 2021), which takes long time and needs a huge storage capacity. 60 In this respect, Google Earth Engine (GEE), a cloud computing platform for geospatial analysis by Google (Gorelick et al., 2017), can be a promising tool to automatize the detection and tracking of icebergs. GEE makes it possible to process a large volume of geospatial data without downloading them onto local computers. Although various satellite images are freely accessible via GEE in near real-time, Sentinel-1 is the only SAR data that has been provided by GEE as individual images in near real-time.
As the constellation of the two twin satellites (Senitnel-1A/B), Sentinel-1 has a remarkable temporal (< 2-3 days in the polar 65 regions) and spatial resolution (≤ 40 m) (Torres et al., 2012). A number of studies have taken advantage of the unprecedented computing performance of GEE for the Sentinel-1 data processing: glacier margin mapping (Lea, 2018), glacier lake mapping (Zhang et al., 2020), estimation of glacier surface speed (Di Tullio et al., 2018), crop mapping (Mandal et al., 2018;Singha et al., 2019;Jin et al., 2019), flood mapping (DeVries et al., 2020;Clement et al., 2018), and wetland mapping (Mahdianpari et al., 2019(Mahdianpari et al., , 2020. Despite the variety of applications of Sentinel-1 data in the GEE platform, there have not yet been any published studies that leverage GEE's large data storage capabilities and computing power for automated tracking of icebergs. Sentinel-1 has been popular for iceberg detection since the launch of Sentinel-1 A and B in 2014 and 2016, respectively, because of its remarkable temporal and spatial resolution. Since each satellite has a 12-day revisit cycle, the constellation of the two satellites (Sentinel-1 A/B) can have a revisit cycle of 6 days. Moreover, because the coverage is more frequent at high latitudes in the Southern Ocean, one can acquire Sentinel-1 data with a temporal resolution of at least 2-3 days (Torres et al., 2012). Sentinel-75 1 also has a high spatial resolution (≤ 40 m depending on resolution modes), enabling resolution of icebergs and iceberg features at a fine spatial scale.
Another advantage of Sentinel-1 data is that it is non-commercial and freely accessible via Google Earth Engine (GEE), a cloud computing platform for geospatial analysis by Google (Gorelick et al., 2017b). GEE makes it possible to process a large volume of geospatial data without downloading them onto local computers. A number of studies have taken advantage of the 80 unprecedented computing performance of GEE for Sentinel-1 data processing: glacier margin mapping (Lea, 2018), glacier lake mapping (Zhang et al., 2020), estimation of glacier surface speed (Di Tullio et al., 2018), crop mapping (Mandal et al., 2018;Singha et al., 2019;Jin et al., 2019), flood mapping (DeVries et al., 2020;Clement et al., 2018), and wetland mapping (Mahdianpari et al., 2019(Mahdianpari et al., , 2020. Despite the variety of applications of Sentinel-1 data in the GEE platform, there have not yet been any published studies that leverage GEE's large data storage capabilities and computing power for automatic tracking of icebergs. 85 In this study, therefore, we track the drift of an iceberg in the Southern Ocean by taking advantage of the Sentinel-1 data and the GEE platform. Our target iceberg is the iceberg B43, as designated by the National Ice Center (NIC, usicecenter.gov), which calved off from the Thwaites Glacier ice shelf in the Amundsen Sea in April 2017 ( Figure 1). This tabular iceberg had a relatively large size (> 100 km 2 in April 2017) and a distinctive shape compared to surrounding icebergs. As a preliminary investigation based on manual downloading of Sentinel-1 images and visual interpretation before applying the tracking method, 90 we found that this icebergIt drifted westward along the coast from the Amundsen Sea to the Ross Sea with a relatively stable flattopped surface and eventually broke into several pieces in the northern Ross Sea in March 2020. Since the iceberg has a distinctive shape that facilitates its identification, we were able to trace its entire journey. The aim of this study is (1) to take advantage of the GEE cloud computing platform for its potential to track the drift and decay of this iceberg from the Amundsen Sea to the Ross Sea and (2) to assess the impact of major environmental factors (e.g. ocean currents, winds, sea ice concentration, sea ice drift, 95 temperature) on changes in its movement and area. In addition, we examine satellite altimeter data to estimate the freeboard and thickness of this iceberg.

Sentinel-1 SAR images
In this study, we use ESA's Sentinel-1 C-band (center frequency of 5.405 GHz, wavelength of 5.6 cm) SAR Ground Range Detected (GRD) data in order to detect the location and area of the target iceberg. This product has three different spatial resolutions of 10, 25, or 40 meters with different resolution modes, but all images are resampled to 40 m. Since each Sentinel-1 satellite (Sentinel-1A/B) has a 12-day revisit cycle, this makes the combination of Sentinel-1 A and B has a revisit cycle of 6 days; 110 benefiting from the high latitudes of the Southern Ocean, we can acquire at least one Sentinel-1 image capturing the iceberg per 2-3 days (Torres et al., 2012). We import and process these images via the GEE Code Editor (web-based integrated development environment (IDE) for the GEE JavaScript API) and GEE Python API. Each Sentinel-1 image scene is released from ESA after being pre-processed with (1) thermal noise removal, (2) radiometric calibration, and (3) terrain correction (sentinel.esa.int). Of four available band combinations of the polarization products (VV, HH, VV+VH, and HH+HV), we use only the HH band because 115 most of the images contain this band. In addition, we also use the angle band that represents the interpolated viewing incidence angle at each cell. During the automated tracking of the iceberg, a total of 433 images from April 2017 to March 2020 are used.

Satellite altimeter data
In order to estimate the freeboard and thickness of the iceberg, we use height measurements from two satellite altimeters: CryoSat-2 and ICESat-2. 120 CryoSat-2 employs Ku band radar (center frequency of 13.575 GHz) to measure surface heights with 400 m of along track footprint and 1.65 km of across track footprint. It operates in SAR mode over sea ice and oceanographic areas and SAR Interferometric (SIN) mode around ice sheet edges and mountain glaciers (ESA, 2019). Since the target iceberg floated in the ocean through sea ice and coastal areas, we search for both Level 2 SAR and SIN products that overlapped with the iceberg from April 2017 to March 2020. 125 ICESat-2's Advanced Topographic Laser Altimeter System (ATLAS) uses laser photons at 532 nm to retrieve surface heights with 171 m of laser footprint and 0.7 m of spacing (Magruder et al., 2020;Neumann et al., 2019). Trillions of photons are emitted for each pulse but the number of returned photons is around seven photons for a typical snow-covered surface (Neumann et al., 2019). ATLAS consists of six multi-beams, each with three strong beams and three weak beams, and the strong beams have four times greater energy than the weak beams (Neumann et al., 2019;Markus et al., 2017). We search for ICESat-2 ATL03 130 geolocated photon products (release 003) (Neumann et al., 2020) that intersected the iceberg during our study period.

Meteorological and sea ice motion data
We compare the drift speed, direction, and area changes of the iceberg with meteorological data from the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5 reanalysis model. We use the 2m air temperature, sea ice concentration (SIC), wind speed/direction, and wave height from the ERA5 hourly product, which has a spatial resolution of about 0.25°. This 135 product was is acquired from the Climate Data Store (cds.climate.copernicus.eu) of the Copernicus Climate Change Service.
In addition, we acquire National Snow and Ice Data Center (NSIDC) sea ice motion vector data (nsidc.org/data/NSIDC-0116) (Tschudi et al., 2019). This dataset contains daily sea ice motion vectors, which are derived from multiple data sources, including AVHRR, AMSR-E, SMMR, SSMI, and SSMI/S sensors, International Arctic Buoy Program (IABP) buoys (in the Arctic only), and National Center for Environmental Prediction (NCEP) and National Center for Atmospheric Research (NCAR) 140 Reanalysis forecasts. These sea ice motions are projected to the 25 km EASE (Equal-Area Scalable Earth)-Grid.

Iceberg detection
To detect the target iceberg from the Sentinel-1 images, we employ a superpixel image segmentation method named 145 simple non-iterative clustering (SNIC) (Achanta and Süsstrunk, 2017). This approach uses a similar concept to k-means based clustering (Simple Linear Iterative Clustering (SLIC) algorithm (Achanta et al., 2012)), but in contrast with SLIC, SNIC enforces connectivity from the start. Since GEE provides this SNIC algorithm as a basic built-in function, we use the SNIC function in the GEE environment. The object-based segmentation processes of SNIC in this study are summarized by the following steps:

(i) Apply Gaussian filter to the raw SAR image
The speckle effects of the original SAR images can introduce errors ( Figure 2a). In particular, the image segmentation process can misinterpret shaded areas associated with the iceberg surface topography as artificial iceberg boundaries and split the iceberg into several pieces. To reduce this effect, we first smooth images by applying a Gaussian filter kernel within a 3-pixel radius ( Figure 2b). 155 (ii) Initialize superpixels In the SNIC image segmentation, all image pixels should be grouped into small clusters of connected pixels, which are named superpixels. The centroids of superpixels, referred to as seeds, are initialized with a given number of pixels chosen a priori on a regular grid. The size of each segment is determined by the seed value. A smaller seed value can distinigush small icebergs 160 but takes a longer computation time ( Figure 2c); a larger seed value cannot distinguish small icebergs but can reduce computation time ( Figure 2d). As shown in Figure 2d, considering the size of the target iceberg, a seed value of 80 (i.e. initial centroid is spaced by 80 pixels) is adequate to detect the target iceberg B43.
(iii) Measure affinity to a centroid 165 The affinity of the j-th pixel to the k-th superpixel centroid is calculated by using the distance between them ( , ): where is a geocoordinate (latitude and longitude) of the j-th pixel, is a geocoordinate of the k-th superpixel centroid, is the HH band backscatter of the j-th pixel, is the HH band backscatter of the k-th superpixel centroid, and s and m are the normalizing factors for spatial and backscatter distance, respectively. If an image has N pixels and K superpixels are expected, s 170 in Equation 1 should be set to √ / . A higher compactness factor m leads to more compacted superpixels and poorer boundary adherence (Achanta and Süsstrunk, 2017), and we set m equal to 1.
(iv) Evolution of centroid Contrary to multiple iterations of SLIC (Achanta et al., 2012), SNIC uses a priority queue to choose the next pixel to be 175 added to a cluster. The priority queue consists of multiple candidate pixels that are 8-connected to a currently growing superpixel cluster. The pixel candidate that has the smallest distance from a centroid is selected among these pixel candidates, and the centroid value is updated online after a new pixel is added to that superpixel. This allows the SNIC algorithm to complete the updating of centroids in a single iteration with lower memory requirements. More details about the SNIC algorithm were explained in Achanta and Süsstrunk (2017). 180 After completing the SNIC image segmentation, we detect the segments of icebergs by applying the mean brightness filter to all segments. Icebergs commonly show higher backscatter intensity compared to the surrounding sea water or sea ice (Silva and Bigg, 2005;Young et al., 1998;Mazur et al., 2017;Wesche and Dierking, 2012). Based on the bright backscatter of icebergs , we identify iceberg segments as the segments satisfying the following equation (Mazur et al., 2017): According to Mazur et al. (2017), 185 we can identify iceberg segments as the segments satisfying the following equation: where γ is the backscatter of SAR images, and α is the incidence angle of the SAR. When we employ this equation, we can successfully distinguish the target iceberg from the surrounding area (Figure 2e and 2f).

Iceberg tracking
In order to automatically detect and track only the target iceberg among multiple icebergs, we adopt a feature extraction method based on the centroid distance function (Mingqiang et al., 2008;Hasim et al., 2016;Arjun and Mirnalinee, 2015). As described in Figure 3, the semi-automated iceberg tracking process consists of the following five steps: (1) We define the target 200 iceberg that will be tracked. We manually digitize the polygon of the reference iceberg from a Sentinel-1 image and calculate the distances from the iceberg centroid for all iceberg pixels. We make a histogram of these centroid distances for all pixels in the iceberg; this centroid distance histogram (CDH) represents the area and shape of the iceberg. (2) We search for an available Sentinel-1 image of the next day, within a 25 km radius from the reference iceberg in the current image. If the image for the next day is not available, the search radius is extended by 25 km each day until a Sentinel-1 image is found (Figure 4a).

(3) Once a 205
Sentinel-1 image is found, all candidate icebergs are detected by the same processes as described in 3.1. (4) We calculate the CDHs for all detected icebergs. (5) By comparing the similarities of the CDHs between these icebergs and the reference iceberg, we determine which iceberg is the target iceberg for that day. The similarity of CDH between iceberg A and reference iceberg where m is the bin size of the histogram, ( ) and ( ) is are the number of pixels in the -th bin for iceberg A and reference iceberg, respectively. We select the iceberg showing similarity greater than 80 % with the reference iceberg ( Figure 4b-4ed, see Appendix A). If multiple icebergs show similarity above 80 % with the reference iceberg, we select the iceberg with the highest similarity as the target iceberg. Once the target iceberg is determined, the centroid of this iceberg becomes the center of search radius for the next day in the second step. However, if the target iceberg is not detected, we return to the second step, expand the 215 search radius, and repeat the same steps until the target iceberg is detected (Figure 3).
It is worth noting that for some days, when a significant mass of the iceberg breaks from the iceberg, the shape of the iceberg changes considerably. In these cases, the initial CDH would no longer be representative of the iceberg and a new reference iceberg would have to be manually set. Three examples of these manual reference adjustments are needed in this study, including days 18 March 2018, 25 March 2019, and 17 January 2020. In addition, if a Sentinel image is not found for more than five 220 successive days, the search radius becomes too large to be processed in the limited memory of GEE. In that case, we must manually find the next available Sentinel image and day, and 40 such images are manually found in this study. Thus, for processing the total of 433 images, 43 manual interventions (3 manual reference adjustments + 40 manual image search) are required (i.e. ~90 % of the total processes are automated). Because of this need of occasional human intervention, we call our overall procedure semiautomated method. 225

Calculation of iceberg thickness
The freeboard of the iceberg (F) is defined as the difference between the heights of the iceberg body (H) and the sea surface heights (H ssh ) by the following equation: For the CryoSat-2 data, we use the sea surface heights from the Level 2 SAR and SIN products, and for the ICESat-2 data, we use the sea surface heights from the ICESat-2 ATL10 sea ice freeboard products. We subtract these sea surface heights from the height measurements of the iceberg. Then the thickness of the iceberg (Z) can be roughly estimated from the freeboard by using Equation 5: where and is the density of water and ice, respectively, and δ is the thickness correction of the upper firn layer. Here we assume 1,025 kg/m 3 and 915 kg/m 3 of seawater and ice density, respectively (Griggs and Bamber, 2011;Chuter and Bamber, 2015;Zhang et al., 2020), and 15 m of typical firn layer correction (Griggs and Bamber, 2011;Lythe and Vaughan, 2001).

Area variations of the iceberg
The iceberg area is retrieved by using our semi-automated iceberg tracking method (Figure 5a). We find three noteworthy periods, all in summers or end-of-summers, when large reductions of the iceberg area are observed: (1)  Although our iceberg tracking method extracts a good monthly mean trend of the iceberg area, there are several anomalous estimates of the individual area at a daily level (Figure 5a). One cause for this is the nearby small icebergs. We observe a number of small icebergs around the coastal Amundsen Sea region, and they are sometimes adjacent to the iceberg B43. In this case, our iceberg-detecting algorithm recognizes them as a single iceberg body, which leads to overestimation of the iceberg area. Second, although we assume a high radar backscatter from the iceberg, this backscatter can vary by sensor incidence angle, iceberg surface, 265 and weather conditions. The variations in backscatter of the iceberg can cause uncertainties in area estimation. In addition, if sea ice has a high backscatter, the similar backscatter between sea ice and the iceberg makes it difficult to distinguish the exact boundary between them (Mazur et al., 2017). Finally, we need to consider the spatial uncertainties introduced when SAR images are projected into the map coordinates and are resampled into 40 m resolution.
We compare the area changes with the ancillary meteorological data (Figure 5b-5d). It is noted that three split-up events 270 occurred only in lower SIC areas and during higher air temperatures (Figure 5b and 5c). In higher SIC areas, the iceberg remains relatively stable because it is surrounded by sea ice. However, when SIC decreases and the iceberg is exposed to open water, the interaction between waves and the iceberg can trigger the breakup of the iceberg (England et al., 2020;Scambos et al., 2008;Wagner et al., 2014). Given that this breakup mechanism is somewhat related to the water temperature and waves (Wagner et al., 2014), we deduce that SIC and temperature can have substantial impacts on the breakup events. In particular, the iceberg was 275 completely broken into several small pieces in March 2020, after being continually exposed to open water since December 2019.
There is a possibility that the rising wave heights during this period accelerate the breakup event ( Figure 5d) (MacAyeal et al., 2006). Nevertheless, considering the complexity of the breakup mechanism, it is challenging to clearly determine what caused the breakup of the iceberg B43.

Trajectory of the iceberg
The trajectory of iceberg B43 resulted from the semi-automated iceberg tracking is shown with solid colored circles in Figure 56. The trajectory agrees well with that from the National Ice Center (NIC)NIC iceberg archive which is reproduced with a black dashed line in Figure 56. In contrast to our semi-automated tracking procedure, the NIC trajectory is manually retrieved 290 by ice analysts every week using a combination of SAR, visible, and infrared remote sensing images (usicecenter.gov). As shown in Figure     As B43 moved away from the starting point, however, its speed increased. From April 2018 to May 2019, the iceberg 320 mainly moved westward parallel to the coastline more than 90 % of the time, which is consistent with the westward sea ice drift and wind velocity (Figure 67c-67e). During this journey from the Amundsen Sea to the eastern Ross Sea, the drift speed was faster than 4 km/day approximately 70 % of the time, and even greater than 8 km/day 10-40 % of the time. This increase of drift speed can mainly be attributed to the faster sea ice movement in this region compared to the starting point in the Amundsen Sea.
When B43 reached the Ross Sea (June 2019), it started to drift more in a northwestward direction (Figure 67e and 67f). 325 From April 2019 to October 2019 the iceberg drifted at its fastest speed as it passed through an area with the fastest sea ice motion ( Figure 67e). It is also noted that the directional change from westward to northwestward during this period agrees with the direction of sea ice movement and wind direction in this region (Figure 67e). It is generally known that sea ice starts to move faster from this region because sea ice is thinner in the Ross Sea compared to the Amundsen Sea, so sea ice is more affected by wind forcing (DeLiberty et al., 2011). Although the drift speed slightly decreased after November 2019, the iceberg was still observed 330 to drift at > 8 km/day around 30% of the time for the last five months (Figure 6f). During the last five months, sea ice may have had less impact on the drift of B43 because it was surrounded by thinner sea ice (DeLiberty et al., 2011) or open water. Figure 78 compares the drift speed and the direction of the iceberg with SIC, wind velocity, and sea ice motion. Here we merely focus on the qualitative comparison of the drift direction and speed with other climatological data. Considering that this iceberg moved along with highly compacted sea ice during most of its journey (SIC > 80 %, Figure 78a), the movement of the 335 iceberg is likely to be related to the sea ice drift and wind velocity (Lichey and Hellmer H., 2001;Schodlok et al., 2006). Wind velocity and ice motion generally show similar variations for each other, and they both show substantial correlation with the iceberg drift (Figure 78b-78e). However, the impact of wind or sea ice drift on the iceberg drift could vary depending on the season, location, SIC, and ocean current.
Although we do not directly analyze ocean current data here due to the lack of reliability and availability of ocean current 340 data under high-concentration sea ice, ocean current would account for a significant part of the iceberg drift because about 80-90 % of the iceberg is under the sea surface. Hence, the westward drift of this iceberg mainly represents the sea current along the Amundsen Sea to the Ross Sea. In the coastal regions of the Amundsen Sea and the Ross Sea, the ocean circulation is dominated by the westward Antarctic Coastal Current (ACoC) which is mainly driven by wind stress and buoyancy (Whitworth III et al., 1985;Orsi et al., 1995;Mathiot et al., 2011;Kim et al., 2016;Stern et al., 2016). 345 north-south velocity of both the iceberg (black crosses) and sea ice motion (magenta circles). Gray shaded areas indicate the splitoff events of the iceberg.

Area variations of the iceberg
The iceberg area is also retrieved by using our semi-automated iceberg tracking method (Figure 8a). We find three 355 noteworthy periods, all in summers or end-of-summers, when large reductions of the iceberg area are observed: (1)  Although our iceberg tracking method extracts a good monthly mean trend of the iceberg area, there are several anomalous estimates of the individual area at a daily level (Figure 8a). One cause for this is due to the nearby small icebergs. A number of 365 small icebergs around the coastal Amundsen Sea region are sometimes adjacent to the iceberg B43 (Appendix A). In this case, our iceberg-detecting algorithm recognizes them as a single iceberg body, which leads to overestimation of the iceberg area. Second, although we assume a high radar backscatter from the iceberg, this backscatter can vary by sensor incidence angle, iceberg surface, and sometimes weather conditions (Wesche and Dierking, 2012). The variations in backscatter of the iceberg can cause uncertainties in area estimation. In addition, if sea ice has a high backscatter, the similar backscatter between sea ice and the iceberg 370 makes it difficult to distinguish the exact boundary between them (Mazur et al., 2017). Finally, we need to consider the spatial uncertainties introduced when SAR images are projected into the map coordinates and are resampled into 40 m resolution.
We compare the area changes with the ancillary meteorological data (Figure 8b-8d). It is noted that three split-up events occurred only in lower SIC areas and during higher air temperatures (Figure 8b and 8c). In higher SIC areas, the iceberg remains relatively stable because it is surrounded by sea ice (Stuart and Long, 2011b). However, when SIC decreases and the iceberg is 375 exposed to open water, the interaction between waves and the iceberg can trigger the breakup of the iceberg (England et al., 2020;Scambos et al., 2008;Wagner et al., 2014). Given that this breakup mechanism is somewhat related to the water temperature and waves (Wagner et al., 2014), we deduce that SIC and temperature can have substantial impacts on the breakup events. In particular, the iceberg was completely broken into several small pieces in March 2020, after being continually exposed to open water since December 2019. There is a possibility that the rising wave heights during this period accelerate the breakup event ( Figure 8d) 380 (MacAyeal et al., 2006). Nevertheless, considering the complexity of the breakup mechanism, it is challenging to clearly determine which factor most contributed to the entire breakup of the iceberg B43 in March 2020.

Freeboard and thickness of the iceberg
To estimate the freeboard and thickness of the iceberg, we search for coincident CryoSat-2 and ICESat-2 tracks overlapping with the iceberg. We find one overlapping CryoSat-2 track and four overlapping ICESat-2 tracks that can be used to estimate the freeboard of the iceberg (Table 1). Although there are time differences between the Sentinel-1 images and the altimeter 395 measurements, we conclude that these CryoSat-2/ICESat-2 tracks include the heights of the iceberg by comparing the drift of the iceberg and the height measurements. For example, Figure 9 shows the Sentinel-1 image on 31 March 2019, and the CryoSat-2 track on the same day. Although the higher points of CryoSat-2 do not exactly correspond to the iceberg in the Sentinel-1 image due to the time differences between the two datasets (around 14 hours), we can assume that these higher points represent the surface of the iceberg when the movement of the iceberg is considered. In particular, in the Sentinel-1 image, we cannot see any potential 400 40 m-height objects except the iceberg. Furthermore, near the iceberg, the CryoSat-2 points are off the straight ground track line, which indicates that there is an object different from the surrounding sea ice or open water. Since the retracking algorithm of the CryoSat-2 SIN product may emphasize a late-returned signal (Wingham et al., 2006), the radar signal from the edge of the iceberg can be biased to the non-iceberg area. Therefore, we assume that the heights represent the surface heights for the iceberg, and the mean freeboard is 42.4 m for this date (Table 1). 405 While CryoSat-2 has only one coincident track with the iceberg, there are four coincident ICESat-2 tracks passing through the iceberg (Table 1). Figure 10 shows an example of the ICESat-2 ATL03 data on 29 November 2019, and the Sentinel-1 image on the next day. Although they have time differences of about 24 hours, the heights of the ATL03 tracks are likely to represent the heights of the iceberg at the time of the Sentinel-1 image. It is interesting that ICESat-2 data describe the surface height of the iceberg in more detail with some ridges and valleys ( Figure 10). This is attributed to a better spatial resolution of the ICESat-2 410 ATL03 product (117 m of footprint spaced by 0.7 m in the along-track direction (Magruder et al., 2020)) than the CryoSat-2 SAR or SIN product (~400 m of resolution in the along-track direction). In addition, the multiple tracks of ICESat-2 can be a considerable advantage for examining the shape of the iceberg surface.
As shown in Table 1, the iceberg freeboard is estimated to be greater than 50 m in November 2018, but less than 50 m from then on. There is not high confidence, hHowever, that the thickness of this iceberg decreased during the drift because the 415 limited number of height measurements is not sufficient to help us to findextract a significant trend of freeboard change. Moreover, since each track passed through a different part of the iceberg, the heights also depend on the sampling points where CryoSat-2 or ICESat-2 track were passing by. Nevertheless, we can safely deduce that the freeboard of this iceberg is was within 30-60 m during the tracking period. This range of freeboard agrees with the previous estimates for other large icebergs over the Antarctic (Tournadre et al., 2015;Scambos et al., 2005;Romanov et al., 2012;Han et al., 2019). As shown in Table 1, this level of freeboard 420 is equivalent to the iceberg thickness of 150-400 m (Equation 5). This implies that about 120-350 m of the iceberg thickness (i.e. 80-90 % fraction of the total thickness) is below the water surface. Thus, considering the average thickness (~265 m), initial area (~105 km 2 ), and 15 m firn layer of this iceberg, the melting of the entire iceberg body will contribute to approximately 24 Gt of freshwater input into the Southern Ocean. This amount of freshwater is equivalent to about 1 % of total annual freshwater flux in the Southern Ocean (Hammond and Jones, 2016). 425

Conclusions
This study demonstrates for the first time the potential of Google Earth Engine (GEE) for iceberg tracking in the Southern 440 Ocean using Sentienl-1 SAR images. It presents a cloud-based computational method for tracking iceberg B43 from when it calved off in the Amundsen Sea coast until it broke up in the Ross Sea. First, iceberg B43 is detected by using a SNIC object-based image segmentation and backscatter threshold algorithm. Then, the trajectory of this iceberg is constructed by comparing its centroid distance histogram with those of icebergs in subsequent images. Although manual tracking iswas used for several dates due to memory limitations of GEE and occasional large changes in the iceberg shape, the method (with a ~90% automation) was is 445 successful in tracking B43 for three years from April 2017 to March 2020.
Factors affecting the decrease in the area of iceberg B43 over time were are studied by comparison with contemporaneous changes in environmental variables from ERA5 reanalysis data and NSIDC sea ice motion data. First, in terms of SIC, we find that split-up events of the iceberg occurred with lower SIC and higher exposure to open water conditions. This suggests that wave action, as calculated in ERA 5, played a role in triggering the break-up of the iceberg. Our analysis suggests that higher water 450 temperature and higher wave heights can also accelerate this break-up mechanism. In addition, we find that iceberg drift could be associated with sea ice drift or wind velocity. The speed and direction of the iceberg drift agrees with the variations of sea ice drift and winds in the Amundsen Sea and Ross Sea regions. Given that this iceberg is was 150-350 m in thickness, with the majority of its volume underwater, the main contribution to the westward drift of the iceberg is thought to be the westward sea current (ACoC) Predicting the behavior of icebergs (e.g. trajectory, speed, decay, and breakup) is a highly complex endeavor that depends on many factors related to dynamic and thermodynamic interactions between ocean currents, the atmosphere, waves, sea ice, and bottom topography (Stern et al., 2016;Rackow et al., 2017;Schodlok et al., 2006;Lichey and Hellmer H., 2001;Gladstone et al., 2001). Thus, developing, validating, and calibrating models for predicting iceberg's variables would require a large and accurate iceberg database containing the life history of many icebergs. The method presented here can be used for these purposes by taking 460 advantage of GEE's large data storage capability and high computing power. Indeed, GEE allowsed us to process 433 Sentinel-1 SAR images in the cloud, taking only 10-30 seconds to process and analyze one day of data and only a few hours to process all three years. Moreover, since all image-processing tools arewere provided by GEE, this approach iswas able to save time, cost, and resources as compared to a similar traditional tracking done in a local computer. Therefore, iceberg tracking based on Sentinel-1 images and the GEE platform would be an efficient option for tracking a large number of icebergs in the polar oceans. 465 To build a large database of icebergs using the GEE-based iceberg tracking approach presented in this paper, however, two three key limitations would need to be overcome. First, in terms of the image segmentation, although the SNIC algorithm in GEE detects iceberg B43 successfully (Figure 2), this superpixel approach should be tuned for a further application to small icebergs. In this study, the surface roughness and shape of iceberg B43 have no significant impacts on the image segmentation because the iceberg is large and superpixels are also large (i.e., seed = 80). However, for detecting small icebergs, the surface 470 shapes of icebergs can have significant impacts on the image segmentation and iceberg detection because the superpixels should be small. Therefore, further enhanced image segmentation process may be required for the detection of small icebergs. Second, although the centroid distance histogram (CDH) based tracking approach successfully tracks works for iceberg B43 (thanks to the an iceberg with a distinctive shape and large size of the B43), this method has limitations in tracking small icebergs that share similar shapes or areas. Indeed, we attempted unsuccessfully to track the small pieces of B43 when it broke up (Figure 1c) because 475 these pieces have similar CDHs. On the other hand, we also note that our tracking method may have difficulty tracking icebergs that are too large to be captured within a single Sentinel-1 scene. Second,Third, our tracking algorithm is not fully-automated. In this study, instead of defining the detected iceberg from the previous day as the reference for the next-day tracking, we defined the reference iceberg at the initial digitizing step and use this reference for all the following steps. Although this approach enables us to avoid tracking errors caused by temporary anomalies (e.g. impacts of surrounding small icebergs), it prevents our algorithm 480 from detecting sudden changes in the iceberg shape or area. By overcoming these limitations, fully-automated tracking of small icebergs would be possible and a large and accurate iceberg database could would be constructed in the future.

Appendix A. Variations of similarity values 485
In this study, we track the target iceberg that has > 80 % of similarity with the reference iceberg. This appendix is for discussing how these similarity scores change and what factors potentially affect the similarity scores. Figure A1 shows the area of iceberg B43 and the similarity of CDH with the reference iceberg. 80 % threshold of CDH similarity is appropriate for tracking this iceberg, but the similarity decreased near the split-off events. Hence, if these split-off events reduce the similarity to < 80 %, it is necessary to newly define the reference iceberg. Figure A2 shows the Sentinel-1 images for six low-similarity dates in Figure  490 A1(b). In Figure A2, we observe three factors that are likely to lower the CDH similarity between the target iceberg and reference iceberg: (1) surrounding small icebergs (A, B, C, E in Figure A2), (2) split-off events (F in Figure A2), and (c) uncertainty from the image segmentation because of confusion between shaded part of the iceberg and surrounding bright sea ice (B, D in Figure   A2).   Figure A2. Sentinel-1 SAR images and SNIC segmentation results for six low-similarity times A-F in Figure A1.
Code and data availability. Python code for this GEE-based semi-automated iceberg tracking is available at: https://github.com/YoungHyunKoo/GEE_iceberg_tracking.git. 505 Author contributions. YK designed the research, conducted programming and analysis, created figures, and drafted the initial manuscript. HX, SA, AM, GM, CH contributed to writing and editing of the manuscript.
Competing interests. The authors declare that they have no conflict of interest. 510