DOI QR코드

DOI QR Code

Tracking Large Tabular Icebergs in Antarctica Using AMSR2 Observations and Random Forest:A Case Study of A23A

  • Minji Cho (Interdisciplinary Program in Earth Environmental System Science & Engineering, Kangwon National University) ;
  • Hyewon Song (Interdisciplinary Program in Earth Environmental System Science & Engineering, Kangwon National University) ;
  • Sumin Jeong (Division of Geology and Geophysics, Kangwon National University) ;
  • Hyangsun Han (Department of Geophysics, Kangwon National University)
  • Received : 2025.04.07
  • Accepted : 2025.04.23
  • Published : 2025.04.30

Abstract

Icebergs can significantly influence sea ice-ocean environments and pose serious hazards to maritime navigation in polar regions. Therefore, detecting and tracking icebergs is essential for scientific and operational purposes. This study investigates the feasibility of using passive microwave observations to detect and track large tabular icebergs in Antarctica, focusing on the iceberg A23A in the Weddell Sea. A Random Forest (RF) iceberg detection model was developed using the Advanced Microwave Scanning Radiometer 2 (AMSR2) observations. Reference datasets were constructed by combining Sentinel-1 SAR images with co-located AMSR2 pixels to label iceberg and non-iceberg. Brightness temperatures from all AMSR2 channels, along with auxiliary parameters commonly used in sea ice concentration retrieval algorithms, were employed as input variables for model development. The RF model showed an F1-score of 0.84 on an independent test dataset, indicating reasonably robust performance. While the grid size of AMSR2 data (10 × 10 km) limited the accurate delineation of iceberg area, the model effectively identified the location of A23A and tracked its drift over time. Among the input variables, the difference between horizontally and vertically polarized brightness temperatures at 89 GHz was significant for distinguishing iceberg from surrounding sea ice and open water. The central positions of A23A estimated by the RF model from July 2012 to September 2023 showed a root mean square deviation of 11.3 km compared with data from the Antarctic Iceberg Database. Considering the size of A23A (~65 × 70 km) and the grid size of the AMSR2 data, this deviation is relatively small. However, the RF model failed to detect A23A after September 2023, likely due to changes in its microwave radiation properties as it drifted into lower latitudes and underwent surface melting.

Keywords

1. Introduction

Icebergs are fragments of ice that have calved from the termini of ice shelves. As they originate from glacial ice, icebergs exhibit distinct physical properties compared to sea ice. Consequently, the drift and melting of icebergs can influence ocean salinity, current circulation, sea ice formation processes, and marine ecosystems (Bigg and Marsh, 2023; Ingels et al., 2021; Wu and Hou, 2017; Merino et al., 2016; Arrigo et al., 2002; Enderlin et al., 2023; Schloesser et al., 2019). The evolution of an iceberg from formation and dissipation not only reflects changes in the polar marine environment but also acts as a valuable proxy for climate change. Moreover, iceberg drift poses significant hazards to navigation in polar oceans. Therefore, the detection and tracking of icebergs are critically important for both scientific and operational applications.

Due to their wide-ranging movement in polar oceans, satellite remote sensing has been widely used for iceberg detection and tracking. Microwave remote sensing is particularly suitable for iceberg monitoring as it can observe the Earth’s surface regardless of weather conditions and sun altitudes. Synthetic aperture radar (SAR), scatterometers, and passive microwave sensors are commonly used in this context. SAR provides high spatial resolution (meterto tens-of-meters)images, allowing for precise detection of icebergs(Han et al., 2019; Barbat et al., 2021;Wesche and Dierking, 2012; Koo et al., 2021).However, despite overlapping swaths in polar regions, the revisit cycles of SAR satellites limit their capability to observe drifting icebergs at short temporal intervals. Scatterometers offer lower spatial resolution (several kilometers), but they can observe the entire polar region every 1–2 days and are effective for iceberg tracking due to the higher backscattering signals from icebergs compared to sea ice and open water(Stuart and Long, 2011; Budge and Long, 2018; Singh et al., 2022). However, when icebergs’ surfaces undergo melting, their microwave scattering characteristics can resemble those of surrounding sea ice or open water, complicating detection and tracking (Budge and Long, 2018).

Passive microwave sensors have low spatial resolution, ranging from several to a few tens of kilometers, but they offer daily, allweather coverage of the polar regions. Their multi-frequency brightness temperature observations provide the potential to detect and track large icebergs, even during the summer melt season when microwave scatterometers face limitations. Despite this potential, the application of passive microwave observations for iceberg monitoring remains underexplored, with the sole notable study being by Phillips and Laxon (1995), who visually tracked the icebergs A20A and A20B using 85 GHz brightness temperature from Special Sensor Microwave/Imager.

Detecting and tracking icebergs using passive microwave observations is also essential for enhancing sea ice products such assea ice concentration and extent, which are critical parameters in climate-related research (Meier et al., 2022; Bliss et al., 2019; Melsheimer et al., 2023; Ivanova et al., 2015). Since icebergs have different microwave radiation signatures compared to sea ice, failure to mask iceberg-contaminated pixels can lead to erroneous sea ice parameter retrievals (Kern et al., 2007; Han and Lee, 2011). Fig. 1 shows this issue using sea ice concentration retrieved from the Arctic Radiation and Turbulence Interaction STudy (ARTIST) Sea Ice (ASI) algorithm (Spreen et al., 2008) based on AMSR2 observations over the iceberg A23A on July 16, 2018, in the Weddell Sea. Although the iceberg should appear as 0% sea ice concentration, values ranging from60%to 85%were incorrectly retrieved. Such erroneous sea ice concentration retrievals can occur across the Antarctic Ocean, emphasizing the importance of detecting iceberg locations from passive microwave observations to improve the accuracy of sea ice products.

OGCSBN_2025_v41n2_1_341_2_f0001.png 이미지

Fig. 1. Iceberg A23A captured by Sentinel-1 SAR and its effect on sea ice concentration retrieved from AMSR2 observations. (a) Sentinel-1 SAR image acquired on July 16, 2018. (b) Sea ice concentration derived from the ASI algorithm using AMSR2 observations on the same date.

The microwave radiative properties of icebergs can vary with seasonal melt and refreeze conditions, and over longer periods, are affected by changes in physical properties such as ice thickness and internal structure due to climate change. Therefore, iceberg detection models must account for these complex and dynamic characteristics. Machine learning approaches can offer promising solutions for this purpose. Machine learning techniques have been applied to retrieve sea ice concentration, sea ice types, and sea ice thickness from passive microwave observations, outperforming traditional physics-based algorithms (Han et al., 2021; Chi et al., 2019; Chi and Kim, 2021; Radhakrishnan et al., 2021; Dong et al., 2022). However, no prior study has combined passive microwave observations with machine learning techniques for iceberg detection and tracking.

In this study, we investigate the feasibility of detecting and tracking large tabular icebergs in Antarctica using multi-frequency brightness temperature observations from passive microwave sensors. Specifically, we develop a Random Forest(RF) model to detect the iceberg A23A, one of the large tabular icebergs in Antarctica, using observations from the AMSR2. We evaluate the performance of the RF model and apply it to the time-series AMSR2 dataset to track the drift of A23A, thereby demonstrating the potential of passive microwave observations for large-scale iceberg monitoring.

2. Materials

2.1.Iceberg A23A

The A23A (Fig. 1a) originated from a larger iceberg, A23, which calved from the Filchner-Ronne Ice Shelf in the Weddell Sea, West Antarctica, in 1986.At the time of its formation, A23A was a massive tabular iceberg, approximately 400 m thick and an area of ~4,000 km² (Stulic et al., 2023; Budge and Long, 2018). The base of the iceberg became grounded on the seafloorin a coastal region typically covered by thick multiyear sea ice, and it remained stationary for decades following 1991 (Schodlok et al., 2006; Stulic et al., 2023; Grosfeld et al., 2001). Over time, A23A gradually thinned, and in May 2022, it was observed to begin drifting westward, carried by the Weddell Sea Gyre. As of March 3, 2025, A23A has been grounded on a shallow seafloor near South Georgia Island (54.6°S, 38.7°W) (BBC, 2025). While grounded along the Antarctic coast, A23A had a substantial influence on the local sea ice-ocean environment (Stulic et al., 2023). During its drift, the iceberg likely altered ocean salinity and affected regional sea ice formation conditions.

2.2. Sentinel-1 SAR Images

To construct the reference dataset for developing the AMSR2- based iceberg detection model, SAR images acquired by the Sentinel-1A and Sentinel-1B satellites were utilized. The European SpaceAgency (ESA)launched Sentinel-1A in April 2014,followed by Sentinel-1B in April 2016. Both satellites are equipped with C-band SAR operating at a center frequency of 5.55 GHz (Torres et al., 2012). Each satellite has a 12-day revisit cycle, which allows for a 6-day temporal resolution when both are operational. However, Sentinel-1B ended its mission due to a power supply issue in December 2021. Sentinel-1C was launched in December 2024 and is currently undergoing satellite commissioning activities. Sentinel-1 provides SAR images in four acquisition modes— Stripmap, Interferometric Wide swath, Extra Wide swath (EW), and Wave—with single or dual polarization.

In this study, a total of 786EW mode SAR images from Sentinel1A/B, acquired between October 3, 2014, and September 25, 2023, were used to define the location and extent of A23A. Only SAR images acquired within 30 minutes of corresponding AMSR2 observations were selected to minimize positional discrepancies due to iceberg drift driven by ocean currents and wind. All SAR images used in this study are HH-polarized Ground Range Detected (GRD) products with a grid spacing of 40 m. These images were preprocessed using standard procedures, including orbital correction, radiometric calibration, thermal noise removal, and speckle noise reduction via multi-looking.

2.3. GCOM-W AMSR2 Data

AMSR2, a successor to the Advanced Microwave Scanning Radiometer for Earth Observing System (AMSR-E), is a passive microwave sensor onboard the Global Change Observation Mission-Water (GCOM-W) satellite, launched in May 2012 by the Japan Aerospace Exploration Agency (Imaoka et al., 2010). AMSR2 measures brightness temperatures in both horizontal(H) and vertical (V) polarizations at multiple frequency channels centered at 6.925, 7.3, 10.65, 18.7, 23.8, 36.5, and 89.0 GHz (Maeda et al., 2016). Table 1 summarizes the specifications of AMSR2 frequency channels.

Table 1. Specifications of AMSR2 frequency channels (Maeda et al., 2016)

OGCSBN_2025_v41n2_1_341_3_t0001.png 이미지

To develop the iceberg detection model, a total of 786 AMSR2 Level-1R products acquired from October 21, 2014, to April 10, 2023, were used. These products were chosen to ensure a time difference of less than 30 minutes from corresponding Sentinel1 SAR images, minimizing potential spatial mismatches caused by iceberg drift. The AMSR2 Level-1R products provide brightness temperatures as swath data that have been radiometrically and geometrically corrected (Maeda et al., 2016). All brightness temperatures in Level-1R products are resampled with equal footprint size for each frequency and provided to a 10 × 10 km grid size. Fig. 2 shows brightness temperatures from each AMSR2 frequency channel acquired within 30 minutes of the Sentinel-1 SAR image shown in Fig. 1(a).

OGCSBN_2025_v41n2_1_341_4_f0001.png 이미지

Fig. 2. Brightness temperatures (TB) from AMSR2 channels around A23A on July 16, 2018. (a) TB at 6.925 GHz horizontal polarization (H-pol). (b) TB at 7.3 GHz H-pol. (c) TB at 10.65 GHz H-pol. (d) TB at 18.7 GHz H-pol. (e) TB at 23.8 GHz H-pol. (f) TB at 36.5 GHz H-pol. (g) TB at 89.0 GHz H-pol. (h) TB at 6.925 GHz vertical polarization (V-pol). (i) TB at 7.3 GHz V-pol. (j) TB at 10.65 GHz V-pol. (k) TB at 18.7 GHz V-pol. (l) TB at 23.8 GHz V-pol. (m) TB at 36.5 GHz V-pol. (n) TB at 89.0 GHz V-pol. The red polygon in each image indicates the boundary of iceberg A23A. The black lines represent the coastline and grounding lines.

To track A23A using the developed iceberg detection model, AMSR2 Level-3 daily brightness temperature products in polar stereographic projection with a 10 × 10 km grid size were compiled at monthly intervals from July 1, 2012, to December 1, 2024. The AMSR2 Level-3 products are gridded versions of the Level-1R swath data, separated by ascending and descending orbital passes. In this study, only descending orbit observations were used.

2.4. Positions of A23A from the Antarctic Iceberg Database

To evaluate the positional accuracy of A23A as detected by the AMSR2-based iceberg detection model, time-series A23A position datawere obtained from the Antarctic Iceberg Tracking Database (https://www.scp.byu.edu/data/iceberg/database1.html). This database provides iceberg track information across the Antarctic region from 1978 to the present, integrating satellite scatterometer observations and iceberg monitoring records from the U.S. National Ice Center (Budge and Long, 2018).

3. Methodology

3.1. Construction of Reference Dataset and Input Variables for the Iceberg Detection Model

To construct a reference dataset for iceberg detection, that is, to distinguish AMSR2 pixels that contain theA23Afrom those that do not, Sentinel-1 SAR images were used. The grid size of the AMSR2 Level-1R products is 10 × 10 km, which means that each pixel can include a mixture of iceberg, sea ice, and open water. Therefore, it is essential to select AMSR2 pixels in which the microwave radiation characteristics of icebergs are dominant. In passive microwave remote sensing, a sea ice concentration threshold of 15–30% is commonly used to distinguish between sea ice and open water pixels (Ji et al., 2018; Parkinson and Cavalieri, 1989;Worby and Comiso, 2004), based on the principle that a minimum fraction of sea ice should be present in a pixel forits microwave radiation signature to be detectable. In a similar manner, we labeled an AMSR2 pixel as ‘iceberg’ if more than 30% of the pixel area was occupied by A23A, and the remaining AMSR2 pixels were labeled as ‘non-iceberg.’ To determine this, the boundary of A23A in each Sentinel-1 SAR image was manually digitized based on the contrast in backscattering intensity between the iceberg and surrounding surfaces. This process allowed for the reliable delineation of the iceberg extent across all SAR images used for the construction of the reference dataset. The digitized A23A region was then spatially overlaid onto the AMSR2 grid, and for each AMSR2 pixel (10 × 10 km), we calculated the proportion of its area covered by the Sentinel1-derived iceberg region.

As input variables for the iceberg detection model, we used brightness temperatures (TB) at all AMSR2 frequency channels for both horizontal(H) and vertical(V) polarizations. In addition, we included five parameters commonly used in sea ice concentration retrieval algorithms. These include a polarization ratio at 18 GHz (PR(18)), spectral gradient ratios (GR) at vertical polarization between37GHz and 18GHz (GR(37V18V)) and between23GHz and 18 GHz (GR(23V18V)), the difference between horizontal and vertical polarization gradient ratios at 89 and 18 GHz (∆GR), and the difference in TB measured at 89 GHz vertically and horizontally polarized channels (P(89)) (Cavalieri et al., 1984; Markus & Cavalieri, 2000; Spreen et al., 2008). PR(18), GR(37V 18V), and GR(23V18V) are key parameters in the NASA Team algorithm for sea ice concentration retrieval (Cavalieri et al., 1984). PR(18),GR(37V18V) are useful for distinguishing between sea ice types and open water, while GR(23V18V) is used to account for atmospheric effects embedded in the TBs (Cavalieri et al., 1984; Markus & Cavalieri, 2000; Spreen et al., 2008; Markus and Dokken, 2002). ∆GR and P(89) are introduced in the NASA Team2 (Markus & Cavalieri, 2000) and ASI (Spreen et al., 2008) algorithms, respectively, to enhance sea ice concentration retrieval accuracy under complex surface conditions such as inhomogeneous surface layering (Markus and Dokken, 2002). These parameters are calculated as

\(\begin{align}P R(18)=\frac{T_{B}(18 \mathrm{~V})-T_{\mathrm{B}}(18 H)}{T_{B}(18 V)+T_{B}(18 \mathrm{H})}\end{align}\)       (1)

\(\begin{align}G R\left(f_{1 p} f_{2 p}\right)=\frac{T_{B}\left(f_{1 p}\right)-T_{B}\left(f_{2 p}\right)}{T_{B}\left(f_{1 p}\right)+T_{B}\left(f_{2 p}\right)}\end{align}\)       (2)

∆GR = GR(89H18H – 89V18V)       (3)

P(89) = TB (89V) – TB (89H)       (4)

where f is the frequency and p is the polarization of the AMSR2 channel.

These parameters have been shown to exhibit distinct characteristics over glacial ice, making them valuable for discriminating icebergs from sea ice and open water (Han and Lee, 2011). Therefore, the combination of AMSR2 TBs and these five derived parameters is expected to provide a robust feature set for developing an effective iceberg detection model.

3.2. Development of Random Forest Iceberg Detection Model

A pixel-based binary classification model was developed using RF (Breiman, 2001), a rule-based machine learning algorithm, to distinguish iceberg pixels from non-iceberg pixels based on AMSR2 brightness temperature data. RF is an ensemble learning method based on bootstrap aggregating, which builds multiple unpruned classification and regression trees using bootstrapped samples from the original training data. For each tree, a random subset of training samples is drawn with replacement, and at each node, a random subset of input variables is selected for splitting. This process helps prevent overfitting and enhances model generalization by reducing sensitivity to data distribution and quality. In classification tasks, RF combines the predictions of multiple decision trees by selecting the class with the majority vote. It also provides the relative importance of input variables via the mean decrease in accuracy, which quantifies the increase in classification error when a specific variable is excluded from the model. RF is also computationally efficient, interpretable, and capable of handling high-dimensional input data. In this study, the model was implemented using the randomForest package in R software.

The performance of an RF model is largely influenced by two main hyperparameters: ntree, which determines the total number of trees constructed, and mtry, which specifies how many input features are randomly selected for consideration at each split. While increasing ntree often enhances model performance, the improvement tends to plateau beyond a certain threshold and may lead to higher computational costs. As such, careful tuning of ntree is necessary. Similarly, tuning mtry is important to ensure an appropriate trade-off between model complexity and generalization capability.

A total of 37,583 iceberg samples and 348,109 non-iceberg samples were compiled, reflecting the relatively small area of A23A compared to the surrounding sea ice and open water. Of these, approximately 70% (281,804 samples in total; 27,432 iceberg and 254,372 non-iceberg samples) were used for model training. The remaining 30% (103,888 samples in total; 10,151 iceberg and 93,737 non-iceberg samples) were used as a test dataset. The test data were sampled from AMSR2 observations in the years 2016, 2019, and 2023, while the training data were derived from other years, ensuring independence between the training and test datasets. To tune the RF hyperparameters, 30% of the training samples from each class(iceberg and non-iceberg) were randomly selected. The value of ntree was varied from 100 to 2,000 in increments of 100, and mtry was varied sequentially from 1 to 10. The optimal combination was selected based on the highest overall classification accuracy. The RF model was then trained using the full training dataset with the selected optimal values of ntree and mtry. To mitigate the effects of class imbalance, we applied stratified random sampling and incorporated class weighting during the RF model training. A higher weight was assigned to the iceberg class(set to 1), while a lower weight (0.5) was assigned to the non-iceberg class, to enhance sensitivity to minority samples.

Model performance was evaluated on the test dataset using standard classification metrics such as overall accuracy, precision, recall, and F1-score. Precision is the proportion of correctly predicted iceberg samples among all samples predicted as icebergs. Recall isthe proportion of true iceberg samples correctly identified by themodel. Overall accuracy represents the percentage of correctly classified samples among all test samples. The F1- score combines precision and recall into a single metric by calculating their harmonic mean. It is especially effective when evaluating model performance on datasets with class imbalance. Additionally, variable importance was analyzed to interpret the contribution of each input variable to the iceberg detection.

The RF model was applied to the monthly AMSR2 Level-3 daily brightness temperature products from July 2012 to December 2024. For each iceberg detection map, the median X and Y coordinates in polar stereographic projection of the detected A23A pixels were determined to define the centroid of the iceberg, and the drift trajectory was constructed. The detected positions of A23A were compared to the corresponding locations from the Antarctic Iceberg Tracking Database to assess the feasibility of the developed RF model for iceberg tracking.

4. Results and Discussion

4.1. Performance of the RF Iceberg Detection Model

Hyperparameter tuning was performed by randomly sampling 30%ofthe original training data. The highest overall classification accuracy (99.9%)was achieved with an ntree value of 100 and an mtry value of 9. Using this optimal configuration, the RF-based iceberg detection model was developed. All other parameters, except for the classwt used to specify class weight, were left at their default values as defined in the randomForest package in R. The model showed an overall accuracy of 97.1%, precision of 89.1%(user’s accuracy for iceberg), and recall of 79.9%(producer’s accuracy for iceberg), resulting in an F1-score of 0.84 (Table 2).

Table 2. Accuracy assessment metrics for the RF-based iceberg detection model

OGCSBN_2025_v41n2_1_341_6_t0001.png 이미지

Although the RF model demonstrated high overall accuracy, largely influenced by the 10:1 class imbalance between noniceberg and iceberg samples, approximately 20% of the A23A samples in the test dataset were misclassified as non-iceberg. This can be attributed to the reference data construction, in which an AMSR2 pixel was labeled as an iceberg only if more than 30% of its areawas occupied by A23A, as observed in the Sentinel-1 SAR images. When the iceberg coverage is close to this 30% threshold within a 10 × 10 km AMSR2 pixel, its microwave radiation signature may not be sufficiently dominant to differentiate it clearly from surrounding sea ice or open water. This ambiguity likely introduced classification uncertainty in the RF model. To improve model performance, the threshold for labeling iceberg pixels during reference dataset construction could be increased. Reconstructing the training data with a higher iceberg area requirement would allow more confident labeling of iceberg pixels and potentially enhance the robustness and accuracy of the iceberg detection model.

While the RF model does not achieve perfect performance in terms of statistical metrics, visual comparisons indicate that the A23A locations detected by the RF model closely matched those observed in the corresponding Sentinel-1 SAR images(Fig. 3).This suggests that, although the RF model may not precisely capture the spatial extent of A23A due to the grid size of AMSR2 Level1R data (10 × 10 km), it is sufficiently accurate for identifying the location of large tabular icebergs. However, some pixels were incorrectly classified as icebergs in regions where no icebergs were present(Fig. 3g). These misclassifications should be masked when applying the model to track A23A over time. Implementing such masking procedures would improve the accuracy and reliability of the iceberg drift trajectories derived from the RF model.

OGCSBN_2025_v41n2_1_341_7_f0001.png 이미지

Fig. 3. Iceberg A23A observed in Sentinel-1 SAR images and detected by the RF model. Sentinel-1 SAR image on (a) December 15, 2015, (b) December 24, 2017, (c) January 22, 2020, and (d) January 4, 2023. The RF model iceberg detection result corresponding to (e) December 15, 2015, (f) December 24, 2017, (g) January 22, 2020, and (h) January 4, 2023. The red polygon in each image indicates the boundary of iceberg A23A. The black lines represent the coastline and grounding lines.

4.2. Variable Importance of the RF Iceberg Detection Model

The five most important variables identified by the RF iceberg detection model were P(89), TB at 89.0 GHz H polarization, TB at 36.5 GHz H polarization, TB at 89.0 GHz V polarization, and ∆GR (Fig. 4). In general, both sea ice and icebergs, when unaffected by surface melt, exhibit higher TBs in the vertically polarized channels compared to the horizontally polarized channels (Kern et al., 2020). The relatively smooth and homogeneous surface of A23A may result in a lower TB in the horizontally polarized high-frequency channels and higher emissivity in the vertically polarized channels due to its high internal density, leading to elevated P(89) values. In contrast, surrounding multiyear sea ice typically exhibits higher TBs in the horizontally polarized channels compared to icebergs, with a smaller difference between horizontally and vertically polarized TBs (Han and Lee, 2011;Kern et al., 2020).Consequently, the P(89) values over A23A are generally higher than those over adjacent sea ice (Fig. 5), explaining its role as the most influential variable in the RF model. Similarly, TBs at 89.0 GHz in both H and V polarizations were ranked second and fourth most important variables, respectively, suggesting their significant contributionsto iceberg discrimination. Another possible reason forthe high importance of the 89 GHz-related variables is the finer native footprint (3 × 5 km) of the 89 GHz channels. Although these data were resampled to a 10 km grid, the original resolution may have preserved greater spatial detail and contrast than lower-frequency channels, thereby enhancing their effectiveness in distinguishing icebergs from surrounding sea ice and open water.

OGCSBN_2025_v41n2_1_341_8_f0001.png 이미지

Fig. 4. Mean decrease in accuracy of the RF model.

OGCSBN_2025_v41n2_1_341_8_f0002.png 이미지

Fig. 5. P(89) distribution around iceberg A23A on July 16, 2018. The red polygon indicates the boundary of iceberg A23A. The black lines represent the coastline and grounding lines.

The third most important variable was the TB at 36.5 GHz H polarization, which tends to show pronounced contrast among icebergs, sea ice, and open water (Fig. 2f), making it highly effective for the classification. The fifth-ranked variable was ∆GR. Icebergs typically exhibit higher values of both GRs compared to sea ice, resulting in a larger ∆GR (Han and Lee, 2011). This behavior can be attributed to the radiative transfer properties of thick glacial ice, where microwave emissions at 18.7 GHz originate from deeper ice layers and are attenuated as they pass through the surface. As a result, 18.7 GHz TBs measured over thick glacial ice such as icebergs are typically lower than those measured over thinner sea ice (Andersen et al., 2007; Jiang et al., 2007). Han and Lee (2011) also found that ∆GR values for ice shelves, the source of icebergs, tend to fall within a unique range that is higher than those of sea ice and lower than those of open water. These distinct physical characteristics explain the high importance assigned to ∆GR in the RF model.

4.3. Tracking of A23A based on the RF Iceberg Detection Model

The RF iceberg detection model developed in this study was applied to monthly subsets of AMSR2 Level-3 daily brightness temperature products from July 2012 to December 2024 over the Weddell Sea. To reduce false detections, contiguous regions classified as icebergs with an area smaller than 400 km² were masked priorto estimating the position of A23A. Fig. 6 presents the monthly positions of A23A as tracked by the RF model, along with corresponding reference positions from the Antarctic Iceberg Database. Between July 2012 and September 2023, the RF model-derived positions showed deviations ranging from 0.4 km to 18.8 km when compared to those in the Antarctic Iceberg Database, with a root mean square deviation of 11.3 km over the entire period. Given the size of A23A (~65 × 70 km) and the grid size of AMSR2 data (10 × 10 km),this deviation is considered remarkably small. However, after September 2023, the RF model failed to detect A23A. As of March 3, 2025, A23A had drifted northward and was grounded near South Georgia Island (54.6°S, 38.7°W), approximately 850 km away from its last detected location. This detection failure is likely due to substantial changes in the microwave radiation properties of A23A as it moved into lower latitudes and experienced increased surface melting, deviating from the microwave radiation signatures on which the model had been trained.

OGCSBN_2025_v41n2_1_341_9_f0001.png 이미지

Fig. 6. Trajectories of the iceberg A23A from July 2012 to September 2023 based on the RF model and positions from the Antarctic Iceberg Database.

Despite seasonal variation in microwave brightness temperature caused by surface melt and refreezing, the RF model successfully tracked A23A consistently from 2012 through 2023. From 2012 to mid-2022, A23A remained stationary due to grounding. It then began drifting northwest at an average speed of 1.6 km/day until February 2023. Between February and June 2023, the drift shifted slightly westward, with an increased speed of 2.5 km/day. From June to September 2023, A23A moved rapidly northward at an average speed of 5.8 km/day. As previously mentioned, the RF model could no longer track the iceberg beyond this point. Although other icebergs were present in the Weddell Sea, the RF model did not consistently detect and track them. This limitation is likely due to the model being trained solely on A23A, while other icebergs may exhibit different microwave radiation characteristics.

An advantage of the RF model lies in its ability to track the location of A23A without requiring explicit labeling or training for differentsea ice types or openwater conditions. This facilitates easier retraining and model updates. By adjusting the iceberg detection threshold (i.e., the minimum fractional coverage of iceberg with in an AMSR2 pixel) and incorporating samplesfrom the post-2023 period, along with data from additional icebergs and their surrounding environments, the model can be retrained to improve its accuracy and generalizability. Passive microwave imagery can be interpreted visually to identify icebergs. However, such manual inspection is inherently limited by subjectivity, low efficiency, and a lack of reproducibility across time and analysts. Compared to visual inspection of passive microwave imagery, the RF-based approach enables consistent, automated detection with improved temporal resolution and applicability to large-scale monitoring. This enhances its utility for large-scale monitoring and operational applications, particularly where human interpretation is infeasible or inefficient. The U.S. National Ice Center (NIC) and Brigham Young University operate an Antarctic iceberg tracking system that combines daily positions derived from satellite scatterometers with weekly iceberg reportsfrom the NIC, which are based on visible, infrared, and SAR imagery (Budge and Long, 2018).This operational database provides daily records of iceberg positions, sizes, and rotation angles. However, trajectory and velocity estimates currently require manual analysis of scatterometer imagery. This process, including the interpretation of the iceberg’s surrounding environment, can limit consistency and automation. In comparison, ourRF model enables automated and consistent detection and can be more readily scaled for largearea, high-frequency monitoring without the need for manual processing.

Future works will focus on expanding the spatial and temporal coverage oftraining and test datasets to improve model robustness. In addition, a passive microwave-based algorithm for retrieving iceberg concentration will be developed, along with correction methods for enhancing sea ice parameter products, such as sea ice concentration, in areas affected by large icebergs.

5. Conclusions

An RF binary classification model was developed to detect and track the iceberg A23A in Weddell Sea, West Antarctica, using multi-frequency brightness temperature observations from AMSR2. Reference datasets for iceberg and non-iceberg classes were constructed based on coincident Sentinel-1 SAR imagery. The RF model achieved an F1-score of 0.84, indicating relatively strong performance in detecting A23A. Although the RF model exhibited a commission error of approximately 20% for iceberg detection, this did not significantly affect its ability to delineate the spatial distribution of A23A and track its location, given the 10 km grid size of the AMSR2 data and the massive size of the iceberg. This error could potentially be reduced by adjusting the threshold of the iceberg area within a single AMSR2 pixel used in reference dataset construction. Among the input variables, brightness temperatures at 89 GHz in both horizontal and vertical polarizations were identified as the most influential for iceberg detection, likely due to the distinct radiation properties of A23A compared to surrounding sea ice and openwater at this frequency. The RF model successfully tracked the drift of A23A from July 2012 to September 2023. However, after this period, the model failed to detect A23A likely due to the limited diversity of microwave radiation characteristics included in the training data. Incorporating additional training samples from icebergs with varying properties could enhance the performance and generalizability of the model.

With further development, including the construction of more comprehensive and diverse training datasets, the proposed RF model could be extended to reliably detect and track large tabular icebergs across the entire Antarctic region. Such advancements would not only contribute to understanding the long-term evolution of Antarctic icebergs and supporting maritime safety but also significantly enhance the accuracy of sea ice parameter retrievals from passive microwave satellite data, particularly in areaswhere large icebergs affect the microwave radiation signals of sea ice and open water.

Acknowledgments

This study was supported by the National Research Foundation of Korea (No. 2019R1A6A1A03033167 and NRF-2021R1C1C 1009621). The authors would like to thank European Space Agency for providing Sentinel-1 SAR data and Japan Aerospace Exploration Agency for providing AMSR2 data.

Conflict of Interest

No potential conflict of interest relevant to this article was reported.

References

  1. Andersen, S., Tonboe, R., Kaleschke, L., Heygster, G., and Pedersen, L. T., 2007. Intercomparison of passive microwave sea ice concentration retrievals over the high‐concentration Arctic sea ice. Journal of Geophysical Research, 112(C8), C08004. https://doi.org/10.1029/2006JC003543
  2. Arrigo, K. R., van Dijken, G. L., Ainley, D. G., Fahnestock, M. A., and Markus, T., 2002. Ecological impact of a large Antarctic iceberg. Geophysical Research Letters, 29(7), 1104. https://doi.org/10.1029/2001GL014160
  3. Barbat, M. M., Rackow, T.,Wesche,C., Hellmer, H. H., and Mata, M. M., 2021.Automated iceberg trackingwith a machine learning approach applied to SAR imagery: A Weddell Sea case study. ISPRS Journal of Photogrammetry and Remote Sensing, 172, 189–206. https://doi.org/10.1016/j.isprsjprs.2020.12.006
  4. BBC, 2025. World's largest iceberg runs aground off remote island. Available online: https://www.bbc.com/news/articles/c20d1xp6046o (accessed on Mar. 30, 2025).
  5. Bigg, G. R., and Marsh, R., 2023. The history of a cluster of large icebergs on leaving the Weddell Sea pack ice and their impact on the ocean. Antarctic Science, 35(3), 176–193. https://doi.org/10.1017/S0954102022000517
  6. Bliss, A. C., Steele, M., Peng, G., Meier, W. N., and Dickinson, S., 2019.Regional variability of Arctic sea ice seasonal change climate indicators from a passive microwave climate data record. Environmental Research Letters, 14(4), 045003.https://doi.org/10.1088/1748-9326/aafb84
  7. Breiman, L., 2001. Random forests. Machine Learning, 45, 5–32. https://doi.org/10.1023/A:1010933404324
  8. Budge, J. S., and Long, D. G., 2018. A comprehensive database for Antarctic iceberg tracking using scatterometer data. IEEE Journal of Selected Topicsin Applied Earth Observations and Remote Sensing, 11(2), 434–442. https://doi.org/10.1109/JSTARS.2017.2784186
  9. Cavalieri, D.J., Gloersen, P., and Campbell, W.J., 1984. Determination of sea ice parameters with the Nimbus 7 SMMR. Journal of Geophysical Research, 89(D4), 5355–5369. https://doi.org/10.1029/JD089iD04p05355
  10. Chi, J., and Kim, H. C., 2021. Retrieval of daily sea ice thickness from AMSR2 passive microwave data using ensemble convolutional neural networks. GIScience & Remote Sensing, 58(6), 812–830. https://doi.org/10.1080/15481603.2021.1943213
  11. Chi, J., Kim, H. C., Lee, S., and Crawford, M. M., 2019. Deep learning based retrieval algorithm for Arctic sea ice concentration from AMSR2 passive microwave and MODIS optical data. Remote Sensing of Environment, 231, 111204. https://doi.org/10.1016/j.rse.2019.05.023
  12. Dong, Z., Shi, L., Lin, M., and Zeng, T., 2022. A suitable retrieval algorithm of Arctic snow depths with AMSR-2 and its application to sea ice thicknesses of Cryosat-2 data. Remote Sensing, 14(4), 1041. https://doi.org/10.3390/rs14041041
  13. Enderlin, E. M., Moffat, C., Miller, E., Dickson, A., Oliver, C., Dryák-Vallies, M. C., et al., 2023. Antarctic iceberg melt rate variability and sensitivity to ocean thermal forcing. Journal of Glaciology, 69(278), 1718–1728. https://doi.org/10.1017/jog.2023.54
  14. Grosfeld, K., Schröder, M., Fahrbach, E., Gerdes, R., and Mackensen, A., 2001. How iceberg calving and grounding change the circulation and hydrography in the Filchner Ice Shelf-Ocean System. Journal of Geophysical Research, 106(C5), 9039–9055. https://doi.org/10.1029/2000JC000601
  15. Han, H., and Lee, H., 2011. Microwave radiation characteristics of glacial ice in the AMSR-E NASA Team2 algorithm. Korean Journal of Remote Sensing, 27(5), 543–553. https://doi.org/10.7780/kjrs.2011.27.5.543
  16. Han, H., Lee, S., Kim, H.-C., and Kim, M., 2021. Retrieval of summer sea ice concentration in the Pacific Arctic Ocean from AMSR2 observations and numerical weather data using random forest regression. Remote Sensing, 13(12), 2283. https://doi.org/10.3390/rs13122283
  17. Han, H., Lee, S., Kim, J.-I., Kim, S. H., and Kim, H.-C., 2019. Changes in a giant iceberg created from the collapse of the Larsen C Ice Shelf,Antarctic Peninsula, derived from Sentinel-1 and CryoSat-2 data. Remote Sensing, 11(4), 404. https://doi.org/10.3390/rs11040404
  18. Imaoka, K., Kachi, M., Fujii, H., Murakami, H., Hori, M., Ono, A., et al., 2010. Global Change Observation Mission (GCOM)for monitoring carbon,water cycles, and climate change. Proceedings of the IEEE, 98(5), 717–734. https://doi.org/10.1109/JPROC.2009.2036869
  19. Ingels,J., Aronson, R. B., Smith, C. R., Baco, A., Bik, H. M., Blake, J.A., et al., 2021. Antarctic ecosystem responses following ice-shelf collapse and iceberg calving: Science review and future research. WIREs Climate Change, 12(1), e682. https://doi.org/10.1002/wcc.682
  20. Ivanova, N., Pedersen, L. T., Tonboe, R. T.,Kern, S., Heygster, G., Lavergne,T., et al., 2015.Inter-comparison and evaluation of sea ice algorithms: towards further identification of challenges and optimal approach using passive microwave observations. The Cryosphere, 9(5), 1797–1817. https://doi.org/10.5194/tc-9-1797-2015
  21. Ji, Q., Li, F., Pang, X., and Luo, C., 2018. Statistical analysis of SSMIS sea ice concentration threshold at the Arctic sea ice edge during summer based on MODIS and ship-based observational data. Sensors, 18(4), 1109. https://doi.org/10.3390/s18041109
  22. Jiang, L., Shi, J., Tjuatja, S., Dozier, J., Chen, K., and Zhang, L., 2007. A parameterized multiple-scattering model for microwave emission from dry snow. Remote Sensing of Environment, 111(2–3), 357–366. https://doi.org/10.1016/j.rse.2007.02.034
  23. Kern, S., Spreen, G., Kaleschke, L., De La Rosa, S., and Heygster, G., 2007. Polynya signature simulation method polynya area in comparison to AMSR-E 89 GHz sea-ice concentrations in the Ross Sea and off the Adélie Coast, Antarctica, for 2002–05: First results. Annals of Glaciology, 46, 409–418. https://doi.org/10.3189/172756407782871585
  24. Koo, Y., Xie, H., Ackley, S. F., Mestas-Nuñez, A. M., Macdonald, G. J., and Hyun, C. U., 2021. Semi-automated tracking of iceberg B43 using Sentinel-1 SAR images via Google Earth Engine. The Cryosphere, 15(10), 4727–4744. https://doi.org/10.5194/tc-15-4727-2021
  25. Maeda, T., Taniguchi, Y., and Imaoka, K., 2016. GCOM-W1 AMSR2 Level 1R product: Dataset of brightness temperature modified using the antenna pattern matching technique. IEEE Transactions on Geoscience and Remote Sensing, 54(2), 770–782. https://doi.org/10.1109/TGRS.2015.2465170
  26. Markus, T., and Cavalieri, D. J., 2000. An enhancement of the NASA Team sea ice algorithm. IEEE Transactions on Geoscience and Remote Sensing, 38(3), 1387–1398. https://doi.org/10.1109/36.843033
  27. Markus, T., and Dokken, S. T., 2002. Evaluation of late summer passive microwave Arctic sea ice retrievals. IEEE Transactions on Geoscience and Remote Sensing, 40(2), 348–356. https://doi.org/10.1109/36.992795
  28. Meier, W. N., Stewart, J. S., Windnagel, A., and Fetterer, F. M., 2022. Comparison of hemispheric and regional sea ice extent and area trends from NOAA and NASA passive microwave-derived climate records. Remote Sensing, 14(3), 619. https://doi.org/10.3390/rs14030619
  29. Melsheimer, C., Spreen, G., Ye, Y., and Shokr, M., 2023. First results of Antarctic sea ice type retrieval from active and passive microwave remote sensing data. The Cryosphere, 17(1), 105–126. https://doi.org/10.5194/tc-17-105-2023
  30. Merino, N., Le Sommer, J., Durand, G., Jourdain, N. C., Madec, G., Mathiot, P., et al., 2016. Antarctic icebergs melt over the Southern Ocean: Climatology and impact on sea ice. Ocean Modelling, 104, 99–110. https://doi.org/10.1016/j.ocemod.2016.05.001
  31. Parkinson, C. L., and Cavalieri, D. J., 1989. Arctic sea ice 1973–1987: Seasonal, regional, and interannual variability. Journal of Geophysical Research, 94(C10), 14499–14523. https://doi.org/10.1029/JC094iC10p14499
  32. Phillips, H. A., and Laxon, S. W., 1995. Tracking of Antarctic tabular icebergs using passive microwave radiometry. International Journal of Remote Sensing, 16(2), 399–405. https://doi.org/10.1080/01431169508954407
  33. Radhakrishnan, K., Scott, K. A., and Clausi, D. A., 2021. Sea ice concentration estimation: Using passive microwave and SAR data with a U-Net and curriculum learning. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14, 5339–5351. https://doi.org/10.1109/JSTARS.2021.3076109
  34. Schloesser, F., Friedrich, T., Timmermann, A., DeConto, R. M., and Pollard, D., 2019.Antarctic iceberg impacts on future Southern Hemisphere climate. Nature Climate Change, 9(9), 672–677. https://doi.org/10.1038/s41558-019-0546-1
  35. Schodlok, M. P., Hellmer, H. H., Rohardt, G., and Fahrbach, E., 2006.Weddell Sea iceberg drift: Five years of observations. Journal of Geophysical Research, 111(C6),C06018. https://doi.org/10.1029/2004JC002661
  36. Singh, S., Tiwari, R. K., Sood, V., Kaur, R., and Prashar, S., 2022. The legacy of scatterometers: Review of applications and perspective. IEEE Geoscience and Remote Sensing Magazine, 10(2), 39–65. https://doi.org/10.1109/MGRS.2022.3145500
  37. Spreen, G., Kaleschke, L., and Heygster, G., 2008. Sea ice remote sensing using AMSR‐E 89‐GHz channels. Journal of Geophysical Research, 113(C2), C02S03. https://doi.org/10.1029/2005JC003384
  38. Stuart,K.M., and Long,D.G., 2011.Tracking large tabular icebergs using the SeaWinds Ku-band microwave scatterometer. Deep Sea Research Part II: Topical Studies in Oceanography, 58(11–12), 1285–1300. https://doi.org/10.1016/j.dsr2.2010.11.004
  39. Stulic, L., Timmermann, R., Paul, S., Zentek, R., Heinemann, G., and Kanzow, T., 2023. Southern Weddell Sea surface freshwater flux modulated by icescape and atmospheric forcing. Ocean Science, 19(6), 1791–1808. https://doi.org/10.5194/os-19-1791-2023
  40. Torres, R., Snoeij, P., Geudtner, D., Bibby, D., Davidson, M., Attema, E., et al., 2012. GMES Sentinel-1 mission. Remote Sensing of Environment, 120, 9–24. https://doi.org/10.1016/j.rse.2011.05.028
  41. Wesche, C., and Dierking, W., 2012. Iceberg signatures and detection in SAR images in two test regions of the Weddell Sea, Antarctica. Journal of Glaciology, 58(208), 325–339. https://doi.org/10.3189/2012J0G11J020
  42. Worby, A. P., and Comiso, J.C., 2004. Studies of the Antarctic sea ice edge and ice extent from satellite and ship observations. Remote Sensing of Environment, 92(1), 98–111. https://doi.org/10.1016/j.rse.2004.05.007
  43. Wu, S.-Y., and Hou, S., 2017. Impact of icebergs on net primary productivity in the Southern Ocean. The Cryosphere, 11, 707–722. https://doi.org/10.5194/tc-11-707-2017