| Citation: | Chengfei Jiang, Mingsen Lin, Ruixue Cao, Hao Wei, Lijian Shi, Bin Cheng, Yongjun Jia, Qimao Wang. Classification of ice and water in the Arctic using radar altimeter and microwave radiometer data from HY-2B satellite[J]. Acta Oceanologica Sinica, 2023, 42(5): 179-191. doi: 10.1007/s13131-022-2067-4
|
The Arctic Ocean is a critical component of Earth’s climate system (Aagaard and Carmack, 1994). The Arctic region affects the global climate and is also sensitive to global climate change (Przybylak, 2007). The air temperature in the Arctic Ocean has increased twice as fast as the global average. This is known as the Arctic Amplification (Screen and Simmonds, 2010). As a result, the sea ice in the Arctic Ocean has significantly decreased in recent decades in terms of both the sea ice extent and the sea ice thickness. Accurate observations of sea ice have become quite important.
Satellite remote sensing is a powerful tool for Earth Observation (EO) in the Arctic region. Various sensors, such as microwave scatterometers, microwave scanning radiometers, radar altimeters, and synthetic aperture radar (SAR) devices, have provided abundant information from various satellites. In addition to single sensor observations, the combined passive and active microwave sensors data can improve sea ice observational capabilities (Zhang et al., 2019; Swan and Long, 2012). The Chinese HaiYang-2B (HY-2B) satellite carried various types of active and passive microwave sensors, which has the capacity to detect the Polar Oceans between latitudes from 66° to 80°.
The radar altimeter (RA) is an important active microwave sensor on the HY-2B satellite. A nadir-looking observe passive microwave sensor is equipped on HY-2B in conjunction with the RA. When the RA signal passes through the atmosphere between the platform and the ground targets, the signal transmission path will be delayed due to the impact of atmospheric water vapor and other substances (Fu et al., 1994). Although the non-specialized calibration microwave radiometers (CMR) or climate models could be used to correct the atmospheric pathway delay, the accuracy is often not very high (Keihm et al., 1995; Fleming et al., 1991). Therefore, a kind of radiometer device, equipped on the same platform with similar observation angles as the radar altimeters, is considered auxiliary equipment to correct atmospheric pathway delay for radar altimeters (Keihm et al., 1995). Many satellites use this type of design at present (Obligis et al., 2006; Louet and Bruzzi, 1999; Pol et al., 1998; Bernard et al., 1993). It is the nadir-looking observe passive microwave sensor on HY-2B performs implements this feature. We refer to it as a “CMR” (Jiang et al., 2012). In addition to providing atmospheric pathway delay corrections, the radiometers also could detect objects as they have the frequency bands for ground observation. There have been some studies on large lakes and bays (Kouraev et al., 2004a, b, 2007, 2008, 2011); therefore, they have a great potential for sea ice observations and can play vital roles in the future polar research.
The RA of the HY-2B satellite has two signal tracking modes: the ocean model Sub-Optimal Maximum Likelihood Estimate (SMLE) mode and the land model Offset Center of Gravity (OCOG) mode (Xu et al., 2007, 2014). The model compatibility tracker on HY-2B’s RA selects the appropriate tracking mode by distinguishing between the echoes of the different features, and thus the signal unlocking phenomenon caused by the sudden changes in the topography detected by a single tracker along the orbit is eliminated. The SMLE tracker can provide more accurate results for ocean measurements than the OCOG tracker (Xu et al., 2013, 2014). However, sometimes the signals from different objects are very similar, so it is difficult for the model compatibility tracker on HY-2B to recognize the types of ground objects 100%, thus losing part of the data. The subsequent processing of RA waveform data returned from different features is different. So, it is of quite significance to distinguish sea ice and water accurately from the original information of RA data.
This study focused on the identification of sea ice and seawater using passive and active microwave sensor data from HY-2B satellite to provide ideas for making accurate ice flags of RA data. Three methods (single parameter threshold, multi-parameter linear segmentation, and K-means) were used for the sea ice identification in this study.
In general, ice identification for radar altimeters is mainly derived from comparisons of differences in tropospheric calibration results between calibration radiometers and models, or from the use of climatic state sea ice data for identification (Batoula, 2011; Bronner et al., 2013). Due to the lack of data on the difference between the model wet tropospheric correction and the wet tropospheric correction retrieved from radiometer brightness temperatures, we select the Satellite Application Facility on Ocean and Sea Ice (OSI-SAF) ice concentration data to compare the sea ice and seawater identification results. The results of the three classification methods compared with those of OSI-SAF show large differences in ice water separation near the ice edge. Therefore, this study developed a method to correct the ice and water separation using waveform parameters based on linear segmentation and extraction of the ice edges using an image processing method. The purpose of this study was to obtain the best results possible. Moreover, the factors influencing the ice water classification were investigated.
We use the RA’s Ku band data. A signal is transmitted through space to a target at earth’s surface and a reflected microwave echo is received by RA. The sea surface height, significant wave height, and backscatter coefficient (B_coe) can be calculated from the echo waveform data (Jiang et al., 2015). The tracker of the RA generates a waveform to detect features approximately every 50 ms. Each waveform has 128 samplings (Xu et al., 2013, 2017).
The CMR equipped on the HY-2B is a nadir-looking three-frequency (18.7 GHz, 23.8 GHz, and 37 GHz) microwave radiometer (Zhang et al., 2015). The observation angles of these bands are 2.4°, 2.2°, and 0°, respectively (Fig. 1). All three bands are linear polarization modules. The linearity is greater than 0.999 (
Two primary data products are at our disposal, i.e., the HY-2B RA Level 1B and the HY-2B CMR Level 1B products. All data are available from
The RA Level 1B product contained 20 Hz waveforms and auxiliary data. This level of data supported the SMLE and OCOG tracking package flags. In this study, we also use the data from the RA Level 1 product to calculate the RA’s B_coe (see Eq. (1)). The B_coe values were calculated using the Automatic Gain Control (AGC) values and some ancillary data from the RA Level 1B products, which were supplied in the Level 1B packages. All the data with AGC values greater than 75 is removed.
| $$ \begin{split}{\bf{\sigma }}^{{{0}}}=& {\rm{AGC}}_{\rm{corrected}}+30\;{\rm{log}}_{{{10}}}\left[{{H}}\left(1+\frac{{H}}{{{R}}_{\rm{e}}}\right)\right]-\\ &30\;{\rm{log}}_{{{10}}}\left[{{H}}_{{{0}}}\left(1+\frac{{{H}}_{0}}{{{R}}_{\rm{e}}}\right)\right]+{\rm{Value}}\_{\rm{modify}} , \end{split}$$ | (1) |
where AGCcorrected denotes the AGC value after the lookup table correction and fine tuning; H is the satellite altitude; H0 is the altitude set for the satellite equals 970 km; and Re is the earth radius (6 371 km). The “Value_modify” is a cross-calibration parameter between the HY-2A/B and Jason 2 satellites and the in-situ observation (
The brightness temperature (BT) is obtained from the HY-2B CMR Level 1B products. The BT sensor has the same observation tracks as the HY-2 ground orbit trajectories because of its a nadir-looking observe passive microwave sensor.
The HY-2B satellite platform operates on a solar synchronous, fixed orbit, with a repetition period of 14 days (Xu et al., 2017) (Fig. 2). The orbital inclination of HY-2B is 99.34015°N (Jiang et al., 2012). Both the RA and CMR are the nadir-looking instruments constrained by the satellite orbit. The instruments cover the northern hemisphere up to about 81°N but leave a sizeable no-data region in central Arctic.
In this study, we select the data from the entire year of 2019, and only the data at latitudes above 60°N are investigated. All the B_coe data and BT data for the sea ice and seawater are selected through the removal of the points with land mask.
BT is often used in studies of sea ice (Markus and Cavalieri, 2000; Comiso, 1995; Lomax et al., 1995). The CMR’s three-band 2019 annual BT and joint histograms are illustrated in Fig. 3. We see clear discrimination of BT using 160 K. Three bands’ BT in the range of about 100 K to 300 K. The different colors of Fig. 3 represent the distribution probability of the sampled points in different BT values.
The CMR’s three-band 2019 annual BT and joint histograms are illustrated in Fig. 3. We can find a clear band distribution of the three bands’ BT across the annual seasons. Among the three bands’ data, the 23.8 GHz and 37.0 GHz BT showed strong similarity. The high BT appears at 200–260 K, and the low value is 100–180 K. They represent the high BT of sea ice and the low BT of seawater, respectively (Eppler et al., 1992). The BT at the lowest band (18.7 GHz) indicated strongest cluster between sea ice and sea water. For all bands, the clusters are better in cold condition compared with the warm condition between early July and late September. Among three bands data, the 23.8 GHz and 37.0 GHz BT showed strong similarity.
The daily histograms of the B_coe values in 2019 are shown in Fig. 4.
The B_coes are related to the roughness of the reflecting surfaces, the dielectric constants, and the reflections off the sub-surfaces (Shokr and Sinha, 2015).
As shown in Fig. 4, unlike the CMR BT values, the RA B_coe values did not always appear as a more obvious phenomenon of two or three clusters. There was only one obvious cluster observed at around 10 dB to 20 dB, and it lasted the entire year. Another cluster (40 dB to 60 dB) became apparent during the summer months from late June until September. There was no obvious differentiation between the two clusters during the other months.
The sea ice surface properties changed when melting occurred (Swan and Long, 2012). The dielectric constant and roughness values of the sea ice surfaces also changed in response to the melting, resulting in changes in the B_coe values detected by the radar altimeter. These changes in the RA’s B_coe values can be used to detect the changes in the sea ice’s surface properties (Woodhouse, 2017).
The OSI-SAF ice concentration product and SAR data are used for the data validation in the subsequent classifications.
(1) OSI-SAF ice concentration product: The European Organization for the Exploitation of Meteorological Satellites (EUMETSAT) OSI-SAF supports the sea ice concentration production (OSI-401-b) (ftp://OSI-SAF.met.no/archive/ice/conc). The resolution of the OSI-SAF ice concentration products is 10 km. As in more research, a threshold is used to identify ice and water, in this study the same threshold of 30% as OSI-SAF ice products is chosen (Aaboe et al., 2021).
(2) Sentinel 1 a/b SAR data: The data source of the Sentinel 1 a/b SAR data was obtained from the Technical University of Denmark (DTU). The DTU provides mosaic images and individual images. The resolution of the mosaic image is 1 km. All the SAR images can be downloaded free of charge from the website (
Since the HY-2B’s model compatible tracker can automatically judge the target object, its different tracking model flags can be used as a ground object division method (Xu et al., 2017). After the land mask, the targets of the remaining feature are sea ice and seawater. The SMLE tracking package flags refer to seawater. The data from the land have been removed as the land flags masked, so the OCOG tracking package flags, which initially referred to the no-seawater target, could be considered sea ice. We select one day from the summertime of 2019 to show the distribution of the tracking package flags when the sea ice is melting.
Although the SMLE packages are mainly distributed in the seawater areas, they were also partially distributed in the sea ice areas (Fig. 5). Similarly, the OCOG packages also appeared in the seawater. Therefore, we needed to find other ways to improve the ice and water classification accuracy using the detection data collected by the satellite’s sensors.
According to the characteristics of the Arctic sea ice variations over an entire year (2019), which we show in the last section, the B_coe and BT values are used to linearly segment the ice water based on a single parameter or multiple parameters. In addition, one machine learning unsupervised classification method (K-means) is selected in the study.
The single parameter threshold method is widely used to classify sea ice and seawater (Swan and Long, 2012; Scheuchl et al., 2001; Laxon, 1994). As BT and B_coe are shown in Figs 3 and 4, BT has a striking characteristic to separate sea ice and seawater compared with the blurred feature of B_coe. Moreover, BT at 18.7 GHz yields more sharp discrimination of sea ice and seawater. It is, therefore, picked up as a single parameter threshold to do the classification. To be more specific in this study, the 18.7 GHz BT at 160 K as the segmentation threshold is used to provide separation of the sea ice and seawater (Fig. 3a).
In contrast to the single parameter threshold, the multi-parameter linear segmentation is to apply multiple satellite parameters to do the classification. In this study, this method refers to combine 18.7 GHz BT of CMR with the B_coe from the RA, since this combination from two different sensors allow us to benefit of the advantages of both data sources.
Joint mappings of the CMR’s BT and the RA’s B_coes are completed every half month, which is approximately equal to a cycle (14-day) of the HY-2B (cf. Fig. 6). Statistical analyses are performed every 1 dB or 1 K.
The sea ice and seawater clusters joint histograms throughout the entire annual 2019 are shown in Fig. 6. The distribution of the sea ice exhibited pronounced seasonal variation characteristics. The seawater is mainly distributed throughout the year in areas with low BT and low B_coe. The distribution area of sea ice, on the other hand, varies with seasonal changes. There is no overlap between the main distribution areas of the sea ice and seawater.
The sea ice and seawater were not only separated from each other, but they were also closely related. A shifting connection tube was observed between the two clusters, and it was wide in September and narrow during the other months. The cluster of sea ice moves towards the high B_coe and low BT in the summertime. In the freezing season, the cluster of sea ice is distributed in the narrow region with a high BT and wide B_coe range.
As is shown in Fig. 6, there were still relatively fixed separation boundaries between the two clusters of sea ice and seawater, regardless of whether the data were for one day or half a month. This type of stable boundary fulfilled a prerequisite for the linear segmentation conducted in this study.
In order to be able to use a simple classification rule for the separation of ice and water, it is, therefore, necessary that the rule can conform to the changing patterns of sea ice and seawater in different seasons. Based on the distribution pattern of sea ice and seawater in Fig. 6, examples were selected every three months. These days covered the four seasons in the Arctic (winter (2019-02-17), spring (2019-05-15), summer (2019-08-14), and autumn (2019-11-15)) including the entirety of the sea-ice freezing and melting processes.
In this study, according to the annual distributions of the sea ice and seawater, a first-degree equation is used as the boundary of the ice water separation. According to the distribution of the sea ice and seawater in Figs 6 and 7 during the winter and summer seasons, the B_coe aggregation point of the seawater was determined to be approximately 5 dB to 10 dB. The summer sea ice backscatter signal value aggregation ranged from 50 dB to 55 dB. In addition, the sea ice’s low BT values in the 18.7 GHz frequency band during the winter season were observed to be greater than 200 K. Moreover, the lowest BT value of the seawater was approximately 125 K. So we chose two points (10 dB, 175 K; and 55 dB, 125 K) that are clearly distinguishable in all seasons for the creation of the split line. A one-degree equation is used to estimate the boundary between sea ice and sea water.
By assuming BT value varies as B_coe, the partition line expression is shown in Eq. (2) below:
| $$ \begin{split} -1\;675/9 + &(10 \times {\rm{backscatter}} \; {\rm{coefficient}}) / 9\; + \\ &{\rm{brightness}}\; {\rm{temperature}} = 0 .\end{split} $$ | (2) |
The K-means is an algorithm that is widely used in unsupervised sea ice classification processes (Gill and Yackel, 2012; Remund et al., 2000). The main concept of the algorithm is to select K initial clustering centers from the datasets randomly, calculate the Euclidian distances between each element and the centers; and then classify them (Forgy, 1965). The average values of each element category are then calculated as the new cluster centers, and the iteration is repeated until the centers no longer change. Unlike threshold segmentation, K-means is an unsupervised classification method and does not need require the definition of a specific threshold to distinguish sea ice from seawater.
The classification results are evaluated according to the classification results of the OSI-SAF sea ice concentration, and the type of evaluation method was based on the calculation strategy for classification accuracy presented by Zygmuntowska et al., (2013), in which the number of correctly classified points were counted as a percentage of such aggregate data:
| $$ {\rm{Accuracy}}=\frac{{\rm{class}}_{{{\rm{class}}}\left({{n}}\right)}\cap {\rm{known}}_{\rm{class}\left({{{n}}}\right)}}{{\rm{known}}_{\rm{class}\left({n}\right)}}\times {{100}}{\%}. $$ | (3) |
In Eq. (3), class refers to the classification results of the different classification methods; known indicates the classification results for reference purposes; and class(n) refers to the types of features.
In this study, the aforementioned calculation method (single parameter threshold, multi-parameter linear segmentation, K-means) are used to correspond to the OSI-SAF ice-concentration production as a known classification, and the classification results are determined to be consistent with the results of the OSI-SAF product.
The daily accuracy ratings of the threshold segmentation using a BT value 160 K in the 18.7 GHz band are shown in Fig. 8. Overall, the classification of sea ice was very consistent with the results of the OSI-SAF ice-type products (higher than 95%). However, the classification results of seawater had a lower consistency from May to September. Starting in May, the correct classification rate gradually decreases, reaches its lowest in July and August (lower than 90%), and gradually increases thereafter.
In addition, the seawater classification has more discrete points between January and May, with some days having a correct rate as low as 70% or less. The overall average correct rates for sea ice and seawater are 99.591 9% and 95.120 7%.
The points above the line were classified as sea ice, and those below the line were classified as seawater. The comparison of the results of this study and the OSI-SAF is shown in Fig. 9.
The sea ice classification results of the multi-parameter linear segmentation were highly consistent with the OSI-SAF products. Similar to single-parameter threshold segmentation, the seawater classification accuracy begins to decrease in May and gradually increases after reaching its lowest value between July and August. The average correct rate of sea ice is 99.517 5%, and seawater is 95.507 8%.
In this study, the unsupervised separation K-means method was used to conduct an ice water separation test. The input parameters included the daily BT values of three bands and the B_coe values of RA. In order to understand which band has a greater impact on the classification, we chose four calculation options in which a total of four parameters were used. Firstly, all four parameters are used, and then one of the three BT bands is removed.
As shown in Fig. 10, the results of the K-means method are also in good agreement with the results of OSI-SAF ice product. However, missing data due to the impact of space particles on the satellite and the pre-processing process resulted in these discrete points appear before May. Although the sea ice matching results are poor for some discrete days throughout the year, the overall correct separation of sea ice and seawater are above 85%. Similar to the threshold segmentation, the correct classification rate of seawater decreases from May to September and reaches its lowest in July and August. However, K-means has a higher correct rate of seawater classification than single-threshold and multi-parameter linear segmentation during this time. Apart from this, results of K-means for sea ice are very different from the others in that it reaches its minimum correct classification result between late August and early September and remains largely above 95% correct for the rest of the year.
The scheme with the 18.7 GHz band removed had somewhat lower correct sea ice classification results than the other three schemes (a phenomenon evident from January to April); the 23.8 GHz as the water vapour band does not show a large difference in classification results after removal. The effect of the water vapour band on the classification results is limited. It can also be seen that the 18.7 GHz band has an important influence on the extraction of sea ice Table 1.
| Sea ice | Seawater | |
| K-means | 96.89% | 98.15% |
| K-means (remove 18) | 96.30% | 97.71% |
| K-means (remove 23) | 97.00% | 98.16% |
| K-means (remove 37) | 96.96% | 98.19% |
DownLoad:
CSV
The classification results of all three methods (single parameter threshold, multi-parameter linear segmentation methods, and K-means) differ significantly from OSI-SAF concentration products between May and September. The sea ice extent changed during this time and was also very regular (Fig. 11). This study uses the sea ice extent data from the National Snow and Ice Data Center (
The correct classification of seawater reached its lowest value in July when the rate of change in the sea ice area was greatest. The rapid melting of sea ice and melting of snow on the ice surface during that period reduced BT values of sea ice (Fig. 3). As BT values of sea ice and seawater were close, a large amount of seawater is classified as sea ice. Therefore, the results of the threshold segmentation and the multi-parameter linear segmentation using only the BT differ significantly from the OSI-SAF product.
Between the end of August and the beginning of October, an abrupt decrease in K-means sea ice classification results was formed. This is an important time when the rate of change in sea ice extent turns negative to positive, and sea ice extent basically reaches near its minimum value (Fig. 12). During this time, the new ice starts to form in large quantities. Because the sea ice constantly moves, the new ice is squeezed and deformed by the dynamics, forming a rough surface and reducing the B_coe. The K-means results are poorer, with only the BT threshold and two-parameter segmentation of the BT and B_coes having better results.
After September, the sea water starts to freeze, the ice properties become more stable, and the correct classification rate of the three sea ice and seawater classification methods gradually increases. The time when the seawater separation differs significantly from the OSI-SAF results is approximately the time of the inflection point at which the rate of change in sea-ice extent decreases, and the time when the sea-ice separation of K-means’ result differs significantly from the OSI-SAF product is approximately the time of the sea-ice extent minimum. The three methods also have the potential to determine the inflection point of sea ice change.
To find out where the classification differences are located and to improve the sea ice and seawater identification method, we chose to improve the results of the multi-parameter threshold method. On the one hand, a single threshold for both summer and winter only uses 18.7 GHz data for classification, which does not take advantage of the combined use of active and passive microwave data. On the other hand, although K-means has the advantage of classifying without specifying segmentation thresholds, from the perspective of satellite operation, the HY-2B satellite is subject to irretrievable data loss due to interference from space particles and so on. If data loss occurs, it will affect the data samples used for classification, which will affect the classification accuracy and make the algorithm less stable.
To differentiate the data used to determine the dividing line above and avoid modeling and verification of the data on the same day at the same time, July 15, which was also significantly different from the OSI-SAF results, was selected to illustrate the different points. As shown in Fig. 13, the points with different flags from the two sources are distributed almost on the sea ice edges.
For the combination of RA and CMR, there are three radiometer bands with wide footprints (24 km, 19 km and 10 km) and one radar altimeter observation with a footprint size of only 1.9 km. In the marginal region, RA provides signals from small size sea ice as it’s small footprint.
Therefore, the RA data have a small footprint along the orbit when just using 30% as the sea ice and seawater separation boundary and is thus not accurate. The resolution of the OSI-SAF ice product footprint is 10 km, while the footprint of the HY-2B RA is smaller than this. So, the HY-2B RA was able to see even smaller features due to its smaller footprint. The primary purpose of this paper is to provide a sufficiently accurate sea ice and seawater identification result for the RA data so the RA signals with small footprints can be used for subdivision in the ice edge area.
In a previous study, Jiang et al. successfully used RA waveforms to classify sea ice and seawater and obtained good classification results (Jiang et al., 2019). Therefore, the waveform parameters are suitable for distinguish the features around the edges of the sea ice.
First of all, the data from the boundary area of the sea ice need be extracted. We use an image morphology method to corrosion the sea ice edges. sea ice concentration of 30% and 0% of as the boundaries for the start of the expansion, respectively. Generally, the ice drift speed is 5–10 km per day (Tandon et al., 2018). So, a 5 × 5 square operator (50 km × 50 km) was used to ensure that the sea ice boundaries were contained within it (Fig. 14).
The parameters of RA waveforms (pulse peaking (PP)) were used as the distinguishing parameters in the ice edge areas. The PP values were first proposed by Laxon as a sea ice mapping technique (Laxon, 1994). Then, Jiang et al. improved the parameters applicability to the HY-2 RA satellite by ignoring the noise prone areas (Jiang et al., 2019). The waveform parameter PP values of the RA can be used as an important index for the classification of sea ice and seawater.
| $$ {\rm{PP}} = \frac{{\max}\left({{\rm{power}}}\right)}{\displaystyle\sum\limits _{21}^{108}{{\rm{power}}}} \times 88 . $$ | (4) |
The improved formula of the PP is shown in Eq. (4), where power denotes the power values of the altimeter echoes.
In order to verify the results combined with the multi-parameter threshold and RA waveform data, high-resolution SAR images were used to verify the classification results. Figure 11 shows the revised results for July 15, 2019. The full day data were selected every three months for mapping (Fig. 15). The data selected for these four different months and different places can represent the four seasons in the Arctic. The small figures in the fifth column in Fig. 15 show the OCOG and SMLE tracker flags distribution. The flags cannot provide a better sea ice and seawater classification than the SAR images.
All the data for the four selected days were for areas near the island (January 15) or near the sea ice edges (April 15, July 15, and October 15), and the single parameter threshold, multi-parameter linear segmentation, and K-means methods were all inaccurate in terms of the sea ice and water separation.
The results of the edge correction with waveform parameters are better than the results of these three methods. The situation on October 17 was similar to that on January 15, and the modified method exhibited a good recognition ability for small areas of seawater between discrete ice masses. On April 15, the data from different orbits crossed at the edge of the sea ice. Although the cross observations of the orbit were of the same object, except for the results of the correction method, the results of the other methods were biased at the intersection (i.e., two different classification results were obtained for the same object). As these are visually interpreted images, there is no data that can be sustained over time for quantitative comparison, so only four images have been selected for comparison.
For the first time, the CMR passive microwave BT and RA active microwave B_coe were used to classify the Arctic sea ice and water for the entire 2019 annual cycle.
We found that the threshold segmentation, linear segmentation, and K-means methods can effectively discriminate between Arctic sea ice and seawater during the cold season from early October until the end of April. Compared with OSI-SAF products, the average accuracy of sea ice-seawater discrimination can be 99.6%, 99.5%, and 96.8% for the threshold segmentation, linear segmentation, and K-means methods, respectively. During the melting season, the sea ice classification results of these methods matched well with the OSI-SAF products except those obtained by the K-means method. The accuracy of ice water discrimination stood just 90%, the difference largely occurred at the sea edge. While during the melting season, the accuracy of sea ice-seawater classification started worsening in May and then, gradually increased after reaching the low levels in July or August. Melting sea ice causes a gradual decrease in the BT value of the sea ice surface when the rate of sea ice melting reaches a maximum. The sea ice B_coe is affected to a greater extent when the sea ice melting rate changes from negative to positive when the primary ice grows in large quantities. The sea ice melting affected the calculation parameters, which subsequently affected the classification results of the sea ice and seawater.
However, the discrimination of sea ice and seawater near the ice edge cannot be carried out effectively by any of those methods due to the mixing of sea ice and seawater properties. This weakness can be tackled and solved by a new method using the radar waveform linear segmentation associated with image processing. This procedure provides a better correction of the sea ice-seawater identification errors close to the ice edge. Quantitative verification of the improvement is challenging. It is common for remote sensing data products to define the edge of ice water as having a sea ice density less than a certain threshold. In fact, there is sea ice present in areas below that threshold (Cavalieri et al., 1991; Gloersen, 1992). The presence of sea ice can be detected by the small footprint of RA. Such data above or equal to the resolution of RA footprint for the corresponding time period are not yet available in these regions. The results have been verified by SAR images of different time phase.
Comparing the results of the four classification methods with the OSI-SAF products has some potential to determine the tipping points of sea ice change. The maximum difference between the four seawater classifications and OSI-SAF results is approximately when the change in sea ice area decreases the most. The maximum difference between K-means sea ice classification and OSI-SAF is around the inflection point where sea ice grows.
Although identification of sea ice and seawater is not a new problem and various methodology have been implemented to tackle this problem (Gloersen, 1992; Remund and Long, 1999; Ochilov and Clausi, 2012). This work can be regarded as an alternative to retrieval sea ice and open water on the seasonal and annual cycle for the entire Arctic. A downside is the permanent data missing in the central Arctic north of 81°N, on the other hand, the large seasonal and annual sea ice-seawater variation occurred within the high-latitude zonal domain in the current state of the Arctic. In a long run, we would need additional data to better classification sea ice and water in the Arctic.
A high-resolution long term ice water identification regional product is still missing. In the near future, the launch of the ocean series satellites will provide us with more SAR data and longer-term Earth observations. Further development and utilization of these data sources will enhance the application of Chinese satellite data in polar sea ice, which is a sustainable long-term mission for us.
|
Aaboe S, Breivik L A, Eastwood S. 2021. Algorithm theoretical basis document for the global sea-ice edge and type product. OSI-403-C & EUMETSAT, 250: 1–34
|
|
Aagaard K, Carmack E C. 1994. The arctic ocean and climate: a perspective. In: Johannessen O M, Muench R D, Overland J E, eds. The Polar Oceans and Their Role in Shaping the Global Environment. Washington: American Geophysical Union, 5–20
|
|
Batoula S. 2011. ENVISAT ALTIMETRY Level 2 user manual V 1.4. Paris: European Space Agency
|
|
Bernard R, Le Cornec A, Eymard L, et al. 1993. The microwave radiometer aboard ERS-1.1. Characteristics and performances. IEEE Transactions on Geoscience and Remote Sensing, 31(6): 1186–1198. doi: 10.1109/36.317444
|
|
Bronner E, Guillot A, Picot N. 2013. SARAL/AltiKa Products Handbook: SALP-MU-M-OP-15984-CN. Toulouse: Centre National d’Etudes Spatiales
|
|
Cavalieri D J, Crawford J P, Drinkwater M R, et al. 1991. Aircraft active and passive microwave validation of sea ice concentration from the defense meteorological satellite program special sensor microwave imager. Journal of Geophysical Research: Oceans, 96(C12): 21989–22008. doi: 10.1029/91JC02335
|
|
Comiso J C. 1995. SSM/I sea ice concentrations using the bootstrap algorithm. Washington: National Aeronautics and Space Administration, 4–7
|
|
Eppler D T, Farmer L D, Lohanick A W, et al. 1992. Passive microwave signatures of sea ice. In: Carsey F D, ed. Microwave Remote Sensing of Sea Ice. Washington, DC: American Geophysical Union, 47–71
|
|
Fleming H E, Grody N C, Kratz E J. 1991. The forward problem and corrections for the SSM/T satellite microwave temperature sounder. IEEE Transactions on Geoscience and Remote Sensing, 29(4): 571–583. doi: 10.1109/36.135819
|
|
Forgy E W. 1965. Cluster analysis of multivariate data: efficiency versus interpretability of classifications. Biometrics, 21: 768–769
|
|
Fu L L, Christensen E J, Yamarone C A Jr, et al. 1994. TOPEX/POSEIDON mission overview. Oceanographic Literature Review, 99(C12): 24369–24381
|
|
Gill J P S, Yackel J J. 2012. Evaluation of C-band SAR polarimetric parameters for discrimination of first-year sea ice types. Canadian Journal of Remote Sensing, 38(3): 306–323. doi: 10.5589/m12-025
|
|
Gloersen P. 1992. Arctic and Antarctic Sea Ice, 1978–1987: Satellite Passive-Microwave Observations and Analysis. Washington: Scientific and Technical Information Program, National Aeronautics and Space Administration, 306
|
|
Jiang Xingwei, Lin Mingsen, Liu Jianqiang, et al. 2012. The HY-2 satellite and its preliminary assessment. International Journal of Digital Earth, 5(3): 266–281. doi: 10.1080/17538947.2012.658685
|
|
Jiang Chengfei, Lin Mingsen, Wei Hao. 2019. A study of the technology used to distinguish sea ice and seawater on the Haiyang-2A/B (HY-2A/B) altimeter data. Remote Sensing, 11(12): 1490. doi: 10.3390/rs11121490
|
|
Jiang Maofei, Xu Ke, Jia Yongjun, et al. 2015. Evaluation of HY-2B altimeter products over ocean. In: 2020 IEEE International Geoscience and Remote Sensing Symposium. Waikoloa, HI: IEEE, 5858–5861
|
|
Keihm S J, Janssen M A, Ruf C S. 1995. TOPEX/Poseidon microwave radiometer (TMR). III. Wet troposphere range correction algorithm and pre-launch error budget. IEEE Transactions on Geoscience and Remote Sensing, 33(1): 147–161. doi: 10.1109/36.368213
|
|
Kouraev A V, Crétaux J F, Lebedev S A, et al. 2011. Satellite altimetry applications in the Caspian Sea. In: Vignudelli S, Kostianoy A, Cipollini P, et al., eds. Coastal Altimetry. Berlin: Springer, 331–366
|
|
Kouraev A V, Papa F, Mognard N M, et al. 2004a. Sea ice cover in the Caspian and Aral Seas from historical and satellite data. Journal of Marine Systems, 47(1–4): 89–100
|
|
Kouraev A V, Papa F, Mognard N M, et al. 2004b. Synergy of active and passive satellite microwave data for the study of first-year sea ice in the Caspian and Aral seas. IEEE Transactions on Geoscience and Remote Sensing, 42(10): 2170–2176. doi: 10.1109/TGRS.2004.835307
|
|
Kouraev A V, Shimaraev M N, Buharizin P I, et al. 2008. Ice and snow cover of continental water bodies from simultaneous radar altimetry and radiometry observations. Surveys in Geophysics, 29(4): 271–295
|
|
Kouraev A V, Semovski S V, Shimaraev M N, et al. 2007. Observations of Lake Baikal ice from satellite altimetry and radiometry. Remote Sensing of Environment, 108(3): 240–253. doi: 10.1016/j.rse.2006.11.010
|
|
Laxon S. 1994. Sea ice altimeter processing scheme at the EODC. International Journal of Remote Sensing, 15(4): 915–924. doi: 10.1080/01431169408954124
|
|
Lomax A S, Lubin D, Whritner R H. 1995. The potential for interpreting total and multiyear ice concentrations in SSM/I 85.5 GHz imagery. Remote Sensing of Environment, 54(1): 13–26. doi: 10.1016/0034-4257(95)00082-C
|
|
Louet J, Bruzzi S. 1999. ENVISAT mission and system. In: IEEE 1999 International Geoscience and Remote Sensing Symposium. Hamburg: IEEE, 1680–1682
|
|
Markus T, Cavalieri D J. 2000. An enhancement of the NASA Team sea ice algorithm. IEEE Transactions on Geoscience and Remote Sensing, 38(3): 1387–1398. doi: 10.1109/36.843033
|
|
Obligis E, Eymard L, Tran N, et al. 2006. First three years of the microwave radiometer aboard Envisat: in-flight calibration, processing, and validation of the geophysical products. Journal of Atmospheric and Oceanic Technology, 23(6): 802–814. doi: 10.1175/JTECH1878.1
|
|
Ochilov S, Clausi D A. 2012. Operational SAR sea-ice image classification. IEEE Transactions on Geoscience and Remote Sensing, 50(11): 4397–4408. doi: 10.1109/TGRS.2012.2192278
|
|
Pol S L C, Ruf C S, Keihm S J. 1998. Improved 20- to 32-GHz atmospheric absorption model. Radio Science, 33(5): 1319–1333. doi: 10.1029/98RS01941
|
|
Przybylak R. 2007. Recent air-temperature changes in the Arctic. Annals of Glaciology, 46: 316–324. doi: 10.3189/172756407782871666
|
|
Remund Q P, Long D G. 1999. Sea ice extent mapping using Ku band scatterometer data. Journal of Geophysical Research: Oceans, 104(C5): 11515–11527. doi: 10.1029/98JC02373
|
|
Remund Q P, Long D G, Drinkwater M R. 2000. An iterative approach to multisensor sea ice classification. IEEE Transactions on Geoscience and Remote Sensing, 38(4): 1843–1856. doi: 10.1109/36.851768
|
|
Scheuchl B, Caves R, Cumming I, et al. 2001. Automated sea ice classification using spaceborne polarimetric SAR data. In: IEEE 2001 International Geoscience and Remote Sensing Symposium. Sydney: IEEE, 3117–3119
|
|
Screen J A, Simmonds I. 2010. The central role of diminishing sea ice in recent Arctic temperature amplification. Nature, 464(7293): 1334–1337. doi: 10.1038/nature09051
|
|
Shokr M, Sinha N K. 2015. Sea Ice: Physics and Remote Sensing. Washington, DC: American Geophysical Union, 615
|
|
Swan A M, Long D G. 2012. Multiyear Arctic sea ice classification using QuikSCAT. IEEE Transactions on Geoscience and Remote Sensing, 50(9): 3317–3326. doi: 10.1109/TGRS.2012.2184123
|
|
Tandon N F, Kushner P J, Docquier D, et al. 2018. Reassessing sea ice drift and its relationship to long-term Arctic sea ice loss in coupled climate models. Journal of Geophysical Research: Oceans, 123(6): 4338–4359. doi: 10.1029/2017JC013697
|
|
Woodhouse I H. 2017. Introduction to Microwave Remote Sensing. Boca Raton: CRC Press, 400
|
|
Xu Ke, Jiang Jingshan, Liu Heguang. 2007. A new tracker for ocean-land compatible radar altimeter. In: 2007 IEEE International Geoscience and Remote Sensing Symposium. Barcelona: IEEE, 3825–3828
|
|
Xu Ke, Jiang Jingshan, Liu Heguang. 2013. HY-2A radar altimeter design and in flight preliminary results. In: 2013 IEEE International Geoscience and Remote Sensing Symposium. Melbourne: IEEE, 1680–1683
|
|
Xu Ke, Liu Heguang, Liu Peng, et al. 2014. System techniques and capabilities of the HY-2A altimeter. In: 2014 IEEE Geoscience and Remote Sensing Symposium. Quebec City, QC: IEEE, 5171–5174
|
|
Xu Ke, Liu Peng, Tang Yueying, et al. 2017. The improved design for HY-2B radar altimeter. In: 2017 IEEE International Geoscience and Remote Sensing Symposium. Fort Worth, TX: IEEE, 534–537
|
|
Zhang Dehai, Wang Zhenzhan, Li Yun, et al. 2015. Preliminary analysis of HY-2 ACMR data. In: 2015 IEEE International Geoscience and Remote Sensing Symposium. Milan: IEEE, 177–180
|
|
Zhang Zhilun, Yu Yining, Li Xinqing, et al. 2019. Arctic sea ice classification using microwave scatterometer and radiometer data during 2002–2017. IEEE Transactions on Geoscience and Remote Sensing, 57(8): 5319–5328. doi: 10.1109/TGRS.2019.2898872
|
|
Zygmuntowska M, Khvorostovsky K, Helm V, et al. 2013. Waveform classification of airborne synthetic aperture radar altimeter over Arctic sea ice. The Cryosphere, 7(4): 1315–1324. doi: 10.5194/tc-7-1315-2013
|
Figures(15) / Tables(1)
Supported by:
Beijing Renhe Information Technology Co. Ltd
Chengfei Jiang, Mingsen Lin, Ruixue Cao, Hao Wei, Lijian Shi, Bin Cheng, Yongjun Jia, Qimao Wang. Classification of ice and water in the Arctic using radar altimeter and microwave radiometer data from HY-2B satellite[J]. Acta Oceanologica Sinica, 2023, 42(5): 179-191. doi: 10.1007/s13131-022-2067-4
| Sea ice | Seawater | |
| K-means | 96.89% | 98.15% |
| K-means (remove 18) | 96.30% | 97.71% |
| K-means (remove 23) | 97.00% | 98.16% |
| K-means (remove 37) | 96.96% | 98.19% |
DownLoad:
CSV
DownLoad:
DownLoad:
