You are on page 1of 17

This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING 1

Antarctic Sea-Ice Classification Based on Conditional


Random Fields From RADARSAT-2
Dual-Polarization Satellite Images
Tingting Zhu, Fei Li, Georg Heygster, Member, IEEE, and Shengkai Zhang

Abstract—In January 2014, Chinese National Antarctic Re- Index Terms—Antarctic, classification, conditional random
search Expedition (CHINARE) 30th cruise raised public concern fields (CRFs), multiple features, RADARSAT-2, sea ice, synthetic
since the Xuelong, the Chinese polar research vessel, was trapped aperture radar (SAR).
in the sea-ice zone (66°39 20.88 S, 144°25 2.28 E) in the vicin-
ity of the Adélie Depression area on the east Antarctic continent.
This event highlighted the importance of an operational sea-ice I. INTRODUCTION
classification map for ice routing to serve ship navigation. In this
INCE the large mass of sea ice in the southern and north-
paper, unprecedented Antarctic sea-ice classification algorithms
from RADARSAT-2 satellite dual-polarization synthetic aperture
radar (SAR) images were developed using the conditional random
S ern hemisphere introduces the exchange of mass, heat, and
momentum between the ocean and atmosphere [1], [2], the pos-
fields (CRF) approach by including multiple features from sea-ice itive feedback role that sea ice plays in the global climate sys-
concentration, gray-level cooccurrence matrix textures, polariza- tem results in increasing interest in sea-ice monitoring. Remote
tion ratio, backscatter coefficients, and intensity data. Coincident
RADARSAT-2 Satellite SAR datasets with five scenes were col-
satellite sensors can provide continuous measurements for early
lected for ice classification into categories such as open water, thin detection of sea-ice formation and melting in remote locations
ice, smooth first year ice, deformed first year ice, and old ice during and can monitor extreme climate conditions [3], [4]. The sig-
the CHINARE-30th cruise. The effects of deformation, rafting, and natures of different sea-ice types can be captured by synthetic
ridging during the spring–summer transition period were over- aperture radar (SAR) based on its all-weather and frequent re-
whelmed by the spatial and contextual CRF models in combina-
tion with the rich features extracted for sea-ice classification. Four
peat observation capabilities, because microwave can penetrate
strategies including statistical distribution and region connection, thick clouds and darkness in Polar regions, indicating that SAR
multiple features and support vector machine (SVM) integrated may be the most effective method for monitoring large-scale
into the CRF model are proposed to describe the sea-ice-type re- changes in sea ice [5].
lationships among pixels. By conducting comparative experiments Sea ice comprises a number of different ice types including
between the proposed methods and state-of-the-art sea-ice classifi-
cation based on the SVM algorithm, the best was obtained from the
open water (OW), thin ice (TI), smooth first year ice (SFY),
SVM-based CRF (SVM-CRF) algorithm for sea-ice classification deformed first year ice (DFY), and old ice (OI). Radar imagery
with respect to the three scenes from the Indian Ocean sector and generates surface roughness information by the microwave sig-
two scenes of Pacific Ocean sectors including medium-resolution natures and sea-ice dielectric properties, making it possible to
dual-polarization SAR imagery with a pixel spacing of 50 m and distinguish with SAR data different ice types. Specifically, C-
higher resolution dual-polarization SAR imagery with a pixel spac-
ing of 6.25 m. Results indicate that the SVM-CRF approach has
band SAR imagery provides good contrast between OW and ice
the capacity for improving sea-ice classification, which can provide categories. The multipolarimetric (full- and dual-polarization)
accurate and reliable sea-ice class information for sea-ice analysis. SAR imagery also offers improvements in OW, TI, and thick ice
discrimination, compared to single-polarization imagery [6]–
[8]. Dual-polarization data with HH and HV polarization can
Manuscript received September 30, 2015; revised January 24, 2016 and significantly enhance the interpretation accuracy and reduce the
March 25, 2016; accepted March 26, 2016. This work was supported in part ambiguities in sea-ice classification [9], [10].
by the State Key Program of National Natural Science of China under Grant
41531069, in part by the National Basic Research Program of China under Nevertheless, the disadvantages of SAR images such as
Grant 2012CB957701, and in part by the Chinese Polar Environment Compre- speckle noise, high interclass and intraclass backscatter vari-
hensive Investigation and Assessment Programs under Grant CHINARE2016. ability degrade the image details and decrease sea-ice class sep-
(Corresponding author: Fei Li)
T. Zhu is with the State Key Laboratory of Information Engineering in Survey- arability [11]. In order to overcome these limitations of SAR data
ing, Mapping and Remote Sensing, Wuhan University, Wuhan 430079, China for improving sea-ice analysis, early studies used SAR backscat-
(e-mail: zhutingting62008@163.com). ter characteristics and the robust classifier combined with texture
F. Li is with the State Key Laboratory of Information Engineering in Survey-
ing, Mapping and Remote Sensing, the Chinese Antarctic Center of Surveying features for discrimination between ice and water in the Arc-
and Mapping, and also with the Collaborative Innovation Center of Geospatial tic. Local binary patterns (LBP) are exploited in sea-ice analysis
Technology, Wuhan University, Wuhan 430079, China (e-mail: fli@whu.edu.cn) [12]. Frequently used texture features are the gray-level cooccur-
G. Heygster is with the Institute of Environmental Physics, University of
Bremen, Bremen 28359, Germany (e-mail: heygster@uni-bremen.de). rence matrix (GLCM) textures for SAR-based sea-ice analysis
S. Zhang is with the Chinese Antarctic Center of Surveying and Mapping, [13]. In sea-ice classification, GLCM textures are widely fed into
Wuhan University, Wuhan 430079, China (e-mail: zskai@whu.edu.cn). the robust classifiers such as artificial neural network and sup-
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. port vector machine (SVM) characterized by its nonlinear func-
Digital Object Identifier 10.1109/JSTARS.2016.2551318 tion [14]–[16]. Pulse-coupled neural networks able to model the
1939-1404 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

spatial and contextual information are utilized by Karvonen [17] where a spring/summer snow layer acts as a protective buffer,
for sea-ice segmentation and classification. Recently, Markov retarding sea-ice surface melt [30].
random field (MRF) has been widely used in sea-ice segmenta- These aforementioned differences suggest that several differ-
tion and classification [15], [18]–[20] since the MRF framework ent effects trigger the variability of the observed radar backscat-
models the contextual information in labels and also shows its ter values of different ice types within the Antarctic sea-ice
efficiency in sea-ice classification. However, using the MRF to zone. Therefore, we used multiple features to characterize the
model the contextual information in observed data is difficult ice appearance under complex sea-ice conditions. In this pa-
since the observed vectors are assumed to be conditionally in- per, multiple features including the backscatter, textures, polar-
dependent. Conditional random fields (CRFs) have the ability ization ratio, intensity data, and sea-ice concentration features
to capture the contextual information in both labels and ob- were incorporated into the CRF model for sea-ice classifica-
served data. CRFs [21], which directly models the posterior as tion. During the spring–summer transition period, deformation
Gibbs distribution and thus allows one to capture the depen- is common and can be resolved from the presented imagery.
dence of the observed data, are applied for sea-ice classifica- Deformation may lead to different forms including rafting and
tion in this paper. In recent research, CRF has been introduced ridging, openings in consolidated ice, and breaking of thicker
in ice–water analysis [11], [22], adopting the multimodality ice. However, dynamic processes such as opening and closing of
Gaussian model as a unary classifier. Convolutional neural net- ice significantly influence the backscatter and lead to amounts
work (CNN) has the robust features of learning capability via of subpixel floes, which are challenging for sea-ice classifi-
multilayer neural networks derived from deep learning frame- cation. Since sea-ice concentration can represent opening and
work. CNN-SFCRF [22] integrated the robust features into the closing dynamic characteristics, we introduced sea-ice concen-
CRF framework for sea-ice concentration retrieval and CNN- tration into the features group. Textures information was also
SFCRF has demonstrated improvement in ice–water bound- extracted by GLCM and a polarization ratio was used for sea-ice
aries. In this paper, multiple features have been exploited and classification.
integrated into the CRF framework. In addition, for the purpose The purpose of our study is to explore the operational sea-
of optimization, the category prior derived from SVM classifi- ice classification algorithm for ship navigation of the China R/V
cation has been introduced in SVM-CRF for improving sea-ice Xuelong in the future. This icebreaker was trapped in the sea-ice
classification. zone (66°39 20.88 S, 144°25 2.28 E) during the 30th expedi-
These earlier studies focused on sea-ice classification in the tion in Antarctic. Unprecedented Antarctic sea-ice classification
Arctic during ice growth. However, limited attention was paid to from RADARSAT-2 dual-polarization images was constructed
investigate the relationships between sea ice types and backscat- on the CRF model and multiple features integrated approaches
ter from dual-polarization radar imagery in the Antarctic. The for a robust sea-ice classification algorithm. The study area
maxima extent of Antarctic sea ice is 20% greater than that in and SAR data are introduced in Section II. The methodology
the Arctic [2], [23]. Nonetheless, early ice–water discrimination is described in Section III. Section IV presents the classifica-
research was focused on the Arctic. Also, sea ice cover spatial tion results and the evaluation. The conclusion is presented in
and temporal variability within the sea-ice evolution process Section V.
differs from the Antarctic to the Arctic [24], [25]. The appear-
ance of ice and water shows differences under different ambient II. STUDY AREA AND SATELLITE DATASETS
sea-ice conditions with respect to diverse dynamic and ther-
modynamic interaction between ocean and atmosphere, thereby A. Study Area
differing in sea-ice physical and microwave properties [26], The two study areas are located in the vicinity of the East
[27]. The Antarctic continent is surrounded by sea ice and OW, Antarctic sea zone including the Indian Ocean sector S1 (65°E to
while the land masses extrude into the Southern Ocean. Arctic 85°E) and the Western Pacific Ocean sector S2 (130°E to 150°E).
sea ice generally grows under quiescent ocean conditions. In The first study area is around Zhongshan station (69°22 24 S,
contrast, Antarctic sea ice grows under turbulent ocean surface 76°22 40 E) near Prydz Bay and the Amery Ice shelf. The sec-
conditions, stimulating deformations into rafting and ridging, ond study area is located around the Mertz Glacier in the Adélie
and ice motion in response to external oceanic and atmospheric Depression, where the R/V Xuelong was trapped in the sea-
forcing. Another eliciting factor of backscatter variation is the ice zone (66°39 20.88 S, 144°25 2.28 E). Navigation is limited
seawater flooding of the snow–ice interface when the load of to the summer shipping season when the ice is weaker be-
snow cover is heavy enough to depress the ice sheet below cause of melt and decay and some areas of OW are present.
its freeboard. This process is ubiquitous in the Antarctic. The On November 22, 2013, the R/V Xuelong started its cruise
counterpart in the Arctic is generally the thick multiyear ice [23], from the Zhongshan station in order to run across the Antarctic
[28], [29], which has survived at least one melt season. Defor- continent. The average temperature by shipboard underway ob-
mation and moistened snow pack subject to flooding of seawater servation along the track was recorded below zero. Coincident
can fundamentally influence the sea-ice scattering characteris- RADARSAT-2 Satellite SAR datasets were collected through
tics. In the Antarctic, surface melting is not common and rarely the R/V Xuelong cruise along east Antarctica from the start until
associated with the presence of melt water on the surface of January 7, 2014. In terms of sea-ice dynamics and thermody-
the ice, when compared to the Arctic. Additionally, the ablation namic forcing [26], Antarctic sea ice cover follows the seasonal
season duration in Antarctica is shorter lived than in the Arctic, cycle. Thus, multiyear ice is less common than in the Arctic.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZHU et al.: ANTARCTIC SEA-ICE CLASSIFICATION BASED ON CONDITIONAL RANDOM FIELDS 3

TABLE I
SEA SURFACE CATEGORIES USED FOR THE SAR SEA-ICE
CLASSIFICATION IN THIS PAPER

SAR satellite RADARSAT-2 over Antarctica around the Amery


shelf and the Mertz Glacier. The original goal of the commer-
cial RADARSAT-1 and the follow-up mission RADARSAT-2
SAR satellites missions, launched in 1995 and 2007, was sea-ice
monitoring. Each scene in Regions A and C with a pixel spac-
ing of 50 m was captured in the ScanSAR wide beam mode,
which provides dual-polarization images in HH (transmitting
horizontally and receiving horizontally polarized signals) and
Fig. 1. Locations of the study regions with red frame showing atmospheric HV (transmitting horizontally and receiving vertically polar-
temperature along the track of the 30th expedition of R/V Xuelong as a blue ized signals) polarizations. Region D (see Table II) is taken
line. in HH mode only. Each scene with a pixel spacing of 6.25 m
was captured in the fine beam mode type, offering HH and HV
During the spring–summer transition period (November– polarizations as presented in Table II. Table II also lists the de-
January), the sea ice will also retreat from the marginal ice tailed acquisition time, coordinate, and image size of each scene.
zone to the sea-ice pack cover along the coast of the Antarc- Table II shows the series of five RADARSAT-2 images acquired
tic continent over Sector_1 and Sector_2, except for the con- over the east Antarctic regions during the early summers of
solidated oldest land-fast ice. By virtue of passive microwave 2013 and 2014, which were used for sea-ice classification. Be-
observations, ice growth typically starts in late March and con- cause these are the most suitable overlapping remote sensing
tinues throughout the Antarctic winter until early melt begins datasets that exist for the Prydz Bay and Adélie Depression,
in November and December. During the melt season, sea ice we restricted the description of the two sectors to the study ar-
begins to decay and it drifts into warmer waters. In addition, eas. Fig. 2 illustrates the locations of the study areas with the
during this period, ice concentration can decrease significantly browse images. Large-scale (a pixel spacing of 50 m resolu-
to values lower than 50%. The sea ice can even retreat to the tion) and small-scale (a pixel spacing of 6.25 m) SAR datasets
coast in one month at the end of January. Until mid-March, the are combined in Section II-A for the analysis of complex sea-
sea water will refreeze since air temperatures begin to drop; ice conditions. The combining of these two kinds of resolution
then, new ice is produced in some OW areas, and finally sea datasets should enable improved understanding of sea-ice clas-
ice advances to the marginal zone through the process of sea- sification, which is affected by the strong inter- and intraregional
ice growth.Table I describes sea-ice classes corresponding to differences noted in SAR images.
the development of stages, according to the Antarctic Sea-Ice The images in Table II were taken at the onset of melt during
Processes and Climate protocol for bridge-based sea-ice obser- the seasonal transition from spring to summer. At this time,
vations and World Meteorological Organization [31]–[33]. The various sea-ice types in the region can be seen as shown in
study region hosts a diversity of ice regimes composed of OW, Fig. 2. The icebreaker R/V Xuelong was trapped in the thick
TI, SFY, DFY, and OI types. The OI includes multiyear ice, ice shown in Region E of Fig. 2. RADARSAT-2 imagery also
icebergs, land-fast ice, and ice shelves. The corresponding color shows the locations of Regions A to E. Regions B and E are the
code is also presented in Table I together with a broad range of Prydz Bay and Adélie Depression, respectively.
sea-ice types. Pixel-based sea-ice classification is different from
the egg code symbols [33], since sea-ice classification maps can III. METHODOLOGY
provide accurate location information for different sea-ice types
CRFs have successfully integrated contextual information
during ship navigation, particularly in the complex sea-ice con-
into remote sensing classification [34]–[38]. Several state-of-
ditions of the Antarctic.
the-art methods under CRFs model have been investigated by
extending the potentials with multiscale region connection cal-
B. Satellite Datasets Description
culus (RCC) for SAR image classification [36], modeling the
In this experiment, the image datasets used for algorithm scattering statistics of SAR for image segmentation [37], and in-
development consist of five scenes from the 5.3-GHz-C-band troducing the class boundary constraint for hyperspectral image
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

TABLE II
LIST OF RADARSAT-2 DUAL-POLARIZATION SATELLITE IMAGES USED FOR SEA-ICE CLASSIFICATION

Scene center

Sector Scene ID Acquisition Time Lon Lat Polarization Pixel spacing (meter) Beam/mode type Image size (pixel)

S1 1.Region A 2013-11-22 75.228, −66.938 HH HV 50 ScanSAR wide 10 693∗ 10 190


2.Region B 2013-11-28 76.542, −69.161 HH HV 6.25 Fine 8766∗ 8515
3.Region C 2013-11-29 75.921, −66.746 HH HV 50 ScanSAR wide 10 695∗ 10 190
S2 4.Region D 2014-1-6 143.411, −65.882 HH 50 ScanSAR wide 10 723∗ 10 242
5.Region E 2014-1-7 144.661, −66.609 HH HV 6.25 Fine 8762∗ 8506

classification [38]. In this paper, multiple features are extracted where w is the model parameter. δ(xi , l) verifies δ(xi , l) = 1
as the observation in the CRF. Considering the sea-ice label for xi = l, and δ(xi , l) = 0 for xi = l. P (xi |φ(y)) represents
interaction based on the stages of sea-ice development, mul- the probability of site i taking the label xi conditioned on the
tiple features are extracted and then introduced into the CRF stacked feature vector φ(y) and can be obtained by multiclass
framework. Furthermore, the SVM classification label image is classifier SVM and logistic classifier [41], [42]. This paper uti-
introduced as a category prior to optimize a final labeled image. lizes the multiclass classifier SVM to construct the unary poten-
The SVM-based CRF (SVM-CRF) developed for operational tials. The texture features and sea-ice concentration estimate is
sea-ice classification is presented in Section III-A. useful for reflecting spatial variation of sea ice [16], [43], [44].
From the literature review of different CRF models, two ad- Then, multiple features are exploited for sea-ice classification,
ditional strategies based on CRF are described in Sections III-B which includes sea-ice concentration estimate, GLCM textures,
and III-C. Under the CRF framework, Section III-B presents a polarization ratio, backscatter values, and intensity data.
statistical distribution method based on CRF (STA-CRF) [37], The pairwise potential provides the interaction strength from
which exploits the SAR statistical property for SAR sea-ice clas- a pair of sites (i, j) and is defined as
sification. Additionally, the contextual information described by 
the RCC reasoning approach [36], [39] is also used for sea-ice fiunary (xi |y) = δ(xi , l)logP (xi |φ(y), w) (3)
classification, which is presented in Section III-C. A schematic l∈C

representation of SVM-CRF sea-ice classification flowchart is where v = {vl,l  , l, l ∈ C} is the model parameter for concate-
presented in Fig. 4 and described in Section III-D. nating the feature vector μij (y), and μij (y) is a feature vector
{φi (y), φj (y)} at a pair of sites (i, j). The pairwise feature vec-
A. SVM-Based CRF tor μij (y) at sites (i, j) is described by their differences in cap-
turing the contextual information. If neighboring regions have
CRFs model the posterior distribution over labels X given the
similar image features, fijpairwise favors the same class label for
observations Y [21]. Y = {yi }i∈S is the observed field from an
them. A pairwise term could enhance the classification results
input image, where S is a set of sites in the input image, yi is
since it could model the relationships among the neighborhood
the observed data of site i ∈ S = {1, 2, . . . , M } and M is the
pixels. Especially in SAR images with intense speckle noise, the
total number of superpixels of the input image. The superpixels
pairwise term is significant in smoothing the classification re-
are obtained by a mean shift algorithm [40]. Let X = {xi }i∈S
sults and reducing the speckle noise by considering the context
be the category labels of the homogenous sea-ice types, where
information in the CRF framework.
xi takes value in the set C = {1, 2, . . . , l . . . , L}, and L is the
The pseudolikelihood algorithm is used to learn the CRF
number of sea-ice types. The standard CRF has the following:
model parameters. After obtaining the posterior probability,

1  unary sea-ice classification can correspond to finding the maximum
P (x|y) = exp fi (xi |y) posteriori of the label image. A maximum a posteriori (MAP)
ZCRF
i∈S [45] estimator is used for inferring the optimal configuration
⎤ for sea-ice labels. A Bayesian MAP defines the minimiza-
  pairwise
+ fij (xi , xj |y)⎦ (1) tion energy that can maximize the posterior probability as
i∈S j ∈N i
E(x|y) = −logP (x|y) − logZCRF and can be represented as
follows:
where ZCRF is the partition function for normalization, and xi
XM AP = argmaxP (x|y) = argminE(x|y). (4)
is the category label of site i. Ni is the set of neighborhood
of site i. fiunary (·) and fijpairwise (·) are the unary and pairwise Based on (1)–(3), the posteriori probability can be
potentials of site i, respectively. The unary potential contributes expressed as
the information from a single site i of observations and can be ⎛ ⎞
defined as   pairwise
 P (x|y) ∝ exp ⎝fi
unary
(xi , yi ) fij (xi , xj )⎠.
fiunary (xi |y) = δ(xi , l)logP (xi |φ(y), w) (2) i∈S j ∈N i

l∈C (5)
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZHU et al.: ANTARCTIC SEA-ICE CLASSIFICATION BASED ON CONDITIONAL RANDOM FIELDS 5

Fig. 2. SAR imagery locations from Regions A (left) to E (right).

are used as the observation to construct the unary and pairwise


potential for every pixel in the superpixel using the formula (2)
and (3). Then, MAP is implemented to get the final classifica-
tion result for each pixel in the inconsistent superpixel. Finally,
the information of the superpixel and pixel level can be inte-
grated using the framework illustrated in Fig. 3. The category
prior is introduced in order to avoid the potential misclassified
superpixel, which tends to make the result worse in the label-
ing smooth method. The architecture of the SVM-CRF method
implementation is described in Section III-D and presented in
Fig. 4.

B. Statistical Distribution Based CRF


Statistical distribution models, e.g., Gamma, Weibull, Log-
Fig. 3. Flowchart of the procedure to recalculate the pixel-wise posterior normal, and Alpha-stable [47]–[50], have been applied to model
probability.
the heavy-tail and sharp-peaked statistical properties of SAR im-
agery backscatter values. In this paper, statistical distributions
The solution of MAP relies on the max-product belief propa- are exploited to be integrated into the CRF framework to dis-
gation algorithm [46] in the inference procedure for this factor- criminate the sea-ice classes from RADARSAT-2 SAR images.
ized form. Finally, the initial label prediction is obtained by the Let ϕ(y) denote SAR intensity data. We have
MAP algorithm. The initial label prediction is compared with 
the category prior. The category prior is derived from the im- P (ϕ(y)|x) = P (ϕ(yi )|xi ). (6)
proved SVM classification result, described in Section III-D. If i∈S

the initial prediction label of the superpixel does not match with The distribution of ϕ(y) is an independent condition to x.
the corresponding category prior, this superpixel is regarded Then, aiming at the incorporation with SAR statistical features
as the inconsistent superpixel. The inconsistent superpixels are for sea-ice classification, the STA-CRF can be defined as
required to recalculate the posterior probability according to the 
following flowchart shown in Fig. 3. 1  unary
P (x|y, ϕ(y)) = exp fi (xi |y)
In Fig. 3, S is the label of superpixel from the initial prediction, ZSTA−CRF
i∈S
and S  is label of the corresponding pixels from the category 
prior obtained by the improved SVM classification result. The + log P (ϕ(y)|xi )
superpixel label is not sufficient to represent the overall prefer- i∈S
ence of pixels, leading to the misclassification. In this flowchart, ⎤
the coherence is measured by the comparison between S and S  
+ fijpairwise (xi , xj |y)⎦ (7)
to find the inconsistent superpixels. The inconsistent superpixel i∈S j ∈N i
is required to recalculate the pixel-wise posterior probability in
the superpixel. The corresponding category prior provided the where ZSTA−CRF is the partition function for the STA-CRF
label information to initialize the graph. Then multiple features model, and y is the observed data, which also used the intensity
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

Fig. 4. SVM-CRF method for dual-polarization RADARSAT-2 imagery sea-ice classification.

data in this paper. log P (ϕ(y)|x) can be modeled by adopting where k denotes the kth iteration. y is the observed data includ-
the Alpha-stable distribution [50] to obtain the distribution pa- ing the polarization ratio, and the intensity data at HH and HV
rameters of different ice types. The unary potential function polarization. The unary potential function and pairwise poten-
adopts formula (2) and the pairwise potential function adopts tial function is represented in formula (2) and (3), respectively.
formula (3). P (ϕ(y)|x) modeled by the following Alpha-stable The RCC model is used to establish the region connection re-
distribution, is expressed as lationships between the k − 1th classification results and the
kth classification results. The k − 1th RCC model is the kth
exp{jμϕ − |γϕ|α [1 − jsign(ϕ)β tan(πα/2)]}, α = 1 prior RCC potential. The region connection information can be
q(ϕ) =
exp{jμϕ − |γϕ|[1 + jsign(ϕ)β π2 ln(|ϕ|)]}, α=1 captured by fiRCC , which can be defined as
(8) 
K
where sign is the sign function. {α, β, γ, μ} are the parameters f RCC xk |y = (ωk ,k −1 · Bin) (10)
of the alpha-stable distribution. α is the characteristic exponent, su =1
and β is the skewed parameter. γ is the dispersion parameter, where su is the superpixel region obtained by the mean shift
and μ stands for the location parameter. Training sample of algorithm and K is the number of superpixel regions at k − 1th
each class represents the realization of the alpha-stable distri- iteration. Bin = {bEC , bPO , bTPP , bEQ , bPO } is the histogram
bution. The distribution parameters can be estimated by using a with 5 × K bins for describing the five RCC relationships in-
regression-type method [51]. The alpha-stable parameters cor- cluding externally connected (EC), partially overlapping (PO),
responding to each class are represented using formula (8). With tangential proper part (TPP), nontangential proper part (NTPP),
the alpha-stable parametric distribution, the statistical features and equal (EQ) relationship. ωk ,k −1 is the iteration potential
can be integrated into the CRF framework. Then, the STA-CRF weight between the previous iteration xk −1 and the current it-
model is able to capture the statistical information from the SAR eration xk . For example, if the i th superpixel xki −1 at k − 1th
images. iteration and the ith superpixel xki at kth iteration have the
relationship of PO, then bPO = δ(xki −1 , xki ) = 1. Otherwise,
C. Region Connection-Based CRF (REG-CRF) bPO = δ(xki −1 , xki ) = 0. The REG-CRF model integrated with
the RCC iteration potential for the SAR sea-ice classification
The spatial logic relationships among superpixels are useful
has the advantage of smoothing classification results. Then, the
for interpreting the complex SAR images. In [36], a multiscale
rich spatial and contextual information can also be captured by
RCC model [39] is exploited to reason out the spatial relation-
the spatial reasoning approach RCC model using the iteration
ships under the CRF framework. In the REG-CRF scheme, the
potential in the CRF framework.
RCC potential is constructed by describing the superpixels re-
lationship between iterations. In order to integrate the unary
potential, pairwise potential and iteration region connection po- D. SVM-Based CRF Method Implementation
tential into the CRF framework, the posterior probability is As in Section III-A, the methodology of the SVM-CRF
reformulated as method is demonstrated. In this implementation section, a
 schematic representation of SVM-based CRF architecture is
k 1  unary illustrated in Fig. 4. For sea-ice classification, the original
P x |y = exp fi (xi |y)
ZREG−CRF RADARSAT-2 Dual-PolSAR intensity data first takes the 22°
i∈S
  pairwise incidence angle from the RADARSAT-2 product file as the refer-
+ fij (xi , xj |y) ence for calculating the normalized backscatter coefficients σ0 at
i∈S j ∈N i HH and HV polarization, respectively. Intensity data, backscat-
 ter, GLCM textures, and polarization ratio are the inputs of

+ fiRCC xki −1 |y (9) the SVM classifier for the initial SVM sea-ice classification
i∈S result for sea-ice concentration estimate in order to construct
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZHU et al.: ANTARCTIC SEA-ICE CLASSIFICATION BASED ON CONDITIONAL RANDOM FIELDS 7

the multiple features group. After construction of multiple fea- as sea-ice concentration of the center pixel. Sea-ice con-
tures, including intensity data, backscatter, polarization ratio, centration is a supervised feature derived from the initial
GLCM textures, and sea-ice concentration estimate, the im- SVM classification result, which will be described below.
proved classification process was conducted on multiple fea- The initial SVM classification and the improved SVM clas-
tures using SVM as the classifier. In this flowchart shown in sification are required for the SVM-CRF algorithm. We need to
Fig. 4, the initial SVM classification was obtained for sea-ice describe the setup for the classifiers of SVM and SVM-CRF.
concentration estimate and the improved SVM classification 1) Implementation of SVM Classification: The initial SVM
was used as the category prior. classification result was obtained by using above backscat-
Multiple features were also used in the CRF framework as ter coefficients, intensity data, GLCM textures, and po-
the feature vectors. Based on the unary potential and pairwise larization ratio as inputs of the SVM classifier. Sea-ice
potential, the posterior probability of CRF was obtained. Then, concentration estimate was retrieved on the initial SVM
MAP is used for predicting the initial label in SAR imagery. The classification result. Then, multiple features including
initial prediction label is required to compare with the category backscatter, polarization ratio, intensity data, GLCM tex-
prior, and then, MAP optimization process is implemented for tures, and sea-ice concentration estimate was yielded for
the final classification result. the improved SVM sea-ice classification result. The im-
Since the features are important for SVM-CRF-based sea-ice proved SVM classification result was used as the category
classification, it has been described in detail as follows: prior of the SVM-CRF algorithm. We employed the ENVI
1) Backscatter Normalization: Backscattering coefficient software v.4.8 for the initial SVM classification and the
is the basis for extracting features including GLCM improved SVM classification. In the SVM process, the ra-
textures, polarization ratio, and sea-ice concentration dial basis function (RBF) was used as the kernel function.
estimate. Backscatter normalization first takes a fixed inci- The training samples including 500 pixels of each class
dence angle of 22° from the RADARSAT-2 product file as for SVM classifier were selected from the visual reference
the reference incidence angle. The normalized backscatter classification in each scene.
coefficient σ0 at HH and HV polarization was calculated 2) SVM-CRF Setup: Ten training samples of each class in
using the linear function σ0 = 0.1553(x0 − 22) + y0 , re- each scene were selected with a window of 64 × 64 pixels
spectively [16]. x0 is the incidence angle of each pixel, for CRF model training. All the testing methods utilize
and y0 is the decibel value from the original data product the same number of training samples per class in each
files. scene. Given the training data, a multiclass SVM classifier
2) GLCM Textures: Textures are responsible for reflecting with RBF kernel was selected to calculate the multiclass
spatial variations of different sea-ice types. GLCM tex- probability in the CRF framework.
tures are useful for sea-ice classification. In this paper, 3) Training and Testing Samples: The visually homogeneous
GLCM textures, including the mean, variance, homogene- sea-ice surfaces were manually labeled as the visual ref-
ity, and correlation, have been calculated for sea-ice anal- erence classification illustrated in the experimental sec-
ysis. All GLCM textures were computed on both HH and tion. We implemented the visual interpretation followed
HV backscatter images. The work in [16] discussed the [52] and [53], which reported the backscatter range of
best window parameter for these GLCM textures. The each ice type and described the characteristic of different
window size defined as 49 × 49 pixels and the sliding ice types. Training samples have two sets including one
window size with 8 pixels were adopted in this paper. set for training initial SVM classification and improved
The gray-scale quantization level was set as 32. GLCM SVM classification with limited number set as 500 pix-
textures were implemented using the existing software En- els per class in each scene. Another set, with ten samples
vironment for Visualizing Images (ENVI software v.4.8). for each ice type for the CRF model, was selected and
3) Polarized Ratio: Polarization ratio can be used for distin- each sample was selected with a window of 64 × 64 pix-
guishing sea-ice types because they have different char- els. Two training sets were selected separately from the
acteristics at each polarization. A polarization ratio was visual reference classification. Additional testing sam-
generated as the ratio of backscatter at HV and HH polar- ples were selected from the remaining visual reference
σHV classification.
ization as PR = σ 0H H . The backscatter have been obtained
0
by the first normalized backscatter coefficients step.
4) Sea-Ice Concentration Estimation: In complex sea-ice
IV. SEA-ICE CLASSIFICATION RESULTS AND DISCUSSION
conditions, sea-ice concentration is introduced for rep-
resenting sea-ice evolution spatial patterns, which is a Classification with the proposed algorithm was carried out
useful feature for the discrimination between OW and for OW, TI, SFY, DFY, and OI, based on RADARSAT-2 dual-
consolidated ice. Sea-ice concentration is the percentage polarization SAR imagery. These images were collected to pro-
of ice within a given area, which uses the defined win- vide two scales of sea-ice information. The experiments were
dow size as its given area. The initial SVM classification constructed on five scenes including three large-scale scenes
result is used to calculate sea-ice concentration with the with pixel spacing of 50-m resolution and two small-scale scenes
defined window size 25 × 25. Within the defined window, with pixel spacing of 6.25-m resolution, and one scene with only
the percentage of the ice area was calculated and assigned HH polarization. The large-scale scenes show various ice types
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

Fig. 5. RADARSAT-2 HH scene example with the subset of the original Region C. Colored bullets indicate the ice type according to Table I. The upper left
shows the test dataset location in Region C as the yellow frame in Fig. 5.

and OW. The small-scale scenes are more important for scien- Sea-ice classification result of FEA-SVM is coarse and the vi-
tists onboard the R/V Xuelong for their landing and trapping sual performance of STA-CRF is moderate, shown in Fig. 6(a)
activities. Both large- and small-scale scenes were analyzed and (b), respectively. REG-CRF classification approach is ac-
using the four CRF-based strategies for Antarctic sea-ice classi- ceptable, which can provide smooth results with less sensitiv-
fication, comprising the following Sections IV-B and IV-C. The ity to speckle noise, as presented in Fig. 6(c). The REG-CRF
four strategies were implemented based on CRF models, includ- method has an advantage by maintaining the integrity of every
ing the STA-CRF method, REG-CRF method, multiple features- type and improves the visualization of the classification results.
based CRF (FEA-CRF), and SVM-CRF method. FEA-CRF was In SVM-CRF, the leads boundary details can be well preserved
used to demonstrate the role of category prior in SVM-CRF. We as can be seen in Fig. 6(d).
named the SVM-CRF without category prior as the multiple 2) Visual Performance of SVM-CRF Method Using Different
FEA-CRF. The comparative algorithm multiple features-based Training Samples: The classification results are sensitive to the
SVM (FEA-SVM) was implemented in LIBSVM to assess the numbers of training samples. Therefore, visual performance of
performance of the proposed algorithm. Since large-scale scenes SVM-CRF using different numbers of training samples is also
cover a variety of ice types in Sectors 1 and 2, it is not easy to presented. As shown in Fig. 7, the classification performance is
show ice class details. Therefore, in Section IV-A, we take an improved as the number of training samples increases. The clas-
example from the section in Region C to illustrate the detailed sification result with three samples of each class for SVM-CRF
visual performance of these testing algorithms. In addition to shows the worst performance in Fig. 7(a). With five training
visual inspection and comparative analysis, quantitative mea- samples, the performance has an improvement in DFY and TI.
sures of the classification accuracy were calculated in order to With respect to the visual performance of 10 and 15 training
evaluate the performance of the proposed methods, presented samples, their classification results are similar. In the follow-
in Sections IV-B and IV-C. The overall accuracy (OA) of our ing experiments, the SVM-CRF method adopted ten training
proposed SVM-CRF results is of the order of 92% and the samples for implementing sea-ice classification.
kappa coefficient approximates 0.9. In Section IV-D, the results 3) Visual Performance of FEA-CRF and SVM-CRF: To
discussion and analysis of different algorithms is presented. demonstrate the role of category prior in the SVM-CRF method,
the classification result of SVM-CRF without category prior
A. Testing Datasets for Comparative Visual Analysis (FEA-CRF) is presented in Fig. 8. From the visual inspection
on Fig. 8, we can see that SVM-CRF is significantly1 better than
In Fig. 5, the section of the SAR image in Region C (shown FEA-CRF. With good preservation of the ice boundary details,
in the upper left of Fig. 5 as the yellow frame) on November SVM-CRF provides better visualization. Furthermore, wrongly
29, 2013, covers the full set of sea-ice types and OW in the classified pixels were corrected by SVM-CRF. This clearly in-
vicinity of the Amery ice shelf near Zhongshan Station. The dicates that category prior is essential in the SVM-CRF frame-
color bullets were manually delineated by visual interpretation work, and also confirms the effectiveness of the category prior
for illustrating different ice types. for preserving the class details.
1) Visual Performance of Four Algorithms: In order to verify
the visual performance of testing methods, their sea-ice classi-
fication results are presented in Fig. 6. The comparative method B. Classification Implemented on Large-Scale Scenes
by FEA-SVM misclassified TI as OW, as seen in the right upper The following experiments used three large-scale SAR
corner of Fig. 6(a), mainly due to the high penetration depth datasets to assess the performance of the proposed algo-
of the radar ray, which reduces the contrast between them. rithm. As an example, the HH polarization imagery with the
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZHU et al.: ANTARCTIC SEA-ICE CLASSIFICATION BASED ON CONDITIONAL RANDOM FIELDS 9

Fig. 6. Example results on comparative algorithms demonstrated on subset of Region C. For the color codes of the sea surface categories see Table I.
(a) FEA-SVM. (b) STA-CRF. (c) REG-CRF. (d) SVM-CRF.

Fig. 7. Example results on SVM-CRF algorithm using different numbers of training samples demonstrated on subset of Region C, (a) 3, (b) 5, (c) 10, and (d) 15
(n per class). For the color codes of the sea surface categories see Table I. (a) n = 3 (b) n = 5 (c) n = 10 (d) n = 15.

Fig. 8. Example results of FEA-CRF and SVM-CRF algorithm demonstrated on subset of Region C. For the color codes of the sea surface categories see Table I.
(a) FEA-CRF. (b) SVM-CRF.

resolution of 50 m for each large scale is presented in Fig. 9(a)– significant mixture of ice and water can also be observed along
(c). Their corresponding visual reference classification is dis- the ice edge, which may lead to classification inaccuracies.
played in Fig. 9(d)–(f) with the color image. The polynya in In order to quantitatively analyze the effectiveness of the pro-
Regions A and C is covered with dispersed TI, as can be seen posed methods, the visual reference classification of each scene
near the bottom of the original data in Fig. 9(a) and (b). The was used to calculate the OA and kappa statistic. Quantitative
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

Fig. 9. HH polarization imagery with a pixel spacing of 50 m for illustrating the large-scale scenes in Region A, C, and D. The color image is their visual
reference classification for each region. For the color codes of the sea surface categories, see Table I. (a) Region A. (b) Region C. (c) Region D. (d) Region A
visual reference classification. (e) Region C visual reference classification. (f) Region D visual reference classification.
TABLE III
REGION A SEA-ICE CLASSIFICATION ACCURACY

Algorithm Class OW/% TI/% SFY/% DFY/% OI/% OA/% Kappa

FEA-SVM OW 91.92 7.79 0.19 0.02 0.08 89.90 0.839


TI 10.67 86.55 2.75 0.03 0.00
SFY 0.00 0.01 99.92 0.06 0.01
DFY 0.00 0.00 2.14 92.96 4.90
OI 0.14 1.33 6.01 1.05 91.47
STA-CRF OW 93.68 6.26 0.02 0.00 0.04 82.58 0.796
TI 15.07 84.89 0.02 0.02 0.00
SFY 0.01 20.49 79.28 0.22 0.00
DFY 0.00 0.24 11.44 86.44 1.88
OI 0.32 4.88 12.78 2.12 79.90
REG-CRF OW 94.12 5.81 0.03 0.00 0.04 85.03 0.821
TI 10.44 89.54 0.00 0.02 0.00
SFY 0.00 22.15 77.83 0.02 0.00
DFY 0.00 0.00 18.41 80.11 1.48
OI 0.28 4.77 8.27 1.42 85.26
FEA-CRF OW 91.83 8.11 0.01 0.04 0.01 86.66 0.838
TI 11.16 88.14 0.70 0.00 0.00
SFY 0.01 11.55 88.44 0.00 0.00
DFY 0.00 1.42 26.82 71.76 0.00
OI 0.45 2.72 2.92 0.78 93.13
SVM-CRF OW 92.95 6.98 0.00 0.07 0.00 90.54 0.872
TI 15.96 83.91 0.10 0.03 0.00
SFY 0.00 1.67 98.13 0.20 0.00
DFY 0.00 0.00 3.40 96.60 0.00
OI 0.11 3.02 2.39 0.83 93.65

Surface class acronyms see Table I.

assessments are provided by the accuracy reports for each type distinction of ice from water in Fig. 10(b). By virtue of visual
calculated in the confusion matrices in Tables III–V, below. inspection, the smoothest result is provided by the REG-CRF
1) Region A Sea-Ice Classification and Evaluation: The dis- method and it can also provide clear boundaries of multiyear
advantage of FEA-SVM is the misclassification of the TI as OW ice located in the middle of Region A. STA-CRF and REG-
along the ice edge and in the polynya. STA-CRF shows robust CRF are deficient in SFY classification, which has been largely
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZHU et al.: ANTARCTIC SEA-ICE CLASSIFICATION BASED ON CONDITIONAL RANDOM FIELDS 11

TABLE IV
REGION C SEA-ICE CLASSIFICATION ACCURACY

Algorithm Class OW/% TI/% SFY/% DFY/% OI/% OA/% Kappa

FEA-SVM OW 96.30 3.36 0.19 0.00 0.15 92.04 0.902


TI 7.51 92.49 0.00 0.00 0.00
SFY 0.08 0.06 97.15 2.61 0.10
DFY 0.14 0.00 17.75 81.19 0.92
OI 0.00 0.00 1.93 0.53 97.54
STA-CRF OW 94.06 5.72 0.01 0.16 0.05 88.98 0.855
TI 10.31 89.68 0.00 0.01 0.00
SFY 0.15 4.49 94.65 0.66 0.05
DFY 0.24 2.90 7.16 89.16 0.54
OI 0.00 0.16 1.94 16.72 81.18
REG-CRF OW 94.78 4.99 0.01 0.16 0.06 91.41 0.882
TI 7.78 92.21 0.00 0.01 0.00
SFY 0.42 2.92 95.21 1.45 0.00
DFY 0.19 1.71 0.35 97.58 0.17
OI 0.00 0.11 1.82 16.71 81.36
FEA-CRF OW 94.18 5.73 0.00 0.02 0.07 88.85 0.843
TI 8.48 91.52 0.00 0.00 0.00
SFY 0.17 33.82 65.99 0.00 0.02
DFY 0.22 1.97 1.88 95.11 0.82
OI 0.00 0.46 0.11 3.72 95.71
SVM-CRF OW 95.53 4.45 0.00 0.00 0.02 94.13 0.919
TI 6.72 92.21 0.51 0.26 0.30
SFY 0.15 2.32 97.5 0.00 0.03
DFY 0.19 2.11 1.35 96.35 0.00
OI 0.00 0.10 0.86 3.06 95.98

Surface class acronyms see Table I.

TABLE V
REGION D SEA-ICE CLASSIFICATION ACCURACY

Algorithm Class OW/% TI/% SFY/% DFY/% OI/% OA/% Kappa

FEA-SVM OW 98.92 0.55 0.17 0.29 0.07 82.11 0.760


TI 12.38 65.48 19.58 2.51 0.05
SFY 0.00 0.07 75.73 24.20 0.00
DFY 0.00 0.00 3.26 96.42 0.32
OI 0.00 13.40 0.00 0.06 86.54
STA-CRF OW 95.54 4.33 0.03 0.06 0.04 89.54 0.861
TI 8.83 83.51 7.17 0.40 0.09
SFY 0.00 0.05 87.29 12.64 0.02
DFY 0.02 0.09 11.31 88.52 0.06
OI 0.01 0.00 0.07 4.59 95.33
REG-CRF OW 95.29 4.70 0.01 0.00 0.00 90.50 0.873
TI 7.61 73.43 14.60 3.83 0.53
SFY 0.00 0.03 87.70 12.27 0.00
DFY 0.00 0.10 1.50 98.39 0.01
OI 0.00 0.00 0.00 5.82 94.18
FEA-CRF OW 93.91 5.98 0.00 0.11 0.00 89.25 0.870
TI 5.57 92.13 1.69 0.29 0.32
SFY 0.00 2.54 92.45 5.01 0.00
DFY 0.00 0.18 20.10 76.92 2.80
OI 0.00 0.01 0.07 15.72 84.20
SVM-CRF OW 97.93 1.59 0.02 0.45 0.01 92.32 0.901
TI 0.92 91.09 3.45 4.51 0.03
SFY 0.00 0.69 94.49 4.81 0.01
DFY 0.00 0.16 10.08 89.75 0.01
OI 0.00 0.00 0.06 14.66 85.28

Surface class acronyms see Table I.

misclassified as TI. The CRF-based methods provide the same is 0.839 less than 0.872 given by the SVM-CRF method. FEA-
good separability between TI and OW presented in Table III. In CRF could not detect the complete iceberg in the center of the
Table III, the SVM-CRF method provides about 96% correct- images. Among the four CRF strategies, the SVM-CRF per-
ness for distinguishing the SFY and DFY ice, and outperforms formed the best in terms of visual inspection and accuracy. The
STA-CRF and REG-CRF. The kappa of the FEA-SVM method sea-ice classification accuracy table (see Table III) indicates
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

Fig. 10. Sea-ice classification results in Region A. For the color codes of the sea surface categories, see Table I. (a) FEA-SVM. (b) STA-CRF. (c) REG-CRF.
(d) FEA-CRF. (e) SVM-CRF.

that the largest classification error is TI, misclassified as OW by highest accuracy. In terms of visual performance, the SVM-CRF
FEA-SVM, which is obvious in the subset of Region C shown method retrieved the best classification results in Region D by
in Fig. 6(a). The confusion matrices in Table III also show that preserving details from each sea-surface-type boundary. From
the CRF-based methods offer good performance in ice–water Table V, we conclude that the CRF model has confirmed the ef-
discrimination. The combined effects of snow cover, sea water fectiveness of the CRF-based methods since it can incorporate
flooding, and deformations observed in Fig. 5 are overwhelmed the spatial and contextual information for classification. When
by the spatial and contextual CRF models. In combination with in combination with category prior, the SVM-CRF method
the rich features extracted for sea-ice classification, the visual- leads to better classification results by modeling the spatial and
ization of sea-ice classification is improved. With category prior, contextual information under CRF framework when compared
SVM-CRF sea-ice classification was much improved compared to the FEA-SVM.
to FEA-CRF. Specifically, SVM-CRF classification approach
can give completely consistent results with the visual inspec-
tion in large-scale scenes experiments. C. Classification Implemented on Small-Scale Scenes
2) Region C Sea-Ice Classification and Evaluation: In this In this section, we present the classification results based
testing area, the FEA-SVM result is still less sensitive to TI, on the small-scale RADARSAT-2 imagery of the two sectors
failing to distinguish TI from OW. The CRF-based approaches in Prydz Bay and the Adélie Depression, respectively. The
can provide good boundary details between TI and OW [see high-resolution RADARSAT-2 imagery with a pixel spacing of
Fig. 11(b)–(e)]. But in the right bottom of image, the West 6.25 m was utilized to analyze sea-ice conditions. The countless
ice shelf, STA-CRF and REG-CRF misclassified OI as DFY small OI floes can be resolved in these fine resolution images.
by 16%, as shown in Table IV. FEA-SVM, FEA-CRF, and Original datasets and corresponding visual reference classifi-
SVM-CRF can provide accurate classification result for the West cation in the small-scale regions of B and E are presented in
ice shelf. This may be due to the contribution of the multiple Fig. 13. The HH polarization imagery of Regions B and E as
features for improving the classification. In the image center, the presented in Fig. 13 indicate that high resolution data does not
iceberg can be detected with whole object by FEA-SVM and necessarily mean good interpretation since the intense speckle
SVM-CRF. The misclassified SFY pixels by FEA-CRF have noise may hinder the sea-ice interpretation.
been corrected by SVM-CRF, which indicates the effectiveness In Fig. 13(a) and (c), three sea-ice types can be observed in-
of category prior in the CRF framework. Compared with the cluding TI, SFY, and OI. The small-scale scenes can provide
testing algorithms classification results in Fig. 11, the best visual high-resolution imagery for each sea surface type with fine de-
performance is observed in Fig. 11(e) given by the SVM-CRF tails. The small-scale scene is relatively simpler than the large
classification method. Table IV also shows that the best accuracy scale, which hosts various sea-ice types. Small-scale deforma-
is available by SVM-CRF with 0.919 kappa and 94.13% OA. tion like ridging (thin bright lines) can be observed in the upper
3) Region D Sea-Ice Classification and Evaluation: Fig. 1 right corner of Fig. 13(c). The deformation leads to an increase
shows that the real-time temperature in the Adélie Depression in the surface roughness, which may make the signature much
was higher than that in the Prydz Bay area. Around the Adélie more complex in intra- and intervariability of the classes. The
Depression, first year ice covered a major area of Region D and testing algorithms were also performed on these two small-
TI was less than in Region A. Fig. 12(a) shows the same defi- scale scenes in Regions B and E. The classification results im-
ciency in poor discrimination between OW and TI as presented plemented by the testing algorithms are presented in Figs. 14
in Figs. 10(a) and 11(a) by the FEA-SVM method. By visual and 15.
inspection, OI has been largely misclassified by all the testing 1) Region B Sea-Ice Classification and Evaluation: From
algorithms. The worst performance in separating OI from DFY visual inspection, the sea-ice classification result given by the
is the SVM-CRF method by 14% errors shown in Table V. FEA-SVM algorithm is most sensitive to speckle noise. How-
The second misclassification error is induced by STA-CRF in ever, OI classification correctness by FEA-SVM is the highest.
classifying the SFY as DFY. Benefiting from the good perfor- Fig. 14 shows that the CRF-based methods are less sensitive to
mance in TI classification, the SVM-CRF method provides the noise, which again demonstrates that CRF-based strategies are
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZHU et al.: ANTARCTIC SEA-ICE CLASSIFICATION BASED ON CONDITIONAL RANDOM FIELDS 13

Fig. 11. Sea-ice classification results in Region C. For the color codes of the sea surface categories, see Table I. (a) FEA-SVM. (b) STA-CRF. (c) REG-CRF.
(d) FEA-CRF. (e) SVM-CRF.

Fig. 12. Sea-ice classification results in Region D. For the color codes of the sea surface categories, see Table I. (a) FEA-SVM. (b) STA-CRF. (c) REG-CRF.
(d) FEA-CRF. (e) SVM-CRF.

Fig. 13. HH polarization imagery with a pixel spacing of 6.25 m for illustrating the small-scale scenes in Regions B and E (a) (c). The color image is their visual
reference classification for each region. For the color codes of the sea surface categories, see Table I. (a) Region B. (b) Region B visual reference classification.
(c) Region E. (d) Region E visual reference classification.

Fig. 14. Sea-ice classification results in Region B. For the color codes of the sea surface categories, see Table I. (a) FEA-SVM (b) STA-CRF (c) REG-CRF
(d) FEA-CRF (e) SVM-CRF.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

14 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

TABLE VI CRF-based strategies have better antinoise capability than the


REGION B SEA-ICE CLASSIFICATION ACCURACY
FEA-SVM algorithms.

Algorithm Class TI/% SFY/% OI/% OA/% Kappa


D. Analysis and Discussion
FEA-SVM TI 82.35 16.46 1.19 85.75 0.783
SFY 4.42 86.97 8.61 In this paper, we have attempted to exploit CRF for modeling
OI 2.43 8.55 89.02
STA-CRF TI 96.63 3.36 0.01 87.64 0.824
contextual information in sea-ice classification using four CRF-
SFY 4.57 95.43 0.00 based strategies including STA-CRF, REG-CRF, FEA-CRF, and
OI 5.07 16.38 78.55 SVM-CRF. The comparative experiment of FEA-SVM was also
REG-CRF TI 96.66 3.34 0.00 89.71 0.843
SFY 2.88 96.02 1.10
conducted in order to highlight the advantages of the proposed
OI 4.05 12.12 83.83 algorithm. The performance of the proposed approach on five
FEA-CRF TI 98.06 1.94 0.00 83.26 0.793 regions demonstrated its efficacy for sea-ice classification in
SFY 19.79 80.21 0.00
OI 23.42 0.03 76.55
Sections IV-A, IV-B, and IV-C. In Fig. 16, the comparisons of
SVM-CRF TI 96.34 3.63 0.03 91.54 0.878 the testing methods in different regions, their sensitivity using
SFY 2.74 97.17 0.09 different numbers of training samples, and their consumption
OI 5.44 8.06 86.50
times are also demonstrated for analysis.
Surface class acronyms see Table I. 1) Accuracy Comparison of Different Algorithms: In order to
compare the performance of the testing algorithms in different
regions and scales, Fig. 16 is presented to show their accuracies
effective in modeling contextual information for sea-ice clas- by OA and kappa. FEA-SVM shows unstable performance on
sification. The best antinoise algorithm is REG-CRF, which different regions. The SVM-CRF algorithm achieved the highest
can provide a much smoother classification map than the other OA and kappa on each scene. With category prior, SVM-CRF
four methods. However, the tiny OI near the right in Region B offers a significant increment in OA and kappa compared to
was misclassified by STA-CRF and REG-CRF as TI. FEA-CRF FEA-CRF. In all five scenes, we observed that STA-CRF pro-
shows the worst performance in distinguishing the OI from TI vided the acceptable accuracy in terms of OA and kappa. In
in the center of this testing area. The OI pixels omitted by FEA- addition, REG-CRF’s accuracy is higher than STA-CRF. This
CRF have been corrected by SVM-CRF due to the introduction may be due to the effectiveness of iteration potential in mod-
of category prior. The best performance for describing the ice eling contextual information. As discussed in the visual perfor-
floes is obtained by the SVM-CRF algorithm, which preserved mance and accuracy report in Sections IV-A, IV-B and IV-C,
many more details when compared to the FEA-SVM algorithm. the SVM-CRF outperforms the other algorithms, which further
Table VI shows the Region B classification accuracy of the five demonstrates its validity in sea-ice classification.
algorithms and indicates that OA of SVM-CRF is highest among 2) Time Consumption of Different Algorithms: Table VIII
the testing algorithms. The largest misclassification is caused by summarizes the computational time for five sea-ice classifi-
the FEA-CRF algorithm, which leads to a 23.42% error in dis- cation algorithms on each test scene. The overall execution
tinguishing the OI from the TI. The larger misclassification by time takes the features and model training time into conside-
STA-CRF leads to a 16.38% error by wrongly classifying DFY ration. The model training time cost is shown in parenthe-
as TI. REG-CRF performed better in small-scale scenes than sis. Among the CRF strategies, REG-CRF used the least time
in large-scale scenes with 89.71% OA. Nonetheless, the SVM- for classification on large-scale scenes, followed by the STA-
CRF algorithm achieved the best result in terms of the accuracy CRF. However, SVM-CRF needs more time for the features
from OA by 91.54% and kappa by 0.878 (see Table VI). calculation. The proposed SVM-CRF method used consider-
2) Region E Sea-Ice Classification and Evaluation: Some able amounts of time in computation and will be studied in our
OI pixels were misclassified by all the testing algorithms. The future work. FEA-SVM utilizes LIBSVM in MATLAB 2011a
largest misclassification can be seen in Fig. 15(a) by the FEA- platform. STA-CRF is run in MATLAB 2011a platform. REG-
SVM algorithm, which failed to discriminate the SFY from CRF, FEA-CRF, and SVM-CRF algorithms are run in Visual
the tiny OI, with a 27.42% error as reported in Table VII. Studio 2010 platform. The features are calculated by ENVI
The subpixel OI floes are a challenge for sea-ice classifica- v.4.8 software. All testing algorithms shown in Section IV are
tion since their small size can be discerned as speckle noise. accomplished by a computer with an Intel Core CPU at 2.4 GHz
In Table VII, STA-CRF and SVM-CRF provided preferable and 48.00 GB of RAM.
performance in OI classification. In terms of the accuracy ta- 3) Sensitivity Analysis of Different Algorithms Using Different
ble and visual inspection for region E, the SVM-CRF clas- Numbers of Training Samples: Table IX presents an analysis of
sification provided the best performance among the testing the sensitivity of different algorithms using different training
algorithms. samples. The OA and kappa of different algorithms using dif-
To summarize, the CRF-based algorithms improve the visual ferent numbers of training samples were calculated. Because
performance of classification results by preserving the details the testing of datasets is time consuming due to their large
of the ice boundary. In terms of OA and kappa, the highest sizes, we only took two regions A and B for sensitivity analysis
accuracies in two small-scale regions are achieved by the SVM- in different numbers of training samples. With a limited num-
CRF algorithm. By visual inspection of small-scale scenes, the ber of training samples, OA and kappa of REG-CRF shows its
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZHU et al.: ANTARCTIC SEA-ICE CLASSIFICATION BASED ON CONDITIONAL RANDOM FIELDS 15

Fig. 15. Sea-ice classification results in Region E. For the color codes of the sea surface categories, see Table I. (a) FEA-SVM. (b) STA-CRF. (c) REG-CRF. (d)
FEA-CRF. (e) SVM-CRF.

TABLE VII
REGION E SEA-ICE CLASSIFICATION ACCURACY

Algorithm Class TI/% SFY/% OI/% OA/% Kappa

FEA-SVM TI 94.85 4.95 0.20 91.04 0.870


SFY 3.27 96.72 0.01
OI 1.56 27.42 71.02
STA-CRF TI 94.97 4.92 0.11 91.32 0.875
SFY 6.44 93.56 0.00
OI 2.01 17.89 80.10
REG-CRF TI 95.48 4.03 0.49 92.88 0.889
SFY 5.67 94.33 0.00
OI 7.06 12.31 80.63
FEA-CRF TI 94.34 5.41 0.25 92.30 0.872
SFY 4.18 95.82 0.00
OI 2.22 23.36 74.42
SVM-CRF TI 95.71 4.27 0.02 94.54 0.905
SFY 4.66 95.34 0.00
OI 1.98 12.94 85.08

Surface class acronyms see Table I.

TABLE VIII
COMPUTATION TIMES (IN HOURS) FOR DIFFERENT METHODS

Method Region A Region B Region C Region D Region E

FEA-SVM 3.2(1.2) 2.6(0.8) 3.1(1.1) 1.5(0.7) 2.0(0.6)


STA-CRF 2.0 1.3 2.0 1.0 1.0
REG-CRF 1.7 1.1 1.8 1.1 1.2
FEA-CRF 4.0(2.0) 3.2(1.4) 4.1(2.1) 2.0(1.2) 3.1(1.7)
SVM-CRF 5.5(2.3) 4.4(1.8) 5.4(2.3) 3.0(1.5) 4.0(2.0)

TABLE IX
OA(%) AND KAPPA OF DIFFERENT NUMBERS OF TRAINING SAMPLES BY Fig. 16. Accuracy of FEA-SVM, STA-CRF, REG-CRF, FEA-CRF, and SVM-
DIFFERENT METHODS FOR REGIONS A AND B CRF algorithms by OA and kappa coefficient. (a) OA. (b) Kappa coefficient.

Labeled samples per class


potential in sea-ice classification compared to FEA-SVM. Us-
Region Method 3 5 10 15 ing different numbers of training samples, STA-CRF shows the
A FEA-SVM 66.15 (0.584) 78.03 (0.667) 88.57 (0.855) 91.45 (0.894)
lowest accuracy among the five algorithms in Regions A and B.
STA-CRF 63.57 (0.531) 71.83 (0.606) 82.58 (0.796) 88.08 (0.852) However, the SVM-CRF shows the highest OA and kappa using
REG-CRF 72.87 (0.618) 80.46 (0.715) 85.03 (0.821) 85.19 (0.828) different numbers of training samples. The proposed SVM-CRF
FEA-CRF 71.87 (0.607) 81.25 (0.720) 86.66 (0.838) 89.24 (0.868)
SVM-CRF 75.83 (0.623) 82.26 (0.729) 90.54 (0.872) 92.61 (0.912)
method demonstrates superior performance compared with the
B FEA-SVM 74.61 (0.633) 81.04 (0.711) 85.75 (0.783) 92.32 (0.905) other four approaches, indicating robust performance.
STA-CRF 70.45 (0.625) 81.47 (0.804) 87.64 (0.8235) 89.98 (0.861)
REG-CRF 77.48 (0.722) 83.47 (0.818) 89.71 (0.8427) 91.27 (0.903)
FEA-CRF 73.9 (0.697) 80.12 (0.768) 83.26 (0.793) 88.47 (0.852) V. CONCLUSION
SVM-CRF 78.54 (0.737) 85.47 (0.831) 91.54 (0.8776) 93.73 (0.910)
We investigated the RADARSAT-2 signatures for sea-ice-
type classification on the dual-polarization SAR imagery issued
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

16 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING

during the R/V Xuelong trapping event. We propose using the [6] E. Rignot and M. Drinkwater, “Winter sea-ice mapping from multiparam-
CRF-based algorithms for providing sea-ice classification maps. eter synthetic-aperture radar data,” J. Glaciol., vol. 40, no. 134, pp. 31–45,
1994.
The performance was assessed by comparing results with those [7] B. Scheuchl, D. Flett, R. Caves, and I. Cumming, “Potential of
of state-of-the-art algorithms on large-scale scenes with medium RADARSAT-2 data for operational sea ice monitoring,” Can. J. Remote
resolution and on small-scale scenes with high resolution. The Sens., vol. 30, no. 3, pp. 448–461, Jun. 2004.
[8] K. Partington, J. D. Flach, D. Barber, D. Isleifson, P. J. Meadows, and
evaluation results from both scales are consistent. The CRF- P. Verlaan, “Dual-polarization C-band radar observations of sea ice in
based algorithms, including STA-CRF, REG-CRF, FEA-CRF, the amundsen gulf,” IEEE Trans. Geosci. Remote Sens., vol. 48, no. 6,
and SVM-CRF, are given and compared with the FEA-SVM pp. 2685–2691, May 2010.
[9] R. De Abreu, D. Flett, B. Scheuchl, and B. Ramsay, “Operational sea ice
algorithm. Although multiple features are useful for improv- monitoring with RADARSAT-2—A glimpse into the future,” in Proc. Int.
ing sea-ice classification results, only the algorithm SVM-CRF Geosci. Remote Sens. Symp., 2003, pp. 1308–1310.
provides the best performance in terms of visual inspection and [10] P. Yu, A. Qin, and D. A. Clausi, “Feature extraction of dual-pol SAR
imagery for sea ice image segmentation,” Can. J. Remote Sens., vol. 38,
accuracy. Compared with FEA-CRF, the category prior intro- pp. 352–366, 2012.
duced into SVM-CRF demonstrates its efficacy in correcting [11] F. Li, D. Clausi, L. Wang, and L. Xu, “A semi-supervised approach for
the misclassified pixels. We found that the SVM-CRF algo- Ice-water classification using dual-polarization SAR satellite imagery,”
in Proc. IEEE Conf. Comput. Vision Pattern Recognit. Workshops, 2015,
rithms can greatly improve the classification results by integrat- pp. 28–35.
ing category prior from the improved SVM classification into [12] J. Karvonen, B. Cheng, T. Vihma, M. Arkett, and T. Carrieres, “A method
the CRF models. The proposed algorithm performs well in dis- for sea ice thickness and concentration analysis based on SAR data and a
thermodynamic model,” Cryosphere, vol. 6, pp. 1507–1526, 2012.
criminating between the TI and OW, which is difficult because [13] L. K. Soh and C. Tsatsoulis, “Texture analysis of SAR sea ice imagery
the high penetration of the radar reduces their contrast. Sea-ice using gray level co-occurrence matrices,” IEEE Trans. Geosci. Remote
conditions during the spring–summer transition highlights the Sens., vol. 37, no. 2, pp. 780–795, Mar. 1999.
[14] A. V. Bogdanov, S. Sandven, O. M. Johannessen, V. Y. Alexandrov,
difficulty of sea-ice classification associated with the high vari- and L. P. Bobylev, “Multisensor approach to automated classification
ability in backscatter. Among the four CRF-based strategies, of sea ice image data,” IEEE Trans. Geosci. Remote Sens., vol. 43, no. 7,
the SVM-CRF approach performs the best and gives spatially pp. 1648–1664, Jul. 2005.
[15] S. Leigh, Z. Wang, and D. Clausi, “Automated ice–water classification us-
and temporally consistent results when compared to the visual ing dual polarization SAR satellite imagery,” IEEE Trans. Geosci. Remote
interpretation of the different SAR scenes. Snow cover and ab- Sens., vol. 52, no. 9, pp. 5529–5539, Sep. 2014.
lation backscatter characteristics during sea-ice growth were [16] H. Liu, H. Guo, and L. Zhang, “SVM-based sea ice classification using tex-
tural features and concentration from RADARSAT-2 Dual-Pol ScanSAR
not investigated in this study. However, as snow thickness is an data,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 8, no. 4,
important parameter for modifying sea-ice backscatter charac- pp. 1601–1613, Apr. 2015.
teristics, the influence of snow on sea ice as it affects backscatter [17] J. Karvonen, “Baltic sea ice SAR segmentation and classification using
modified pulse-coupled neural networks,” IEEE Trans. Geosci. Remote
needs to be investigated in future studies by extending the tem- Sens., vol. 42, no. 7, pp. 1566–1574, Jul. 2004.
poral series analysis on SAR scenes. This future research will [18] Q. Yu and D. A. Clausi, “IRGS: Image segmentation using edge penalties
also contribute to improve our understanding of sea-ice growth and region growing,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 30,
no. 12, pp. 2126–2139, Dec. 2008.
processes, which may benefit accurate prediction of sea-ice dis- [19] H. Deng and D. Clausi, “Unsupervised segmentation of synthetic aperture
tribution and extent combined with remote sensing, meteoro- radar sea ice imagery using a novel Markov random field model,” IEEE
logical, and climate data. Trans. Geosci. Remote Sens., vol. 43, no. 3, pp. 528–538, Mar. 2005.
[20] S. Ochilov and D. Clausi, “Operational SAR sea-ice image classifica-
tion,” IEEE Trans. Geosci. Remote Sens., vol. 50, no. 11, pp. 4397–4408,
Nov. 2012.
ACKNOWLEDGMENT [21] J. Lafferty, A. McCallum, and F. C. Pereira, “Conditional random fields:
Probabilistic models for segmenting and labeling sequence data,” in Proc.
The authors would like to thank Polar Research Institute of Int. Conf. Mach. Learn., 2001, pp. 282–289.
China and Chinese National Arctic and Antarctic Data Center [22] L. Wang, A. Wong, K. A. Scott, D. A. Clausi, L. L. Xu,
for providing the RADARSAT-2 images and temperature data. M. Shafiee, and F. Li. (2015). Sea ice concentration estimation
from satellite SAR imagery using convolutional neural network and
They would also like to thank the associate editor and six anony- stochastic fully connected conditional random field [Online]. Available:
mous reviewers for the constructive comments and suggestions http://vip.uwaterloo.ca/files/publications/
on the revision of this paper. [23] R. Galley, E. Key, D. Barber, B. Hwang, and J. Ehn, “Spatial and temporal
variability of sea ice in the southern Beaufort sea and Amundsen gulf:
1980–2004,” J. Geophys. Res., vol. 113, 2008, Art. no. C05S95.
REFERENCES [24] I. Simmonds, “Comparing and contrasting the behaviour of Arctic and
Antarctic sea ice over the 35 year period 1979–2013,” Ann. Glaciol,
[1] G. A. Maykut, “Energy exchange over young sea ice in the central Arctic,” vol. 56, pp. 18–28, 2015.
J. Geophys. Res., vol. 83, no. C7, pp. 3646–3658, 1978. [25] M. Drinkwater and X. Liu, “Seasonal to interannual variability in Antarctic
[2] G. A. Maykut, “Large-scale heat exchange and ice production,” J. Geo- sea-ice surface melt,” IEEE Trans. Geosci. Remote Sens., vol. 38, no. 4,
phys. Res., vol. 87, pp. 7971–7984, 1982. pp. 1827–1842, Jul. 2000.
[3] J. Yackel and D. Barber, “Melt ponds on sea ice in the Canadian [26] S. W. Laxon, N. Peacock, and D. Smith, “High interannual variability
Archipelago: 2. On the use of RADARSAT-1 synthetic aperture radar of sea ice thickness in the Arctic region,” Nature, vol. 425, no. 6961,
for geophysical inversion,” J. Geophys. Res., vol. 105, pp. 22061–22070, pp. 947–949, Oct. 2003.
2000. [27] A. P. Worby and J. C. Comiso, “Studies of the Antarctic sea ice edge and
[4] G. Heygster et al., “Remote sensing of sea ice: Advances during the ice extent from satellite and ship observations,” Remote Sens. Environ.,
DAMOCLES project,” Cryosphere, vol. 6, pp. 1411–1434, 2012. vol. 92, no. 1, pp. 98–111, Jul. 2004.
[5] O. M. Johannessen, E. V. Shalina, and M. W. Miles, “Satellite evidence [28] H. J. Zwally, J. C. Comiso, C. L. Parkinson, W. J. Campbell,
for an Arctic sea ice cover in transformation,” Science, vol. 286, no. 5446, F. D. Carsey, and P. Gloersen, “Antarctic sea ice, 1973–1976: Satellite
pp. 1937–1939, 1999. passive-microwave observations,” NASA, SP-459, p. 206, 1983
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZHU et al.: ANTARCTIC SEA-ICE CLASSIFICATION BASED ON CONDITIONAL RANDOM FIELDS 17

[29] M. Steele, S. Dickinson, J. Zhang, and R. W Lindsay, “Seasonal ice loss [53] J. L. Lieser, R. A. Massom, and P. Heil, “Technical report: Sea ice reports
in the Beaufort sea: Toward synchrony and prediction,” J. Geophys. Res., for the season 2013–2014,” Antarctic Climate and Ecosystems Coopera-
vol. 120, pp. 1118–1132, 2015. tive Research Centre, Hobart, Tasmania, Australia, 2014, pp. 1–176.
[30] R. A. Massom, M. R. Drinkwater, and C. Haas, “Winter snow cover on sea
ice in the Weddell Sea,” J. Geophys. Res., vol. 102, no. C1, pp. 1101–1117,
1997.
[31] World Meteorological Organization (WMO). (1970). [Online]. Available:
http://www.wmo.int/
[32] A. P. Worby and I. Allison, “A technique for making ship-based obser-
vations of Antarctic Sea ice thickness and characteristics: Part I. Obser- Tingting Zhu received the B.S. degree in survey-
vational technique and results,” Antarctic Cooperative Research Centre ing and mapping from Nanjing Normal University,
Research, Hobart, Tas., Australia, Rep. 14, May 1999. Nanjing, China, in 2012, and the M.S. degree in sur-
[33] D. A. Clausi, A. K. Qin, M. S. Chowdhury, P. Yu, and P. Maillard, veying and mapping from the State Key Laboratory
“MAGIC: MAP-guided ice classification system,” Can. J. Remote Sens., of Information Engineering in Surveying, Mapping,
vol. 36, no. 1, pp. S13–S25, Dec. 2010. and Remote Sensing (LIESMARS), Wuhan Univer-
[34] Y. Zhong, X. Lin, and L. Zhang, “A support vector conditional random sity, Wuhan, China, in 2014. She is currently working
fields classifier with a Mahalanobis distance boundary constraint for high toward the Ph.D. degree at LIESMARS.
spatial resolution remote sensing imagery,” IEEE J. Sel. Topics Appl. Earth Her research interests include the polar remote
Observ. Remote Sens., vol. 7, no. 4, pp. 1314–1330, Apr. 2014. sensing and change detection.
[35] J. Zhao, Y. Zhong, L. Zhang, “Detail-preserving smoothing classifier based
on conditional random fields for high spatial resolution remote sensing
imagery,” IEEE Trans. Geosci. Remote Sens., vol. 53, no. 5, pp. 2440–
2452, Dec. 2015.
[36] X. Su, C. He, Q. Feng, X. Deng, and H. Sun, “A supervised classification Fei Li received the B.S. degree from the Univer-
method based on conditional random fields with multiscale region con- sity of Science and Technology of China, Hefei,
nection calculus model for SAR image,” IEEE Geosci. Remote Sens. Lett., China, in 1982, and the Ph.D. degree from the Insti-
vol. 8, no. 3, pp. 497–501, May 2011. tute of Geodesy and Geophysics of Chinese Sciences
[37] P. Zhang, M. Li, Y. Wu, and H. Li, “Hierarchical conditional random Academy, Wuhan, China, in 1992.
fields model for semisupervised SAR image segmentation,” IEEE Trans. He is currently the Director of the Chinese Antarc-
Geosci. Remote Sens., vol. 53, no. 9, pp. 4933–4951, Sep. 2015. tic Research Center of Surveying and Mapping,
[38] G. Zhang and X. Jia, “Simplified conditional random fields with class Wuhan University, Wuhan. He is also the Vice Pres-
boundary constraint for spectral-spatial based remote sensing image clas- ident of Wuhan University. His research interests in-
sification,” IEEE Geosci. Remote Sens. Lett., vol. 9, no. 5, pp. 856–860, clude the polar remote sensing, geodesy and geo-
Sep. 2012. physics in polar research, geophysical inversion in
[39] J. Inglada and J. Michel, “Qualitative spatial reasoning for high-resolution lunar gravity and internal structure.
remote sensing image analysis,” IEEE Trans. Geosci. Remote Sens.,
vol. 47, no. 2, pp. 599–612, Feb. 2009.
[40] D. Comaniciu and P. Meer, “Mean shift: A robust approach toward feature
space analysis,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 24, no. 5, Georg Heygster (M’00) received the Diploma degree
pp. 603–619, May 2002. in solid-state physics and the Ph.D. degree in digital
[41] T. Toyoda and O. Hasegawa, “Random field model for integration of local image processing from the University of Göttingen,
information and global information,” IEEE Trans. Pattern Anal. Mach. Göttingen, Germany, in 1976 and 1979, respectively.
Intell., vol. 30, no. 8, pp. 1483–1489, Aug. 2008. From 1979 to 1988, he was a Consultant with
[42] P. Zhong and R. S. Wang, “A multiple conditional random fields ensemble the Computer Center, University of Bremen, Bremen,
model for urban area detection in remote sensing optical images,” IEEE Germany. Since then, after working for one year on
Trans. Geosci. Remote Sens., vol. 45, no. 12, pp. 3978–3988, Dec. 2007. the imaging mechanisms of scanning acoustic mi-
[43] D. A. Clausi and B. Yue, “Comparing cooccurrence probabilities and croscope, he has been the Head of the Geophysi-
Markov random fields for texture analysis of SAR sea ice imagery,” IEEE cal Analysis of Satellite Images Group, Institute of
Trans. Geosci. Remote Sens., vol. 42, no. 1, pp. 215–228, Jan. 2004. Environmental Physics, University of Bremen. His
[44] N. Y. Zakhvatkina, V. Y. Alexandrov, O. M. Johannessen, S. Sandven, research activities include passive and active microwave remote sensing, partic-
and I. Y. Frolov, “Classification of sea ice types in ENVISAT synthetic ularly both surface and atmospheric parameters in high latitudes, various aspects
aperture radar images,” IEEE Trans. Geosci. Remote Sens., vol. 51, no. 5, of the hydrological cycle, long-term trends, and retrieval techniques. He was
pp. 2587–2600, May 2013. or has been the Principal Investigator or Co-investigator of numerous research
[45] C. Bouman and M. Shapiro, “A multiscale random field model for projects funded by the European Union, the European Space Agency, the Ger-
Bayesian image segmentation,” IEEE Trans. Image Process., vol. 3, man Research Foundation, and the Japan Aerospace Exploration Agency. These
no. 2, pp. 162–177, Mar. 1994. projects include the development of sensor software and hardware, conducting
[46] W. T. Freeman and Y. Weiss, “On the optimality of solutions of the max- campaigns, the final data analysis from single- and multisensor data to geophys-
product belief propagation algorithm in arbitrary graphs,” IEEE Trans. Inf. ical parameters, and the interpretation and application of these results in various
Theory, vol. 47, no. 2, pp. 736–744, Feb. 2001. areas such as meteorology, climatology, and oceanography.
[47] H. C. Li, W. Hong, Y. R. Wu, and P. Z. Fan, “On the empirical-statistical
modeling of SAR images with generalized gamma distribution,” IEEE
J. Sel. Topics Signal Process., vol. 5, no. 3, pp. 386–397, Jun. 2011.
[48] M. di Bisceglie and C. Galdi, “CFAR detection of extended objects in
high-resolution SAR images,” IEEE Trans. Geosci. Remote Sens., vol. 43, Shengkai Zhang received the B.S. degree from the
no. 4, pp. 833–843, Apr. 2005. Wuhan Technical University of Surveying and Map-
[49] W. Szajnowski, “Estimators of log-normal distribution parameters,” IEEE ping, Wuhan, China, in 1998, and the Ph.D. degree in
Trans. Aerosp Electron. Syst., vol. AES-13, no. 5, pp. 533–536, Sep. 1977. geodesy from Wuhan University, Wuhan, in 2006.
[50] T. Wan, N. Canagarajah, and A. Achim, “Segmentation-driven image Since then, he was with the Chinese Antarctic
fusion based on alpha-stable modeling of wavelet coefficients,” IEEE Center of Surveying and Mapping, Wuhan Univer-
Trans. Multimedia, vol. 11, no. 4, pp. 624–633, Jun. 2009. sity.
[51] I. A. Koutrouvelis, “Regression-type estimation of the parameters of stable He is currently an Associate Professor at the Chi-
laws,” J. Amer. Statist. Assoc., vol. 75, pp. 918–928, Dec. 1980. nese Antarctic Center of Surveying and Mapping,
[52] C. Jackson and J. Apel, Synthetic Aperture Radar Marine User’s Manual, Wuhan University. His research interests include the
U.S. Dept. Commerce, Washington, DC, USA, 2004. [Online]. Available: polar geodesy and remote sensing.
http://www.sarusersmanual.com/

You might also like