Next Article in Journal
Comparative Analysis of Carbon Density Simulation Methods in Grassland Ecosystems: A Case Study from Gansu Province, China
Next Article in Special Issue
Evaluation of Sea Ice Motion Estimates from Enhanced Resolution Passive Microwave Brightness Temperatures
Previous Article in Journal
Assessing the Impact of Land Use and Land Cover Change on Environmental Parameters in Khyber Pakhtunkhwa, Pakistan: A Comprehensive Study and Future Projections
Previous Article in Special Issue
Seasonal and Interannual Variations in Sea Ice Thickness in the Weddell Sea, Antarctica (2019–2022) Using ICESat-2
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Remote Sensing
Volume 17
Issue 1
10.3390/rs17010171
remotesensing-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Validation of an AMSR2 Thin-Ice Thickness Algorithm for Global Sea-Ice-Covered Oceans Using Satellite and In Situ Observations

by
Kazuki Nakata
1,*,
Misako Kachi
1,
Rigen Shimada
1,
Eri Yoshizawa
1,
Masato Ito
2 and
Kay I. Ohshima
3,4
1
Earth Observation Research Center, Japan Aerospace Exploration Agency, Tsukuba 305-8505, Japan
2
National Institute of Polar Research, Tachikawa 190-8518, Japan
3
Institute of Low Temperature Science, Hokkaido University, Sapporo 060-0819, Japan
4
Arctic Research Center, Hokkaido University, Sapporo 001-0021, Japan
*
Author to whom correspondence should be addressed.
Remote Sens. 2025, 17(1), 171; https://doi.org/10.3390/rs17010171
Submission received: 9 November 2024 / Revised: 23 December 2024 / Accepted: 27 December 2024 / Published: 6 January 2025
(This article belongs to the Special Issue Monitoring Sea Ice Loss with Remote Sensing Techniques)
Browse Figures
Versions Notes

Abstract

:
The detection of thin-ice thickness using satellite microwave radiometers is a strong tool for estimating sea-ice production in coastal polynyas, which leads to dense water formation driving ocean thermohaline circulation. Thin-ice areas are classified into two ice types: active frazil, comprising frazil ice and open water, and thin solid ice, areas of relatively uniform thin ice. A thin-ice algorithm for AMSR-E has been developed to classify these two ice types and estimate ice thickness of <20 cm. In this study, we validate the applicability of the algorithm to the successor, AMSR2, using validation data of ice types identified from Sentinel-1 and ice thickness derived from MODIS. The validation results show an ice-type misclassification rate of ~3% and mean absolute errors in ice thickness of 2.0 cm and 5.0 cm for active frazil and thin solid ice, respectively. These values are similar to those for AMSR-E, indicating that the thin-ice algorithm can be applied to AMSR2. Further validations with the moored ADCP backscattering data capturing underwater frazil ice signals demonstrate that the algorithm can accurately distinguish between two ice types and effectively detect deep-penetrating frazil ice. The AMSR2 thin-ice thickness data has been released as a JAXA research product.

1. Introduction

Information on thin-ice thickness is essential for understanding the dynamics and thermodynamics of coastal polynyas, which are regions of thin ice. Coastal polynyas are formed by sea-ice divergence due to prevailing offshore winds during winter [1]. A large heat loss to the atmosphere occurs in coastal polynyas because the heat-insulating effect of thin ice is much smaller. In shallow coastal polynya areas where the whole water column can reach the freezing point, a large part of this heat loss contributes to sea-ice production. Dense shelf water formed by brine rejection during sea-ice formation plays a key role in the formation of Antarctic Bottom Water (e.g., [2]), North Pacific Intermediate Water [3], and the Cold Halocline Layer in the Arctic Ocean [4,5]. Therefore, the detection of thin-ice regions and the estimation of sea-ice production are crucial for understanding global thermohaline circulation and the associated biogeochemical cycles [6].
Retrieving thin-ice thickness from satellite data enables the estimation of global sea-ice production through the heat budget calculation. Thermal infrared sensors, such as the Moderate Resolution Imaging Spectroradiometer (MODIS), have sometimes been used to retrieve thin-ice thickness (e.g., [7]). Since typical coastal polynyas have a width of a few to one hundred kilometers, thermal infrared sensors with a resolution of about one kilometer can resolve even small coastal polynyas. However, coastal polynyas appear and disappear within a few days due to fluctuations in atmospheric forcing. Such temporal variability cannot be captured sufficiently by thermal infrared sensors, as their retrieval capability is heavily limited by cloud cover. On the other hand, satellite microwave radiometers are suitable for resolving temporal variability and obtaining seamless data because these radiometers can observe sea-ice areas globally every day, regardless of daytime or nighttime and cloud cover. Although microwave radiometers with a resolution of around 10 kilometers cannot detect narrow polynyas, they are capable of capturing relatively large polynyas, which hold significant importance for climate and ocean research.
When sea ice grows, surface salinity decreases due to the brine rejection [8]. The decreased surface salinity results in an increase in the surface emissivity of sea ice, leading to an increase in brightness temperature and depolarization [9]. Therefore, ice thickness has a negative correlation with the polarization ratio (PR) of the horizontal and vertical brightness temperatures (TBh and TBv), defined as PR = (TBv − TBh)/(TBv + TBh). In situ microwave observations have confirmed that PR is highly sensitive to changes in ice thickness of ~20 cm (e.g., [10]).
Based on these microwave characteristics, several studies have developed algorithms to derive thin-ice thickness from satellite microwave sensors such as the Special Sensor Microwave/Imager (SSM/I), the Advanced Microwave Scanning Radiometer for the Earth Observing System (AMSR-E), and the Advanced Microwave Scanning Radiometer 2 (AMSR2) (e.g., [11,12,13,14,15,16,17,18,19]). In these thin-ice algorithms, an ice thickness of <~20 cm is inferred by using an empirical equation based on a direct comparison between the PR and ice thickness. The ice thickness for the comparison is estimated using a heat flux calculation, using the ice-surface temperatures from satellite thermal infrared sensors. This ice thickness is referred to as thermal ice thickness and is defined as the thickness at which the calculated total heat flux would be realized if the ice thickness were uniform in the satellite footprint. Thermal ice thickness is a useful quantity for the estimation of sea-ice production because the heat flux or sea-ice production can be directly estimated from the thermal ice thickness in the footprint without information on detailed ice-thickness distribution [12].
The relationship between PR and thermal ice thickness differs depending on the oceans. Therefore, the PR–thickness relationship for AMSR-E has been derived independently for each ice-covered ocean: the Arctic Ocean by Iwamoto et al. [13], the Chukchi Sea by Martin et al. [14,15], the Sea of Okhotsk by Nihashi et al. [16], the Sea of Japan by Nihashi et al. [17], and the Southern Ocean by Nihashi and Ohshima [18]. The PR–ice thickness relationship for AMSR2 has also been obtained for the Southern Ocean by Nihashi et al. [19,20]. Moreover, mapping of sea-ice production has been conducted for each ocean from a heat flux calculation using the ice thickness derived from the microwave radiometer data (e.g., [15,16,17,18,19,20,21,22,23,24]).
On the other hand, a polynya or thin-ice area includes various types of thin ice, specifically frazil/grease ice, nilas, pancake ice, and thin-level ice. At the resolution scale of satellite microwave radiometers, which ranges from several to a few tens of kilometers, thin-ice regions can be roughly classified into two ice types. One is active frazil, a mixture of frazil ice, grease ice, pancake ice, and open water formed under turbulent conditions, and the other is thin solid ice, a relatively uniform thin-ice region such as nilas formed under relatively calm conditions. Nakata et al. [25] examined the PR–thickness relationship for active frazil and thin solid ice separately and found that the relationship is quite different between these two ice types, causing a difference in ice thickness of up to ~15 cm for the same PR. Based on these findings, Nakata et al. [25] developed an algorithm to obtain thin-ice thickness for active frazil and thin solid ice separately after classification of the ice types. This algorithm could be applied to global ice-covered oceans because the difference in the PR–thickness relationship among oceans is mainly caused by the difference in the dominant ice type due to regional differences in atmospheric and geographic conditions. Nakata et al. [26] and Nakata and Ohshima [27] have applied this algorithm to map sea-ice production in both hemispheres. These mappings have demonstrated that the use of this thin-ice algorithm results in up to 50% higher sea-ice production in frazil-dominant polynyas than previous estimates such as those by Nihashi and Ohshima [18], highlighting the importance of taking account of sea-ice type in the algorithm.
The algorithm by Nakata et al. [25] was designed for AMSR-E, whose operational period was from June 2002 to October 2011. AMSR2 succeeded AMSR-E from July 2012 to the present day. The applicability of the AMSR-E algorithm proposed by Nakata et al. [25] to AMSR2 has not been validated. If the algorithm can also be applied to AMSR2, we will be able to obtain more than 20-year sea-ice products derived from the algorithm, which will be very useful for climate-change studies, particularly those related to the recent decline in Arctic and Antarctic sea ice.
To validate the thin-ice thickness algorithm proposed by Nakata et al. [25], it is essential to examine both the sea-ice-type classification and thin-ice thickness estimation. In synthetic aperture radar (SAR) images, active frazil ice is represented by areas with high backscatter streaks, whereas solid ice is represented by relatively low-backscatter areas without such streaks. Therefore, SAR images from Copernicus Sentinel-1 observed during the AMSR2 period can be used to obtain ground truth data for sea-ice-type classification. To validate the PR–ice thickness relationship, the thermal ice thickness derived from MODIS data can be used as the ground truth.
Furthermore, an acoustic Doppler current profiler (ADCP) moored in the Cape Darnley polynya successfully detected signals of active frazil ice, which extends to depths of several tens of meters, through volume backscattering strength (SV) data [28]. The use of such ADCP SV data enables the first in situ validation of the AMSR series (hereafter referred to as AMSR) algorithm for active frazil detection. Although these data were obtained during 2010, the AMSR-E period, if the AMSR-E ice-type classification can be applied to AMSR2, the validation of the AMSR-E algorithm would result in the validation of the AMSR2 algorithm.
In this study, we validate whether the thin-ice thickness algorithm developed for AMSR-E can be applied to AMSR2 using Sentinel-1 SAR and MODIS images. In addition, we further validate the degree to which the ice-type classification accurately represents active frazil by utilizing the ADCP SV data capturing the presence of underwater frazil ice. Through such robust validations, we have developed a reliable AMSR2 algorithm to create the thin-ice thickness product, which will be a useful dataset for long-term monitoring and investigation of global coastal polynyas when combined with the AMSR-E thin-ice thickness product. This research is part of the validation work for the GCOM-W/AMSR2 Thin-Ice Thickness (Thermal Ice Thickness) product, which has been released as a new sea-ice product by the Japan Aerospace Exploration Agency (JAXA).

2. Data and Method

JAXA provides a GCOM-W/AMSR2 Level-1R (L1R) TB product that matches the center position and size of the footprint among different frequency channels using the Backus–Gilbert method [29]. Measured TBs can be expressed as the convolution of the antenna pattern with the radiation components. The Backus–Gilbert method is an optimal interpolation technique used to synthesize TBs for a desired antenna pattern from measured TBs. In this product, the TBs at high-frequency channels are adjusted so that they are observed with the antenna patterns of the lower-frequency channels. Although these modified TBs have a lower spatial resolution than the original TBs, they are suitable for operations between TBs at different frequency channels. Because the thin-ice algorithm requires the calculation of the gradient ratio (GR) from 36 to 89 GHz vertically polarized TBs, defined as GR = (TB89v – TB36v)/(TB89v + TB36v), this Level 1R TBs product was used in this study.
The direct use of TBs is not appropriate for comparison between AMSR-E and AMSR2 because of sensor biases. To reduce such biases, a comparison of the TBs between AMSR2 and AMSR-E was conducted by the “Intercomparison results between AMSR2 and TMI/AMSRE/GMI (AMSR2 ver. 2.0) EORC JAXA”. This intercomparison was conducted by slowly rotating the antenna after the end of the AMSR-E mission. Following the approach of Ohshima et al. [24], we used conversion equations to adjust the AMSR2 TBs to be consistent with the AMSR-E TBs. This adjustment reduces the difference between AMSR2 and AMSR-E TBs over the ocean to ~3.3 K at 36 GHz and ~1.7 K at 89 GHz.
The thin-ice algorithm of Nakata et al. [25] masks areas with sea-ice concentrations of <30% as open water, where the ice concentration from the enhanced NASA Team (NT2) algorithm [30] is used. In this study, to maintain consistency with other JAXA sea-ice products, we used the sea-ice concentration from the JAXA AMSR2 standard product, derived from the bootstrap algorithm [31,32,33]. This sea-ice concentration sometimes shows a lower concentration of ~10% or more in coastal polynyas than that of the NT2 algorithm. Therefore, we adopted a lower threshold of 15% for open-water masks, which reduces the misclassification of polynya areas as open water.

2.1. Ice Types Derived from Sentinel-1 SAR

To demonstrate the performance of the AMSR algorithm for ice-type classification, Ohshima et al. [28] used Sentinel-1 SAR images in cases of dominant-frazil ice for four Antarctic polynyas in the Ross Sea and around East Antarctica. In this study, we also use the same SAR images to validate the ice-type classification more quantitatively. In addition to the data for the Southern Ocean, we newly used four SAR images obtained in the Northern Hemisphere for the validation. The locations and details of the Sentinel-1 images used in this study are shown in Table 1 and Figure 1. These Sentinel-1 images were searched using Google Earth Engine, targeting regions where large active frazil areas frequently appear. Only the images that allowed for the objective inspection of active frazil were then used.
From the SAR images, we visually classify the areas of active frazil and solid ice, which are used as validation data for the AMSR2 algorithm. The discrimination of thin and thick solid ice from SAR images is difficult and was not performed in this study. As shown in the Sentinel-1 image obtained in the Shackleton polynya (Figure 1), active frazil appears in a SAR image as areas with high backscatter streaks. Therefore, ice types can be discriminated from SAR images by visually interpreting the presence or absence of streaks. It should be noted that the observational timings of the Sentinel-1 images used in this study differ from those of AMSR2 by up to five hours. In the case of active frazil, which appears under strong wind conditions, the boundary of the ice types can be shifted by the strong wind during the difference in the observational timings. Previous studies indicated that active frazil areas typically expand offshore at a rate of ~2 km/h (e.g., [34]). In this study, assuming that sea ice moves offshore at a rate of 3 km/h, areas within the distance of ice moving from the ice-type boundary were excluded from the validation data. We used a somewhat higher rate to minimize the amount of incorrect data for the ground truth. An example of ice-type classification through visual inspection of the Shackleton polynya is shown in Figure 1. Because Sentinel-1 observed the area two hours earlier than AMSR2 did in this case, areas within a distance of 6 km offshore from the boundary were excluded from the validation data.

2.2. Thin-Ice Thickness Derived from MODIS

The Aqua/MODIS MYD29 sea-ice product contains the sea-ice surface temperature (Ts), which can be used to calculate the thermal ice thickness [35]. It has been reported that the MODIS cloud mask frequently misclassifies thin-ice areas as cloud cover [36,37]. Hence, we used unmasked Ts by applying the method of Key et al. [38] to channels 32 and 33 of the MYD021KM product. We focused on six regions with high occurrence of thin ice (offshore of Cape Darnley, Vincennes Bay, the Ross Sea, the Sea of Okhotsk, the Bering/Chukchi Sea, and North of Baffin Bay), selecting approximately 30–40 scenes with clear-sky conditions for each region during September–May 2013–2015 in the Northern Hemisphere and April–October 2013–2015 in the Southern Hemisphere. The total number of scenes used in this study is 226. The difference in observation time between GCOM-W/AMSR2 and Aqua/MODIS can be neglected because these satellites have been flying in the Afternoon Constellation (A-train) to synchronize their observations. After deriving the Ts from these scenes, we smoothed the MODIS Ts data to match the resolution of AMSR2 using a weighted average with a Gaussian function approximating the AMSR2 36 GHz antenna pattern.
To create high-quality validation data for thin-ice thickness, the following masking processes were conducted. Following Drucker et al. [39], we masked pixels with a solar zenith angle of >75°. Additionally, to avoid resolution degradation and atmospheric effects caused by large sensor scan angles, we only used pixels with a MODIS sensor scan angle of <40°, following Mäkynen and Similä [40]. Finally, we visually judged the presence of cloud cover and used only cloud-free pixels for the validation.
Next, we describe the calculation method for thin-ice thickness from MODIS data. Because we only used pixels with a small solar zenith angle, the shortwave radiation can be neglected. Then, the net heat flux QNET is expressed as the sum of sensible and latent heat fluxes and the net longwave radiation at the ice surface. The net longwave radiation and turbulent heat fluxes are functions of the surface temperature of the Ts. Since thermal inertia can be neglected for thin ice under typical meteorological conditions during the winter, the QNET is balanced by the conductive heat flux through the ice [41]. Thus, the sea-ice thickness (hj) can be calculated using the following equation:
Q N E T = k i ( T B − T S ) h i ,
where ki represents the thermal conductivity of sea ice (2.03 W m−1 K−1), and TB denotes the ice-bottom temperature, assumed to be the freezing point of seawater (=−1.86 °C; no oceanic heat flux below). In this study, we used the bulk formulas of Kondo [42] for the sensible and latent heat-flux calculations. To calculate net longwave radiation, we used the empirical formula of Guest [43]. As atmospheric input data, we used the air temperature at 2 m, dew point temperature at 2 m, wind speed at 10 m, and surface sea level pressure from the 1-h ECMWF (European Centre for Medium-Range Weather Forecasts) Reanalysis v5 (ERA5) product with a spatial resolution of 0.25° × 0.25° [44,45]. We used ERA5 data at the closest time to the MODIS observations.
MODIS ice thicknesses of <30 cm and >30 cm were used to assess the accuracy of thin-ice thickness and thin-ice detection from AMSR2, respectively. Figure 2 shows the frequency map of all MODIS data points used in this study and those with ice thicknesses of <30 cm, indicating that the validation data were comprehensively collected in regions where coastal polynyas frequently occur, i.e., the Sea of Okhotsk, Bering Sea, Chukchi Sea, Ross Sea, and East Antarctica, as well as some marginal ice zones. This dataset is sufficient for validation, with the total number of pixels of sea ice and thin sea ice being ~260,000 and ~40,000, respectively.

2.3. ADCP Data

The ADCP observations were conducted off Cape Darnley, Antarctica (the location is indicated by the red cross in the upper panel of Figure 1), from February 2010 to February 2011 [28]. The ADCP (Workhorse Sentinel 300 kHz, Teledyne RD Instruments, Poway, CA, USA) measured the acoustic backscatter strength at 15 min intervals with a cell length of 8 m from the surface to a depth of 80 m. The acoustic frequency of the emitted ping is 307 kHz. The dynamic range is ±80 dB, and the echo intensity accuracy is ±1.5 dB. To correct for the influence of the distance from the transducer to the scatterer, battery power, and other factors, the echo intensity received by the ADCP transducer was converted to the target strength per unit volume in dB, called the volume backscatter strength, SV, following Deines [46] and Ito et al. [47]. The detailed processing has been described by Ohshima et al. [28]. We used the hourly averages of the ADCP SV for comparison with the AMSR2 data in Section 4.

3. Validation of the Thin-Ice Thickness Algorithm Using Satellite Data

3.1. Ice-Type Classification

We validate whether the ice-type classification developed by Nakata et al. [25] can be applied to AMSR2 using the Sentinel-1 SAR images. Nakata et al. [25] employed a linear discriminant function defined as G(x) = wTx for classifying into active frazil and solid ice, where x is the observation vector (PR, GR, 1) consisting of the PR at 36 GHz and the GR of TBv at 36 GHz and 89 GHz, and w is the weighted coefficient vector. We used a value of w = (−193, 1002, −0.7), which ensures that the posterior probabilities P(active frazil|x) and P(thin solid ice|x) are equal when G(x) = 0. Assuming that active frazil has a PR of >0.05, active frazil and solid ice are discriminated using the following equation:
I c e   t y p e = a c t i v e   f r a z i l               i f         G x > 0           a n d           P R > 0.05 s o l i d   i c e                           i f             G x < 0           o r         P R < 0.05
Nakata et al. [25] also introduced a PR threshold to further subdivide solid ice into thin solid ice and thick solid ice. This validation focuses solely on classification into active frazil and solid ice without applying that threshold.
Figure 3a shows the scatter plot of each ice type on the PR–GR plane obtained by comparison with the SAR images and AMSR2. According to a similar analysis for AMSR-E by Nakata et al. [25], plots of active frazil show a quasi-linear relationship on the PR–GR plane because they are mainly influenced by the open water fraction. In contrast, plots of solid ice are more sensitive to the real ice thickness. As the sea ice thickens, the GR of the solid ice decreases significantly. Figure 3a confirms that such characteristics also hold for AMSR2 data. The background colors in Figure 3a indicate the classification range for each ice type by the AMSR-E algorithm of Nakata et al. [25], which well explains the plots of the ice types obtained from Sentinel-1 SAR and AMSR2. The classification accuracy, precision, and recall were 0.98, 0.72, and 0.98, respectively. The lower precision relative to the recall indicates an over-detection of active frazil (see the caption of Table 2 for more details). However, these metrics can be influenced by the sample size for each ice type. To provide a more objective assessment of the performance of the classification method, it is preferable to use metrics that assume equal weighting for the sample size of each ice type. When this assumption was applied, all metrics were calculated to be 0.98, which aligns closely with the error of 3% estimated for AMSR-E. When the evaluation was made separately for the Northern and Southern Hemispheres, the metrics were 0.97 and 0.98, respectively, indicating no significant bias in the performance of ice-type classification between the two hemispheres. In conclusion, the ice-type classification method of AMSR-E is applicable to the AMSR2 data.

3.2. Estimation of Thin-Ice Thickness

Next, we validate the thin-ice thickness estimation method of Nakata et al. [25] by using ice thickness data obtained from MODIS. In the method of Nakata et al. [25], thin-ice thicknesses up to 20 cm are empirically calculated using an exponential relationship between PR and ice thickness for each ice type. The exponential equation is given by
h = e 1 a P R + b − c ,
where a, b, and c represent the regression coefficients obtained by the least squares fitting of the PR–thickness plots. In Nakata et al. [25], the coefficients for AMSR-E were estimated to be (a, b, c) = (596, −11.8, 1.008) for active frazil and (72, 0, 1.06) for solid ice, using the least squares fitting to minimize the distance between the regression curve of (3) and the data points.
Figure 3b shows the scatter plots of the AMSR2 PR and MODIS ice thicknesses for active frazil and solid ice collected in this study. The regression curve for each ice type obtained by the least squares fitting of the PR–thickness plots is shown by the solid lines. The regression curves of AMSR-E by Nakata et al. [25] are also indicated by dashed lines. In this study, we newly derived the regression curves for AMSR2 using all MODIS thin-ice thickness data as reference. Nakata et al. [25] derived the PR–thickness relationship for each ice type using the matchup data of MODIS ice thickness, SAR, and AMSR-E PR that were observed at nearly the same time. They used data only when the ice type could be clearly identified from the SAR images through visual inspection. However, such a comparison cannot be made in this study because we did not have matchup data for MODIS ice thickness and SAR. In the plots of PR versus ice thickness (Figure 3b), the ice type is derived from the AMSR ice-type classification using Equation (2) (Figure 3a) and not from the SAR images. Therefore, some of the plots in Figure 3b include misclassifications of the ice type. Although this approach differs from that of Nakata et al. [25], it is reasonable because the thin-ice thickness estimate depends on the ice-type classification, and fitting (determination of the regression coefficients) should be performed for plots that include misclassifications of ice-type.
Although the plots of active frazil in Figure 3b mostly show ice thicknesses of <5 cm, some plots show a thickness of >20 cm. These outliers could be due to misclassification of the ice type, mixed pixels containing thicker ice, or thin clouds that cannot be excluded through visual inspection. If the contamination of cloud signals is significant, it may affect the estimation of the regression coefficients. To minimize such contamination, we first removed data where the ice thickness was more than 2σ from the mean within each PR interval of 0.02. We then adopted a robust fitting method that minimizes the absolute errors to estimate the regression coefficients. The fitting curve was obtained by assuming that AMSR2 ice thickness is 0 cm at PR = 0.23 to prevent a negative ice thickness. For the fitting in the case of thin solid ice, the normal from each datapoint to the curve was calculated for all data with a thickness of <30 cm, and then a least squares fitting was applied to minimize the distance.
The regression coefficients of AMSR2 for active frazil and thin solid ice were estimated to be (a, b, c) = (353, −5.7, 1.013) and (70, −0.3, 1.093), respectively, which show some deviation from those of AMSR-E by Nakata et al. [25], as shown in Figure 3b. We compared the MODIS thin-ice thickness with the AMSR2 thin-ice thickness obtained from these regression curves, showing good correspondence between the two (Figure 4c). Table 2 summarizes the validation results using an AMSR2 ice thickness of <20 cm and MODIS ice thickness of <30 cm. As illustrated in Figure 4, the error distribution of the ice thickness varies with the thickness and does not follow a normal distribution. We then used the mean absolute error (MAE), which is less sensitive to outliers and is a more appropriate error metric than the root mean squared error (RMSE). The MAEs were calculated to be less than 3 cm for active frazil and less than 10 cm for thin solid ice, indicating that this AMSR2 thin-ice algorithm meets the target accuracy of the JAXA research product. The RMSE is calculated to be 6.5 cm, which is comparable to the value reported for AMSR-E (~6 cm) by Nakata et al. [25]. This validation result is derived from the same reference data used for calculating the regression coefficients, and it does not represent an evaluation with independent data. To confirm the validity of these validation results, we conducted a five-fold cross-validation, a method for evaluating the performance of a regression model that involves splitting the dataset into five folds. The results showed that for thin solid ice, the RMSE was consistently stable at approximately 6.6 cm across the folds, indicating a high reliability of the validation results. However, for active frazil, the RMSE had some variability, ranging from 3.4 cm to 4.4 cm, likely due to the limited number of data for this ice type.
Nihashi et al. [20] also developed an AMSR2 thin-ice thickness algorithm specifically for the Southern Ocean. This algorithm estimates thin-ice thicknesses of <10 cm and 10–20 cm from PR at 89 GHz and 36 GHz, respectively. When our validation data for the Southern Ocean were used for validation of their algorithm, the RMSE, MAE, and bias were calculated to be 0.76 cm, 6.1 cm, and 4.6 cm, respectively. These values are comparable to those of our algorithm.
In addition to the ice thickness error, we here evaluate how accurately thin ice with a thickness of <20 cm can be detected. The validation results in Table 2 reveal that the false-detection rate of thin ice was approximately 42%, whereas the missed-detection rate was approximately 28% for thin-ice detection (“Unbalanced” in the lower section of Table 2). Considering the sample sizes of thin and thick ice, the false-detection rate was calculated as approximately 6%, with an accuracy of 84% (“Balanced” in the lower section of Table 2). There is little difference in all accuracy metrics between the Northern and Southern hemispheres.

3.3. Application of the Algorithm

Figure 5 shows the application results of our thin-ice algorithm for the case in which large thin-ice areas were formed in Vincennes Bay, Antarctica, on 9 August 2016. On this day, MODIS captured both thermal infrared and visible images under a clear sky with very little solar radiation (Figure 5a,b). Additionally, Sentinel-1 also acquired data (Figure 5c), making this case ideal for a detailed intercomparison of the ice type and ice thickness from SAR, visible/infrared, and AMSR2 data. The MODIS visible image (Figure 5a) and Sentinel-1 SAR image (Figure 5c) show a large area with streaks of frazil/grease and pancake ice from the coast, indicative of an active frazil area in the coastal polynya. Adjacent to this area, the SAR image shows a low-backscatter area without streaks, which appears as a low-reflectivity area in the visible image. These characteristics suggest that this area consists of relatively uniform, newly formed solid ice with low surface roughness and no snow cover. The thermal ice thickness derived from MODIS (Figure 5b,d) further supports this interpretation, indicating that, except for the lower right area contaminated by thin clouds, the ice thickness in this area is less than 20 cm and is classified as thin solid ice.
The active frazil and thin solid ice identified by AMSR2 coincide with those identified by MODIS and Sentinel-1 SAR (Figure 5e), demonstrating that the AMSR2 algorithm is accurate for detecting thin-ice areas and distinguishing between active frazil and thin solid ice. Moreover, a comparison with the AMSR2 thin-ice thickness estimates (Figure 5f) demonstrates that the algorithm accurately represents the actual thermal ice thickness (Figure 5d), particularly for active frazil areas. In the region without cloud cover shown in Figure 5, the bias, MAE, and RMSE were calculated to be −0.4 cm, 1.3 cm, and 1.8 cm for active frazil; −1.1 cm, 1.6 cm, and 2.7 cm for thin solid ice; and −0.6 cm, 1.4 cm, and 2.2 cm for total thin ice, respectively. The accuracy, precision, and recall for thin-ice detection were 0.87, 0.83, and 0.98, respectively. The mean ice thickness was 2.5 cm for active frazil areas and 8.6 cm for thin solid-ice areas, indicating that the algorithm effectively captures the difference in ice thickness between ice types, with active frazil being several centimeters thinner than thin solid ice.
Additionally, we present the application results for a case in which active frazil was largely present in Anadyr Bay, the Bering Sea, on 9 August 2022 (Figure 6). All three sensors (Sentinel-1, MODIS, and AMSR2) observed the area under a nearly clear sky, although their observational timings did not coincide, and the MODIS image has a coarser resolution owing to its large scan angles. It is noted that, in the lower part of the images, the sea-ice region was concealed by cloud cover, which can be identified by the high reflectivity in the MODIS visible image (Figure 6a). Both the MODIS visible and SAR images (Figure 6a,c) suggest that a large active frazil area with streaks of high reflectivity and backscatter extended to ~50 km from the northern coast in Anadyr Bay. Ice type and thin-ice thickness distributions from AMSR2 (Figure 6e,f) well represent the spatial features of the active frazil regions, although there is an 8 h difference in observational timing between MODIS and AMSR2, making pixel-by-pixel comparison difficult. Thin solid-ice distribution from AMSR2 can also capture the MODIS thin-ice areas with thicknesses of ~10–20 cm.

4. Validation of Active Frazil Using In Situ Observation

Under the Japanese Antarctic Research Expedition, moored ADCP observations were conducted throughout 2010 off Cape Darnley (see Figure 1 for the location) to clarify the processes of high ice production and subsequent dense water formation in the Cape Darnley polynya, where the precursor of Antarctic Bottom Water is generated [2,21]. An ADCP can observe acoustic scatterers in the water column, specifically marine organisms such as zooplankton, resuspended sediment, and underwater frazil ice. Regarding the ADCP SV data acquired near the coast off Cape Darnley, most of the observed backscatter signals are regarded as those from underwater frazil ice, except for slight signals of diurnal vertical migration of zooplankton [28]. As demonstrated by Ohshima et al. [28], these ADCP SV data provide a valuable validation of the AMSR algorithm for detecting active frazil areas, although precise validation from such data has not been conducted so far. As confirmed in Section 3.1, the AMSR-E sea-ice classification method developed by Nakata et al. [25] can be applied to the AMSR2 data with almost the same accuracy, implying that the validation of the AMSR-E ice-type algorithm from the ADCP SV data during 2010 results in validation of the ice-type classification by AMSR2. Thus, in this study, we conducted further validation by comparing active frazil detection from AMSR-E L2A data at the ADCP location with hourly averaged ADCP SV data.
The time series of the vertical profile of the ADCP SV data (Figure 7b) indicates that high SV signals that extend to depths of several tens of meters appear occasionally, with a typical duration of 1–3 days. These signals are interpreted as underwater frazil ice formed by strong turbulence and cooling by winds [28]. From April to October, the AMSR ice-type classification algorithm captured these underwater frazil ice events with a high SV as active frazil (Figure 7a). However, during the early freezing period of March, missed detections occurred, probably because of the mask-out of areas with sea-ice concentrations of <15%. During the melting period starting from November, both missed and false detections sometimes occurred. According to these comparisons, the AMSR ice-type algorithm is likely to be applied to most of the winter season, while its effectiveness would be limited during early freezing and ice-melting periods.
Assuming that the AMSR ice-type classification can be applied to the period of April–October, we validated the accuracy of the ice-type classification algorithm quantitatively by utilizing the ADCP SV data at the uppermost bin (approximately 0–5 m depths from the surface). We first created a histogram of the ADCP SV in the uppermost bin for active frazil (indicated by orange) and thin solid ice (blue) from the AMSR algorithm (Figure 8a). As expected, the histogram clearly demonstrates that active frazil and solid ice show distinctly different ranges of SV with peaks of −52 dB and −81 dB, respectively.
For validation, we used simple thresholding of the SV at the uppermost bin to obtain the time series of the ice types from the SV and compared the results with those obtained by the AMSR algorithm. To determine the optimal threshold, we supplementarily used eight Envisat Advanced SAR (ASAR; C-band) images acquired by the European Space Agency’s Environmental Satellite in July 2010, which were also used by Nakata et al. [25]. First, we visually identified the predominant ice type at the ADCP location in each SAR image. Figure 7d,e show the enlarged time series of the vertical profile of the ADCP SV and surface SV at the uppermost bin in July, respectively, superimposed by the ice types (dashed lines) visually classified from the SAR images. Among the eight SAR images, two show the active frazil at the ADCP location, with SV values of −56.0 dB and −53.5 dB, and four show the solid ice with a maximum SV value of −69.2 dB. Therefore, the optimal threshold is considered to lie between −69.2 dB and −56.0 dB. Figure 8b illustrates the recall, precision, and accuracy for the AMSR active frazil detection, assuming that the threshold SV value of active frazil is set to the range of −100 dB to −30 dB with an interval of 1 dB. This analysis demonstrates that the highest accuracy lies within the range of −69.2 dB to −56.0 dB, which is consistent with the detection by the SAR images. According to the SAR images on July 9 and July 19, the ADCP station was located at the boundary between active frazil and solid ice, and the SV shows −63.4 dB and −64.3 dB, respectively. If we use the corresponding value of −64 dB as the optimal threshold value, the algorithm achieves an overall accuracy of 80%, although active frazil is over-detected with a recall of 0.85 and a precision of 0.72. A similar accuracy assessment was conducted using the SV at the second bin (5–13 m depth) of the ADCP. A similar optimal threshold value was obtained with the accuracy decreased to 0.7.
The penetration depth of frazil ice is considered to reflect the activity or ice production in the coastal polynyas. Thus, in this study, we attempted to estimate the penetration depth of frazil ice from the SV profiles of ADCP and related it to AMSR active frazil detection. First, we assumed that the histogram of the ADCP SV from all data from April to October consists of two classes with a normal distribution: a class of frazil ice with a high SV and a class of no frazil ice with a low SV. We then estimated the mean and variance for each class and its mixing ratio using the expectation–maximization (EM) algorithm, which refines the model parameters by maximizing the complete data likelihood [48,49]. Using these parameters, we derived optimal threshold values for classification into the two classes for all the bins. Based on the threshold values, we judge the presence or absence of frazil ice signals for all time-series data of SV, starting from the uppermost bin. If frazil ice is judged to be present at the uppermost bin, the judgment of frazil ice is made for the second bin. Further judgment will be continued for the deeper bin until the absence of frazil ice is judged. As such, the penetration depth of frazil ice is determined for all time-series data, as shown in Figure 7c.
Figure 9 compares the penetration depth of frazil ice with the ice-type classification range on the PR–GR plane, showing that the penetration depth of frazil ice clearly differs between the two ice types. The AMSR ice-type classification can effectively detect deeper underwater frazil ice as active frazil with a detection rate of 77% for a penetration depth of 69 m (the deepest layer of the ADCP). Because the AMSR satellite sensor measures the surface, this detection capability could be indirectly caused by the relationship between the surface condition and penetration depth of frazil ice. In fact, the SV at the uppermost bin, which represents the near-surface condition, reflects the penetration depth of frazil ice with a strong correlation coefficient of 0.91. Such a strong relationship likely enables AMSR to detect deep underwater frazil ice.

5. Concluding Remarks

Satellite microwave radiometers are powerful tools for deriving thin-ice thickness information, which is necessary for estimating sea-ice production. The thin-ice algorithm developed by Nakata et al. [25] for AMSR-E can provide a better estimation of thin-ice thickness by discriminating thin-ice types compared with other algorithms. However, this algorithm has not been validated for AMSR2. In this study, we validated its applicability to AMSR2 using Sentinel-1 SAR, MODIS, and in situ ADCP observation data.
The algorithm by Nakata et al. [25] first classifies sea-ice areas into active frazil and (thin) solid ice based on a linear discriminant method using PR and GR and then estimates thin-ice thickness using a PR–thickness relationship separately for active frazil and thin solid ice. Synthetic aperture radar can distinguish between active frazil and solid ice. In this study, four Sentinel-1A images obtained around Antarctica were used to validate the ice-type classification. We obtained validation data for the sea-ice type from visual inspection of the Sentinel-1A SAR images. A comparison of these data with the AMSR2 classification results indicates that the algorithm can also discriminate these ice types from AMSR2 with comparable accuracy to that of AMSR-E.
To compare with ice thickness estimation from AMSR2, we obtained validation data of thin-ice thickness from ~230 clear-sky MODIS images collected around the Sea of Okhotsk, Bering Sea, Chukchi Sea, North of Baffin Bay, Ross Sea, offshore of Cape Darnley, and Vincennes Bay. Using this large global dataset, we newly derived the AMSR2 PR–thickness relationship for active frazil and thin solid ice separately. Then, we evaluated the performance of the AMSR2 ice thickness algorithm, which uses these new PR–thickness relationships.
The MAE between thicknesses from the algorithm and MODIS is calculated to be 2.0 cm and 5.0 cm for active frazil and thin solid ice, respectively. The calculated RMSE between them is almost the same as that for AMSR-E described by Nakata et al. [25]. The validation results show that the algorithm meets the target accuracies (active frazil: ±3 cm, thin solid ice: ±10 cm) for release as a JAXA research product. The validation areas cover global sea-ice regions during winter, indicating that the algorithm can be applied to global thin-ice regions: all coastal polynyas and also marginal ice zones for the early freeze-up period. The thin-ice thickness data calculated using this algorithm have been released from JAXA as the Thin-Ice Thickness (Thermal Ice Thickness) Version 1.0 of the AMSR2 Research Product.
Additionally, we validated the ice-type classification using in situ ADCP data obtained off Cape Darnley, Antarctica, in 2010. According to the comparison with the sea-ice type derived from the ADCP SV data of the uppermost bin (0–5 m depth), this algorithm can discriminate between the two ice types with an accuracy of 80%. Furthermore, the active frazil detection algorithm is effective in capturing deep-penetrating frazil ice.
There are several points to note when using this algorithm. During the initial ice-melting period, low ice-concentration areas are sometimes mistakenly identified as active frazil in the algorithm. Thus, sea-ice-type classification cannot be used during the melting season. The concept of our study is to develop a thin-ice thickness algorithm that can be applied to global ice-covered oceans. As shown in Table 2, there is a bias for the ice thickness estimates only in the Southern Hemisphere; the AMSR2 thickness is smaller than the MODIS thickness by ~4 cm for thin solid ice and ~2 cm for active frazil. This suggests that regional dependency still remains due to factors that are not accounted for in the algorithm, such as the effects of atmospheric vapor and snow on the ice surface. As indicated in Kashiwase et al. [12], in situ sea-ice thickness data such as those from an ice profiling sonar (IPS) will be able to be used to quantitatively assess the above effects. To reduce the regional dependency and improve the accuracy of thin-ice thickness estimates, there is room for improvement by considering atmospheric corrections and the effects of snow cover. One possible way to reduce the effect of atmospheric water vapor is the utilization of atmospheric reanalysis data as input for a radiative transfer model [12,50,51,52,53]. To mitigate issues related to snow cover, the use of TBs at lower-frequency channels may be helpful [54]. The errors in ice thickness estimates are also caused by the mixing of various surface types (open water, active frazil, thin solid ice, and thick ice) within an AMSR2 footprint. These errors could be reduced by using an algorithm developed by Kashiwase et al. [12,55], which accounts for a mixture of active frazil and thin solid ice.
In this study, we validated active frazil detection only for the Southern Ocean. For the Arctic Ocean, moored ADCP and IPS observations have been conducted in the Chukchi Sea coastal polynya [56]. These in situ data will be potentially utilized to further validate our algorithm and assess its global applicability. Additionally, in the Northern Hemisphere, frazil ice production plays a potentially significant role in transporting particulate matter. When frazil ice reaches the seabed or comes into contact with resuspended sediments, it could incorporate materials, including micro-nutrients such as iron [47,57,58]. The sediments are subsequently transported by ice floes and released when ice melts, which could lead to phytoplankton blooms [59]. The validation of AMSR2 active frazil detection using in situ observations in the Northern Hemisphere would be desirable to broadly identify coastal polynyas as drivers of bio-related material cycles and ecosystems.

Author Contributions

Conceptualization, K.N. and K.I.O.; methodology, K.N., M.K., R.S., E.Y., M.I. and K.I.O.; software, K.N.; validation, K.N.; formal analysis, K.N.; investigation, K.N.; resources, M.K.; data curation, K.N.; writing—original draft preparation, K.N.; writing—review and editing, M.K., R.S., E.Y., M.I. and K.I.O.; visualization, K.N.; supervision, K.I.O. and M.K.; project administration, K.I.O.; and funding acquisition, K.I.O. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by a research fund for the Global Change Observation Mission-Water (GCOM-W) of the Japan Aerospace Exploration Agency (JAXA) (PI Nos. ER2GWF404, ER3AMF424). This work was also supported by Grants-in-Aids for Scientific Research (20H05707, 20K20933) and the ArCS II project from the Ministry of Education, Culture, Sports, Science, and Technology in Japan.

Data Availability Statement

The GCOM-W/AMSR2 Level-1R brightness temperature and Level-2 ice concentration data were obtained from the Japan Aerospace Exploration Agency website (https://gportal.jaxa.jp/gpr/, accessed on 1 November 2024). Copernicus Sentinel-1 imagery used in this study is copyrighted by the European Space Agency and was downloaded through the Google Earth Engine (https://earthengine.google.com/, accessed on 1 November 2024). The Aqua/MODIS MYD021KM product was obtained from NASA’s Earth Observation System Data and Information System (https://ladsweb.modaps.eosdis.nasa.gov/, accessed on 1 November 2024). The ECMWF ERA5 product was downloaded from the Copernicus Climate Change Service Climate Data Store (https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-single-levels, accessed on 1 November 2024). The ADCP data used in this study were uploaded to the Arctic Data Archive System of the National Institute of Polar Research in Japan (https://doi.org/10.17592/001.2022081902, accessed on 1 November 2024). The AMSR-E brightness temperature and ice-concentration data were obtained from the website of the National Snow and Ice Data Center at the University of Colorado (https://nsidc.org/data/ae_l2a/versions/4, accessed on 1 November 2024; https://nsidc.org/data/ae_si12/versions/3, accessed on 1 November 2024). The GCOM-W/AMSR2 thin-ice thickness (thermal ice thickness) product can be obtained from the GCOM-W Research Product Distribution Service (https://suzaku.eorc.jaxa.jp/GCOM_W/research/, accessed on 1 November 2024).

Acknowledgments

We thank the three anonymous reviewers for their insightful comments and corrections that have improved this paper.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Maqueda, M.A.M.; Willmott, A.J.; Biggs, N.R.T. Polynya dynamics: A review of observations and modeling. Rev. Geophys. 2004, 42, RG1004. [Google Scholar] [CrossRef]
  2. Ohshima, K.; Fukamachi, Y.; Williams, G.; Nihashi, S.; Roquet, F.; Kitade, Y.; Tamura, T.; Hirano, D.; Herraiz-Borreguero, L.; Hindell, M.; et al. Antarctic Bottom Water production by intense sea-ice formation in the Cape Darnley polynya. Nat. Geosci. 2013, 6, 235–240. [Google Scholar] [CrossRef]
  3. Shcherbina, A.Y.; Talley, L.D.; Rudnick, D.L. Direct observations of North Pacific ventilation: Brine rejection in the Okhotsk Sea. Science 2003, 302, 1952–1955. [Google Scholar] [CrossRef]
  4. Aagaard, K.; Coachman, L.; Carmack, E. On the halocline of the Arctic Ocean. Deep Sea Res. Part A 1981, 28, 529–545. [Google Scholar] [CrossRef]
  5. Cavalieri, D.J.; Martin, S. The contribution of Alaskan, Siberian, and Canadian coastal polynyas to the cold halocline layer of the Arctic Ocean. J. Geophys. Res. 1994, 99, 18343–18362. [Google Scholar] [CrossRef]
  6. Hoppema, M.; Anderson, L.G. Biogeochemistry of polynyas and their role in sequestration of anthropogenic constituents. In Polynyas: Windows into Polar Oceans; Smith, W.O., Barber, D.G., Eds.; Elsevier Oceanography Series; Elsevier: Amsterdam, The Netherlands, 2007; Volume 74, pp. 193–221. [Google Scholar] [CrossRef]
  7. Yu, Y.; Rothrock, D.A. Thin ice thickness from satellite thermal imagery. J. Geophys. Res. 1996, 101, 25753–25766. [Google Scholar] [CrossRef]
  8. Cox, D.F.N.; Weeks, W.F. Salinity variations in sea ice. J. Glaciol. 1974, 13, 109–120. [Google Scholar] [CrossRef]
  9. Naoki, K.; Ukita, J.; Nishio, F.; Nakayama, M.; Comiso, J.C.; Gasiewski, A. Thin sea ice thickness as inferred from passive microwave and in situ observations. J. Geophys. Res. 2008, 113, C2. [Google Scholar] [CrossRef]
  10. Hwang, B.J.; Ehn, J.K.; Barber, D.G. Impact of ice temperature on microwave emissivity of thin newly formed sea ice. J. Geophys. Res. 2008, 113, C2. [Google Scholar] [CrossRef]
  11. Tamura, T.; Ohshima, K.I.; Markus, T.; Cavalieri, D.J.; Nihashi, S.; Hirasawa, N. Estimation of thin ice thickness and detection of fast ice from SSM/I data in the Antarctic Ocean. J. Atmos. Ocean. Technol. 2007, 24, 1757–1772. [Google Scholar] [CrossRef]
  12. Kashiwase, H.; Ohshima, K.I.; Fukamachi, Y.; Nihashi, S.; Tamura, T. Evaluation of AMSR-E thin ice thickness algorithm from a mooring-based observation: How can the satellite observe a sea-ice field with non-uniform thickness distribution? J. Atmos. Ocean. Technol. 2019, 36, 1623–1641. [Google Scholar] [CrossRef]
  13. Iwamoto, K.; Ohshima, K.I.; Tamura, T.; Nihashi, S. Estimation of thin ice thickness from AMSR-E data in the Chukchi Sea. Int. J. Remote Sens. 2013, 34, 468–489. [Google Scholar] [CrossRef]
  14. Martin, S.; Drucker, R.; Kwok, R.; Holt, B. Estimation of the thin ice thickness and heat flux for the Chukchi Sea Alaskan coast polynya from Special Sensor Microwave/Imager data, 1990–2001. J. Geophys. Res. 2004, 109, C10012. [Google Scholar] [CrossRef]
  15. Martin, S.; Drucker, R.; Kwok, R.; Holt, B. Improvements in the estimates of ice thickness and production in the Chukchi Sea polynyas derived from AMSR-E. Geophys. Res. Lett. 2005, 32, L05505. [Google Scholar] [CrossRef]
  16. Nihashi, S.; Ohshima, K.I.; Tamura, T.; Fukamachi, Y.; Saitoh, S. Thickness and production of sea ice in the Okhotsk Sea coastal polynyas from AMSR-E. J. Geophys. Res. 2009, 114, C10025. [Google Scholar] [CrossRef]
  17. Nihashi, S.; Ohshima, K.I.; Saitoh, S. Sea-ice production in the northern Japan Sea. Deep Sea Res. Part I Oceanogr. Res. Pap. 2017, 127, 65–76. [Google Scholar] [CrossRef]
  18. Nihashi, S.; Ohshima, K.I. Circumpolar mapping of Antarctic coastal polynyas and landfast sea ice: Relationship and variability. J. Clim. 2015, 28, 3650–3670. [Google Scholar] [CrossRef]
  19. Nihashi, S.; Ohshima, K.I.; Tamura, T. Sea-ice production in Antarctic coastal polynyas estimated from AMSR2 Data and its validation using AMSR-E and SSM/I-SSMIS data. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2017, 10, 3912–3922. [Google Scholar] [CrossRef]
  20. Nihashi, S.; Ohshima, K.I.; Tamura, T. Reconstruct the AMSR-E/2 thin ice thickness algorithm to create a long-term time series of sea-ice production in Antarctic coastal polynyas. Polar Sci. 2024, 39, 100978. [Google Scholar] [CrossRef]
  21. Tamura, T.; Ohshima, K.I.; Nihashi, S. Mapping of sea ice production for Antarctic coastal polynyas. Geophys. Res. Lett. 2008, 35, L07606. [Google Scholar] [CrossRef]
  22. Comiso, J.C.; Kwok, R.; Martin, S.; Gordon, A.L. Variability and trends in sea ice extent and ice production in the Ross Sea. J. Geophys. Res. 2011, 116, C04021. [Google Scholar] [CrossRef]
  23. Iwamoto, K.; Ohshima, K.I.; Tamura, T. Improved mapping of sea ice production in the Arctic Ocean using AMSR-E thin ice thickness algorithm. J. Geophys. Res. Oceans 2014, 119, 3574–3594. [Google Scholar] [CrossRef]
  24. Ohshima, K.I.; Tamura, N.; Kashiwase, H.; Nihashi, S.; Nakata, K.; Iwamoto, K. Estimation of sea ice production in the Bering Sea from AMSR-E and AMSR2 data, with special emphasis on the Anadyr polynya. J. Geophys. Res. Oceans 2020, 125, e2019JC016023. [Google Scholar] [CrossRef]
  25. Nakata, K.; Ohshima, K.I.; Nihashi, S. Estimation of thin ice thickness and discrimination of ice type from AMSR-E passive microwave data. IEEE Trans. Geosci. Remote Sens. 2019, 57, 263–276. [Google Scholar] [CrossRef]
  26. Nakata, K.; Ohshima, K.I.; Nihashi, S. Mapping of active frazil for Antarctic coastal polynyas, with an estimation of sea-ice production. Geophys. Res. Lett. 2021, 48, e2020GL091353. [Google Scholar] [CrossRef]
  27. Nakata, K.; Ohshima, K.I. Mapping of active frazil and sea ice production in the Northern Hemisphere, with comparison to the Southern Hemisphere. J. Geophys. Res. Oceans 2022, 127, e2022JC018553. [Google Scholar] [CrossRef]
  28. Ohshima, K.I.; Fukamachi, Y.; Ito, M.; Nakata, K.; Simizu, D.; Ono, K.; Nomura, D.; Hashida, G.; Tamura, T. Dominant frazil ice production in the Cape Darnley polynya leading to Antarctic Bottom Water formation. Sci. Adv. 2022, 8, eadc9174. [Google Scholar] [CrossRef]
  29. Maeda, T.; Taniguchi, Y.; Imaoka, K. GCOM-W1 AMSR2 Level 1R product: Dataset of brightness temperature modified using the antenna pattern matching technique. IEEE Trans. Geosci. Remote Sens. 2016, 54, 770–782. [Google Scholar] [CrossRef]
  30. Markus, T.; Cavalieri, D.J. An enhancement of the NASA team sea ice algorithm. IEEE Trans. Geosci. Remote Sens. 2000, 38, 1387–1398. [Google Scholar] [CrossRef]
  31. Comiso, J.C. Sea ice microwave emissivities from satellite passive microwave and infrared observations. J. Geophys. Res. 1983, 88, 7686–7704. [Google Scholar] [CrossRef]
  32. Comiso, J.C.; Cavalieri, D.; Parkinson, C.; Gloersen, P. Passive microwave algorithms for sea ice concentration: A comparison of two techniques. Remote Sens. Environ. 1997, 60, 357–384. [Google Scholar] [CrossRef]
  33. Comiso, J.C.; Nishio, F. Trends in the sea ice cover using enhanced and compatible AMSR-E, SSM/I, and SMMR data. J. Geophys. Res. 2008, 113, C023S07. [Google Scholar] [CrossRef]
  34. Nakata, K.; Ohshima, K.I.; Nihashi, S.; Kimura, N.; Tamura, T. Variability and ice production budget in the Ross Ice Shelf Polynya based on a simplified polynya model and satellite observations. J. Geophys. Res. Oceans 2015, 120, 6234–6252. [Google Scholar] [CrossRef]
  35. Riggs, G.; Hall, D. MODIS Sea Ice Products User Guide to Collection 6; National Snow and Ice Data Center, University of Colorado: Boulder, CO, USA, 2015; p. 50. [Google Scholar]
  36. Fraser, A.D.; Massom, R.A.; Michael, K.J. A method for compositing polar MODIS satellite images to remove cloud cover for landfast sea-ice detection. IEEE Trans. Geosci. Remote Sens. 2009, 47, 3272–3282. [Google Scholar] [CrossRef]
  37. Paul, S.; Huntemann, M. Improved machine-learning-based open-water–sea-ice–cloud discrimination over wintertime Antarctic sea ice using MODIS thermal-infrared imagery. Cryosphere 2021, 15, 1551–1565. [Google Scholar] [CrossRef]
  38. Key, J.R.; Collins, J.B.; Fowler, C.; Stone, R.S. High-latitude surface temperature estimates from thermal satellite data. Remote Sens. Environ. 1997, 61, 302–309. [Google Scholar] [CrossRef]
  39. Drucker, R.; Martin, S.; Moritz, R. Observations of ice thickness and frazil ice in the St. Lawrence Island polynya from satellite imagery, upward looking sonar, and salinity/temperature moorings. J. Geophys. Res. 2003, 108, 3149–3166. [Google Scholar] [CrossRef]
  40. Mäkynen, M.; Similä, M. Thin ice detection in the Barents and Kara seas with AMSR-E and SSMIS radiometer data. IEEE Trans. Geosci. Remote Sens. 2015, 53, 5036–5053. [Google Scholar] [CrossRef]
  41. Ohshima, K.I.; Watanabe, T.; Nihashi, S. Surface heat budget of the Sea of Okhotsk during 1987–2001 and the role of sea ice on it. J. Meteorol. Soc. Jpn. 2003, 81, 653–677. [Google Scholar] [CrossRef]
  42. Kondo, J. Air-sea transfer coefficients in diabatic conditions. Bound.-Layer Meteorol. 1975, 9, 91–112. [Google Scholar] [CrossRef]
  43. Guest, P.S. Surface longwave radiation conditions in the eastern Weddell Sea during winter. J. Geophys. Res. 1998, 103, 30761–30771. [Google Scholar] [CrossRef]
  44. Hersbach, H.; Bell, B.; Berrisford, P.; Biavati, G.; Horányi, A.; Muñoz Sabater, J.; Nicolas, J.; Peubey, C.; Radu, R.; Rozum, I.; et al. ERA5 Hourly Data on Single Levels from 1959 to Present. Copernicus Climate Change Service (C3S) Climate Data Store (CDS). 2018. Available online: https://cds.climate.copernicus.eu/datasets/reanalysis-era5-single-levels?tab=overview (accessed on 1 November 2024). [CrossRef]
  45. Hersbach, H.; Bell, B.; Berrisford, P.; Hirahara, S.; Horányi, A.; Muñoz-Sabater, J.; Nicolas, J.; Peubey, C.; Radu, R.; Schepers, D.; et al. The ERA5 global reanalysis. Q. J. R. Meteorol. Soc. 2020, 146, 1999–2049. [Google Scholar] [CrossRef]
  46. Deines, K.L. Backscatter estimation using broadband acoustic Doppler current profilers. In Proceedings of the IEEE Sixth Working Conference on Current Measurement (Cat. No. 99CH36331), San Diego, CA, USA, 11–13 March 1999; IEEE: Piscataway, NJ, USA, 1999; pp. 249–253. [Google Scholar]
  47. Ito, M.; Ohshima, K.I.; Fukamachi, Y.; Hirano, D.; Mahoney, A.R.; Jones, J.; Takatsuka, T.; Eicken, H. Favorable conditions for suspension freezing in an Arctic coastal polynya. J. Geophys. Res. 2019, 124, 8701–8719. [Google Scholar] [CrossRef]
  48. Dempster, A.P.; Laird, N.M.; Rubin, D.B. Maximum Likelihood from Incomplete Data Via the EM Algorithm. J. Royal Stat. Soc. B 1977, 39, 1–22. [Google Scholar] [CrossRef]
  49. Moon, T.K. The expectation-maximization algorithm. IEEE Signal Process Mag. 1996, 13, 47–60. [Google Scholar] [CrossRef]
  50. Mathew, N.; Heygster, G.; Melsheimer, C. Surface Emissivity of the Arctic Sea Ice at AMSR-E Frequencies. IEEE Trans. Geosci. Remote Sens. 2009, 47, 4115–4124. [Google Scholar] [CrossRef]
  51. Andersen, S.; Tonboe, R.; Kern, S.; Schyberg, H. Improved Retrieval of Sea Ice Total Concentration from Spaceborne Passive Microwave Observations Using Numerical Weather Prediction Model Fields: An Intercomparison of Nine Algorithms. Remote Sens. Environ. 2006, 104, 374–392. [Google Scholar] [CrossRef]
  52. Lavergne, T.; Sørensen, A.M.; Kern, S.; Tonboe, R.; Notz, D.; Aaboe, S.; Bell, L.; Dybkjær, G.; Eastwood, S.; Gabarro, C.; et al. Version 2 of the EUMETSAT OSI SAF and ESA CCI Sea-Ice Concentration Climate Data Records. Cryosphere 2019, 13, 49–78. [Google Scholar] [CrossRef]
  53. Mäkynen, M.; Similä, M. Arctic Thin Ice Detection Using AMSR2 and FY-3C MWRI Radiometer Data. Remote Sens. 2024, 16, 1600. [Google Scholar] [CrossRef]
  54. Shokr, M.; Asmus, K.; Agnew, T.A. Microwave Emission Observations from Artificial Thin Sea Ice: The Ice-Tank Experiment. IEEE Trans. Geosci. Remote Sens. 2009, 47, 325–338. [Google Scholar] [CrossRef]
  55. Kashiwase, H.; Ohshima, K.I.; Nakata, K.; Tamura, T. Improved SSM/I Thin Ice Algorithm with Ice Type Discrimination in Coastal Polynyas. J. Atmos. Oceanic Technol. 2021, 38, 823–835. [Google Scholar] [CrossRef]
  56. Fukamachi, Y.; Simizu, D.; Ohshima, K.I.; Eicken, H.; Mahoney, A.R.; Iwamoto, K.; Moriya, E.; Nihashi, S. Sea-ice thickness in the coastal northeastern Chukchi Sea from moored ice-profiling sonar. J. Glaciol. 2017, 63, 888–898. [Google Scholar] [CrossRef]
  57. Dethleff, D.; Kuhlmann, G. Entrainment of fine-grained surface deposits into new ice in the southwestern Kara Sea, Siberian Arctic. Cont. Shelf Res. 2009, 29, 691–701. [Google Scholar] [CrossRef]
  58. Ito, M.; Ohshima, K.I.; Fukamachi, Y.; Mizuta, G.; Kusumoto, Y.; Kikuchi, T. Underwater frazil ice and its suspension depth detected from ADCP backscatter data around sea ice edge in the Sea of Okhotsk. Cold Reg. Sci. Technol. 2021, 192, 103382. [Google Scholar] [CrossRef]
  59. Krumpen, T.; Belter, H.J.; Boetius, A.; Damm, E.; Haas, C.; Hendricks, S.; Nicolaus, M.; Nöthig, E.M.; Paul, S.; Peeken, I.; et al. Arctic warming interrupts the Transpolar Drift and affects long-range transport of sea ice and ice-rafted matter. Sci. Rep. 2019, 9, 5459. [Google Scholar] [CrossRef] [PubMed]
Remotesensing 17 00171 g001
Figure 1. Locations of four Sentinel-1 images acquired at Terpenia Bay, Anadyl Bay, St. Lawrence Island, the Chukchi Sea (top left panel), Mackenzie Bay, Shackleton polynya, Vincennes Bay, and the Ross Sea (top right panel), indicated by blue rectangles. The red cross marks the location of the ADCP observation off Cape Darnley (68.84°E, 67.63°S), which is utilized in the analysis in Section 4. Main large panel: a Sentinel-1 SAR image of the Shackleton polynya. The regions enclosed by the red and blue lines represent areas of active frazil and solid ice, respectively, identified by visual inspection.
Figure 1. Locations of four Sentinel-1 images acquired at Terpenia Bay, Anadyl Bay, St. Lawrence Island, the Chukchi Sea (top left panel), Mackenzie Bay, Shackleton polynya, Vincennes Bay, and the Ross Sea (top right panel), indicated by blue rectangles. The red cross marks the location of the ADCP observation off Cape Darnley (68.84°E, 67.63°S), which is utilized in the analysis in Section 4. Main large panel: a Sentinel-1 SAR image of the Shackleton polynya. The regions enclosed by the red and blue lines represent areas of active frazil and solid ice, respectively, identified by visual inspection.
Remotesensing 17 00171 g001
Remotesensing 17 00171 g002
Figure 2. (a,c) Frequency map of all MODIS validation data points, and (b,d) frequency map of MODIS data points with ice thickness of <30 cm for the Northern (a,b) and Southern (c,d) Hemispheres.
Figure 2. (a,c) Frequency map of all MODIS validation data points, and (b,d) frequency map of MODIS data points with ice thickness of <30 cm for the Northern (a,b) and Southern (c,d) Hemispheres.
Remotesensing 17 00171 g002
Remotesensing 17 00171 g003
Figure 3. Scatterplots of (a) PR versus GR and (b) PR versus MODIS thickness. Red and blue marks indicate the data for active frazil and solid ice, respectively, in both panels (a,b). The ice types in Figure 3a,b are identified based on visual inspections of Sentinel-1 and Equation (2), respectively. The background colors in Figure 3a indicate the classification range of active frazil (red) and solid ice (blue) using the AMSR-E algorithm of Nakata et al. [25]. The solid and dashed lines in Figure 3b show the fitted exponential curves for active frazil (yellow) and thin solid ice (light blue) derived in this study and by Nakata et al. [25], respectively.
Figure 3. Scatterplots of (a) PR versus GR and (b) PR versus MODIS thickness. Red and blue marks indicate the data for active frazil and solid ice, respectively, in both panels (a,b). The ice types in Figure 3a,b are identified based on visual inspections of Sentinel-1 and Equation (2), respectively. The background colors in Figure 3a indicate the classification range of active frazil (red) and solid ice (blue) using the AMSR-E algorithm of Nakata et al. [25]. The solid and dashed lines in Figure 3b show the fitted exponential curves for active frazil (yellow) and thin solid ice (light blue) derived in this study and by Nakata et al. [25], respectively.
Remotesensing 17 00171 g003
Remotesensing 17 00171 g004
Figure 4. Two-dimensional histograms of AMSR2 versus MODIS ice thickness for the (a) Northern Hemisphere, (b) Southern Hemisphere, and (c) global ocean.
Figure 4. Two-dimensional histograms of AMSR2 versus MODIS ice thickness for the (a) Northern Hemisphere, (b) Southern Hemisphere, and (c) global ocean.
Remotesensing 17 00171 g004
Remotesensing 17 00171 g005
Figure 5. (a) MODIS visible image, and spatial distributions of (b) MODIS surface temperature, (c) Sentinel-1 SAR backscatter, (d) MODIS thin-ice thickness, (e) AMSR2 ice type (AF: active frazil, TN: thin solid ice, TK: thick solid ice), and (f) AMSR2 thin-ice thickness in Vincennes Bay, Antarctica, on 9 August 2016.
Figure 5. (a) MODIS visible image, and spatial distributions of (b) MODIS surface temperature, (c) Sentinel-1 SAR backscatter, (d) MODIS thin-ice thickness, (e) AMSR2 ice type (AF: active frazil, TN: thin solid ice, TK: thick solid ice), and (f) AMSR2 thin-ice thickness in Vincennes Bay, Antarctica, on 9 August 2016.
Remotesensing 17 00171 g005
Remotesensing 17 00171 g006
Figure 6. (a) MODIS visible image, and spatial distributions of (b) MODIS surface temperature, (c) Sentinel-1 SAR backscatter, (d) MODIS thin-ice thickness, (e) AMSR2 ice type (AF: active frazil, TN: thin solid ice, TK: thick solid ice), and (f) AMSR2 thin-ice thickness in Anadyr Bay, the Bering Sea, on 3 February 2022.
Figure 6. (a) MODIS visible image, and spatial distributions of (b) MODIS surface temperature, (c) Sentinel-1 SAR backscatter, (d) MODIS thin-ice thickness, (e) AMSR2 ice type (AF: active frazil, TN: thin solid ice, TK: thick solid ice), and (f) AMSR2 thin-ice thickness in Anadyr Bay, the Bering Sea, on 3 February 2022.
Remotesensing 17 00171 g006
Remotesensing 17 00171 g007
Figure 7. Time series of (a) ADCP backscatter strength (SV) at the uppermost bin (0–5 m depth) (green lines) and AMSR-E active frazil detection (yellow bar), (b) vertical profile of ADCP SV, and (c) penetration depth of frazil ice extracted from our method, from March to November 2010 off Cape Darnley, Antarctica. (d) Enlarged (July) time series of vertical profile and (e) surface (uppermost bin) ADCP SV. The dashed lines in (e) indicate the ASAR-interpreted sea-ice types (red: active frazil, blue: solid ice, green: boundary). Scale marks with month labels indicate the first day of each month.
Figure 7. Time series of (a) ADCP backscatter strength (SV) at the uppermost bin (0–5 m depth) (green lines) and AMSR-E active frazil detection (yellow bar), (b) vertical profile of ADCP SV, and (c) penetration depth of frazil ice extracted from our method, from March to November 2010 off Cape Darnley, Antarctica. (d) Enlarged (July) time series of vertical profile and (e) surface (uppermost bin) ADCP SV. The dashed lines in (e) indicate the ASAR-interpreted sea-ice types (red: active frazil, blue: solid ice, green: boundary). Scale marks with month labels indicate the first day of each month.
Remotesensing 17 00171 g007
Remotesensing 17 00171 g008
Figure 8. (a) Histogram of the ADCP SV for active frazil (orange) and solid ice (blue) detected by the AMSR algorithm, based on hourly averaged SV data from April to October. The red and blue lines indicate the kernel density estimates of the probability density distributions for the active frazil and solid ice. (b) “Recall”, “Precision”, and “Accuracy” of active frazil detection as a function of the threshold value of ADCP SV. The gray shading and black lines in (b) indicate the optimal threshold range and the boundary between active frazil and solid ice, respectively, inferred from the SAR images.
Figure 8. (a) Histogram of the ADCP SV for active frazil (orange) and solid ice (blue) detected by the AMSR algorithm, based on hourly averaged SV data from April to October. The red and blue lines indicate the kernel density estimates of the probability density distributions for the active frazil and solid ice. (b) “Recall”, “Precision”, and “Accuracy” of active frazil detection as a function of the threshold value of ADCP SV. The gray shading and black lines in (b) indicate the optimal threshold range and the boundary between active frazil and solid ice, respectively, inferred from the SAR images.
Remotesensing 17 00171 g008
Remotesensing 17 00171 g009
Figure 9. Scatterplots of PR versus GR colored by the penetration depth of frazil ice estimated from the vertical profile of the ADCP SV. The background colors indicate the classification ranges of active frazil (orange) and solid ice (purple) using the AMSR algorithm.
Figure 9. Scatterplots of PR versus GR colored by the penetration depth of frazil ice estimated from the vertical profile of the ADCP SV. The background colors indicate the classification ranges of active frazil (orange) and solid ice (purple) using the AMSR algorithm.
Remotesensing 17 00171 g009
Table 1. Locations, dates, and times of Sentinel-1 and AMSR2 data.
Table 1. Locations, dates, and times of Sentinel-1 and AMSR2 data.
LocationModeDateTime (UTC)
Sentinel-1AMSR2
Northern HemisphereTerpenia Bay IW24 January 201808:1202:42
Anadyr BayIW3 February 202018:0614:41
St. Lawrence IslandIW12 February 202117:4314:46
Chukchi SeaIW27 January 201917:1615:04
Southern HemisphereRoss SeaEW 11 July 201610:3610:20
Vincennes BayEW 9 August 201612:27 07:22
Shackleton Ice ShelfEW 11 August 201613:5016:13
Mackenzie BayEW 19 June 201715:2811:14
Note: “Mode” indicates the acquisition mode of Sentinel-1, Interferometric Wide Swath (IW) and Extra Wide Swath (EW). The IW mode provides a swath width of 250 km with a spatial resolution of 5 m × 20 m. The EW mode provides a swath width of 400 km with a spatial resolution of 20 m × 40 m.
Table 2. Validation results for the thin-ice thickness estimation.
Table 2. Validation results for the thin-ice thickness estimation.
CategoryMetricsNHSHAll
Thin-ice thickness estimation (cm)Active frazilMAE1.42.82.0
RMSE2.94.83.8
Bias 0.12.01.6
Thin solid iceMAE4.05.85.0
RMSE5.37.56.6
Bias0.14.22.7
Total thin iceMAE3.85.74.8
RMSE5.27.46.5
Bias0.14.12.6
Thin-ice detectionUnbalancedRecall0.71 0.73 0.72
Precision0.86 0.42 0.58
Accuracy0.95 0.92 0.93
BalancedRecall0.71 0.73 0.72
Precision0.98 0.92 0.94
Accuracy0.85 0.83 0.84
Note: The upper section represents the error in thin-ice thickness estimation, and the lower section represents the performance metrics for thin-ice detection. The error in thin-ice thickness estimation is represented by MAE, RMSE, and “Bias” for each ice type. The “Bias” is calculated by subtracting AMSR2 ice thickness from MODIS ice thickness. The metrics for thin-ice detection are as follows: “Recall” (true positive rate) represents the proportion of the number of thin-ice cases correctly detected by AMSR2 among the total number of MODIS-detected thin-ice cases; a high “Recall” indicates a low missed-detection rate. “Precision” represents the proportion of the number of thin-ice cases correctly detected by AMSR2 among the total number of AMSR2-detected thin-ice cases; a high “Precision” means a low false-detection rate (i.e., misclassifying thick ice as thin ice). “Accuracy” represents the proportion of correct classifications (sum of true positives and true negatives) among the total number of samples. “Balanced” represents the result when the value of the metric is modified, assuming that the numbers of thick-ice and thin-ice samples are equal.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Nakata, K.; Kachi, M.; Shimada, R.; Yoshizawa, E.; Ito, M.; Ohshima, K.I. Validation of an AMSR2 Thin-Ice Thickness Algorithm for Global Sea-Ice-Covered Oceans Using Satellite and In Situ Observations. Remote Sens. 2025, 17, 171. https://doi.org/10.3390/rs17010171

AMA Style

Nakata K, Kachi M, Shimada R, Yoshizawa E, Ito M, Ohshima KI. Validation of an AMSR2 Thin-Ice Thickness Algorithm for Global Sea-Ice-Covered Oceans Using Satellite and In Situ Observations. Remote Sensing. 2025; 17(1):171. https://doi.org/10.3390/rs17010171

Chicago/Turabian Style

Nakata, Kazuki, Misako Kachi, Rigen Shimada, Eri Yoshizawa, Masato Ito, and Kay I. Ohshima. 2025. "Validation of an AMSR2 Thin-Ice Thickness Algorithm for Global Sea-Ice-Covered Oceans Using Satellite and In Situ Observations" Remote Sensing 17, no. 1: 171. https://doi.org/10.3390/rs17010171

APA Style

Nakata, K., Kachi, M., Shimada, R., Yoshizawa, E., Ito, M., & Ohshima, K. I. (2025). Validation of an AMSR2 Thin-Ice Thickness Algorithm for Global Sea-Ice-Covered Oceans Using Satellite and In Situ Observations. Remote Sensing, 17(1), 171. https://doi.org/10.3390/rs17010171

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop