3830 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL.

15, 2022

AIR-PolSAR-Seg: A Large-Scale Data Set for Terrain

Segmentation in Complex-Scene PolSAR Images
Zhirui Wang , Member, IEEE, Xuan Zeng , Zhiyuan Yan , Member, IEEE, Jian Kang , Member, IEEE,
and Xian Sun , Senior Member, IEEE

Abstract—Polarimetric synthetic aperture radar (PolSAR) ter- Index Terms—Benchmark data set, polarimetric synthetic
rain segmentation is a fundamental research topic in PolSAR image aperture radar (PolSAR), terrain segmentation.
interpretation. Recently, many studies have been investigated to
handle this task. However, the existing data for PolSAR terrain seg-
mentation have relatively limited scale and their scene complexity
I. INTRODUCTION
is relatively simple. These issues greatly restrict the development of YNTHETIC aperture radar (SAR) is an active sensor
algorithms. Therefore, there is a strong requirement for establish-
ing a large-scale data set for terrain segmentation in complex-scene
PolSAR images. In this paper, we present a benchmark data set
S imaging radar system which can provide high-resolution
observation data under all-time and all-weather conditions
containing a PolSAR amplitude image with a 9082×9805-pixel [1]–[3]. Different from single-polarization radar, Polarimet-
region and 2000 image patches with a size of 512×512 for PolSAR ric SAR (PolSAR) can obtain richer information by trans-
terrain segmentation, which is called AIR-PolSAR-Seg. We collect mitting and receiving orthogonally polarized electromagnetic
the PolSAR image with a resolution of 8 m from the GaoFen-3 waves [4]–[6]. This working mechanism brings polarization
satellite, and it is equipped with pixel-wise annotation which cov-
ers six categories. Compared with the previous data resources, diversity, which can reveal different electromagnetic scattering
AIR-PolSAR-Seg preserves some specific properties. First, AIR- characteristics of the earth terrain. Due to the superior quality
PolSAR-Seg owns a large-size PolSAR image and provides a large in the acquisition of full polarization information [7], PolSAR
quantity of image patches. It offers the research community a com- images collected by airborne and satellite sensors have been
plete data resource with adequate training examples and reliable widely used in many geo-science and remote sensing applica-
validation results. Second, AIR-PolSAR-Seg is established upon a
PolSAR image with high scene complexity. This characteristic mo- tions, such as disaster prediction, environmental monitoring, and
tivates robust and advanced segmentation approaches to facilitate other related fields [5], [8]–[12]. Hence, it is of great significance
complex-scene PolSAR image analysis. Based on AIR-PolSAR-Seg, to interpret PolSAR images [13]–[16].
three tasks are introduced: multi-category segmentation, water In the field of PolSAR image interpretation, terrain segmenta-
body segmentation, and building segmentation. Moreover, a perfor- tion is an important and hot research topic, since it has significant
mance analysis of traditional approaches and deep learning-based
approaches are conducted, which can be regarded as baselines and application value [17]–[21]. In particular, many practical appli-
provide references for future research. cations need an intelligent system that can correctly classify
each pixel in PolSAR images with a specific type of terrain [5],
[11], [22], [23]. This demand motivates researchers to develop
Manuscript received November 6, 2021; revised January 3, 2022, March 1, effective methods for PolSAR terrain segmentation [24]–[28]. In
2022, and April 14, 2022; accepted April 22, 2022. Date of publication April the process of scientific research, a benchmark data set is an in-
28, 2022; date of current version May 20, 2022. This work was supported dispensable component [29], [30]. That is because a benchmark
by the National Natural Science Foundation of China under Grant 61725105,
Grant 62076241, Grant 62171436, and Grant 62101371, and in part by the data set not only provides training samples to build innovative
Jiangsu Province Science Foundation for Youths under Grant BK20210707. models, but also presents reliable evaluation results to testify
(Corresponding author: Xian Sun.) the new developed methods. Therefore, it is of great research
Zhirui Wang and Zhiyuan Yan are with the Aerospace Information Research
Institute, Chinese Academy of Sciences, Beijing 100190, China, and also value to establish a high-quality and reliable data set for terrain
with the Key Laboratory of Network Information System Technology (NIST), segmentation in PolSAR images.
Aerospace Information Research Institute, Chinese Academy of Sciences, Bei- In recent years, a large amount of PolSAR data has been
jing 100190, China (e-mail: [email protected]; [email protected]).
Xuan Zeng and Xian Sun are with the Aerospace Information Research introduced to aid in the advancement of terrain segmentation
Institute, Chinese Academy of Sciences, Beijing 100190, China, with the Key in PolSAR images [29]. These data were used as supporting
Laboratory of Network Information System Technology (NIST), Aerospace resources in previous research works, hastening the development
Information Research Institute, Chinese Academy of Sciences, Beijing 100190,
China, with the University of Chinese Academy of Sciences, Beijing 100190, of many excellent PolSAR terrain segmentation approaches
China, and also with the School of Electronic, Electrical and Communication [4], [5], [31]. With the evolution of aerospace remote sensing
Engineering, University of Chinese Academy of Sciences, Beijing 100190, technology, higher requirements are put forward in practical
China (e-mail: [email protected]; [email protected]).
Jian Kang is with the School of Electronic and Information Engineer- applications [32]–[34]. However, there are several constraints on
ing, Soochow University, Suzhou 215006, China (e-mail: kangjian_1991@ the existing data, which brings challenges to the development
outlook.com). of PolSAR terrain segmentation and will further influence its
AIR-PolSAR-Seg is downloadable at https://github.com/AICyberTeam/AIR-
PolSAR-Seg. application in large-scale and complex real-world scenes. The
Digital Object Identifier 10.1109/JSTARS.2022.3170326 details are discussed as follows.
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
WANG et al.: AIR-POLSAR-SEG: A LARGE-SCALE DATA SET FOR TERRAIN SEGMENTATION IN COMPLEX-SCENE POLSAR IMAGES 3831

Fig. 1. Examples of some open-source PolSAR data [22]. (a) Pauli pseudo image and the annotation, Flevoland. (b) Pauli pseudo image and the annotation, San
Francisco.

Constraint 1 Limited Data Scale: The majority of open-source 1) We establish a data set, i.e., AIR-PolSAR-Seg, for ter-
PolSAR data have a common issue: the data scale is relatively rain segmentation in PolSAR images. Compared with the
limited, which may preclude PolSAR image interpretation re- existing data, AIR-PolSAR-Seg has a large data scale.
search on large-scale scenes. In recent years, the evolution of Specifically, AIR-PolSAR-Seg contains a PolSAR image
computer science and deep learning has opened up an inno- with a 9082×9805-pixel region and 2000 image patches
vative approach for PolSAR terrain segmentation. Adopting with a size of 512×512. In addition, AIR-PolSAR-Seg
deep learning-based approaches to advance the task of PolSAR presents high scene complexity in terms of terrain target
terrain segmentation has become the mainstream [18], [19], [35]. arrangement and spatial distribution.
These deep learning-based methods depend on large-scale and 2) Based on AIR-PolSAR-Seg, three tasks, including multi-
well-annotated data to train and validate the networks, since category segmentation, water body segmentation, and
deep learning is a data-driven approach [36]–[39]. However, the building segmentation, are constructed. Besides, perfor-
existing PolSAR data present relatively limited scale in image mance evaluation and detailed analysis are conducted on
size and image quantity [4], [5], [29], [31], [40], [41]. Therefore, a set of typical approaches with AIR-PolSAR-Seg. The
the existing data may not be appropriate for the research work experimental results demonstrate that the AIR-PolSAR-
on deep learning and PolSAR image analysis. Seg is a proper data set to develop robust algorithms for
Constraint 2 Simple Scene Complexity: Another problem PolSAR terrain segmentation, which can be taken as the
encountered in most open-source PolSAR data is that they baseline results for the subsequent research.
usually present simple scene complexity. As shown in Fig. 1, 3) The data set is provided in an open format. Readers can
the terrain targets in different categories have a regular and obtain these references for future research on PolSAR
uniform distribution across the images. In practical situations, image interpretation. The AIR-PolSAR-Seg can provide
the spatial distributions of different terrain targets have a large the research community with a better data resource to
diversity. This phenomenon leads to high scene complexity of evaluate and develop advanced algorithms in PolSAR
terrain regions on the earth’s surface. For example, the building image analysis.
regions are regular but appear to have complex scattering results. The rest of this article is organized as follows. We first provide
Though the scattering intensity of water is weak, the irregular a comprehensive review of related PolSAR data by summariz-
shape brings great challenges for fine-grained segmentation. In ing the main characteristics in Section II. Then, the details of
this case, the methods based on the existing data can only settle AIR-PolSAR-Seg are described in Section III. In Section IV,
simple situations and show limited performance in complex- three segmentation tasks on AIR-PolSAR-Seg are provided,
scene situations. and the performance of the baseline algorithms on three seg-
With the consideration of the above constraints, this article mentation tasks are given. Finally, Section V concludes this
contributes to the research field of PolSAR terrain segmentation article.
by establishing a large-scale PolSAR data set in complex-scene
situations (i.e., AIR-PolSAR-Seg). Compared with the exist- II. A REVIEW OF RELATED POLSAR DATA
ing data, AIR-PolSAR-Seg is empowered with a larger scale
Previous studies collected PolSAR data using various sensors,
both in terms of the image size and the number of image
including airborne sensors (AIRSAR, ESAR) and spaceborne
patches. It can provide sufficient training examples and re-
sensors (RADARSAT-2). In this section, the related PolSAR
liable evaluation results to support the scientific research of
data will be reviewed according to the types of sensors.
PolSAR terrain segmentation. In addition, the PolSAR image in
AIR-PolSAR-Seg presents high scene complexity. This prop-
erty encourages researchers to promote robust terrain seg- A. Airborne Platform
mentation approaches in complex-scene situations. Briefly, 1) Airsar: The PolSAR data of Flevoland farmland [5] is
the main contributions of this article are summarized captured by the NASA/JPI AIRSAR sensor in L-band. The
as follows. image from the area of Flevoland farmland has 750×1024 pixels,
3832 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 15, 2022

Fig. 4. Oberpfaffenhuofen data, E-SAR. (a) Pauli pseudo image. (b) The
ground truth. (c) The color codes.
Fig. 2. Flevoland data, AIRSAR. (a) Pauli pseudo image. (b) The ground truth.
(c) The color codes.

Fig. 3. San Francisco data, AIRSAR. (a) Pauli pseudo image. (b) The ground Fig. 5. Flevoland data, RADARSAT-2. (a) Pauli pseudo image. (b) The ground
truth. (c) The color codes. truth. (c) The color codes.

and the spatial resolution is about 10 m. Fig. 2 illustrates the Pauli 2) E-SAR: The multi-look PolSAR data of Oberpfaffen-
pseudo image and the corresponding annotations. As shown hofen [4], [31] is obtained from the E-SAR platform in L-band.
in Fig. 2(b), this PolSAR data is annotated with 15 terrain Fig. 3 shows the Pauli pseudo image and the annotations. The
categories. selected area of Oberpfaffenhofen contains 1300×1200 pixels,
The PolSAR data of San Francisco Bay [5] is provided by the and the spatial resolution is 3×3 m. As shown in Fig. 4(b), the
AIRSAR platform in L-band. The image size is 900×1024, and PolSAR data of Oberpfaffenhofen has 3 terrain categories, i.e.,
the spatial resolution is 10 m. There are 5 terrain categories to be woodland areas, open areas, and built-up areas.
identified, i.e., mountain, ocean, urban, vegetation, and bare soil.
Fig. 3 demonstrates the Pauli pseudo image and the annotations.
The ground truth of San Francisco data was labeled by the Key B. Spaceborne Platform
Laboratory of Intelligent Perception and Image Understanding 1) RADARSAT-2: The PolSAR data of Flevoland area [5] is
of Ministry of Education, Xidian University (IPIU team) [29]. provided by the RADARSAT-2 satellite in C-band. Fig. 5 dis-
In the annotation procedures, the research team first collected plays the Pauli pseudo image and the annotations. The selected
the high-resolution Google imagery of the same period and then region covers 1635×2375 pixels, and the spatial resolution is
manually labeled the color map on the Google imagery using the 8 m. As shown in Fig. 5(b), there are 4 terrain categories to be
Labelme software. Next, the annotated color map was remapped identified, including urban, water, forest, and cropland.
to get the ground truth. Both the San Francisco data of AIRSAR This fully PolSAR data of San Francisco [5] is provided by
and RADARSAT-2 in the following contents were annotated the RADARSAT-2 satellite. Fig. 6 presents the Pauli pseudo
through the above procedures. image and the annotations. The region of this data contains
WANG et al.: AIR-POLSAR-SEG: A LARGE-SCALE DATA SET FOR TERRAIN SEGMENTATION IN COMPLEX-SCENE POLSAR IMAGES 3833

TABLE I
COMPARISON ON THE RELATED POLSAR DATA AND AIR-POLSAR-SEG

the building procedures of the AIR-PolSAR-Seg is described.

Then, Section III-B presents the properties of AIR-PolSAR-Seg
in detail and illustrates that AIR-PolSAR-Seg is an appropriate
data set for PolSAR terrain segmentation.

A. The Building Procedures of AIR-PolSAR-Seg

1) Image Collection: The PolSAR amplitude image in AIR-
PolSAR-Seg was captured by the GaoFen-3 (GF-3) satellite on
April 29, 2019. The GF-3 satellite is a spaceborne platform
launched in August 2016, which is a full-polarimetric SAR
sensor in C-band. The GF-3 satellite captured this image over
the area of Hangzhou in Zhejiang province, China and worked
in the mode of quad-polarized strip I (QPSI).
The spatial resolution of this image data is 8 m, and the
selected region contains 9082×9805 pixels. There are four
Fig. 6. San Francisco data, RADARSAT-2. (a) Pauli pseudo image. (b) The polarization modes in this image data, including vertical-vertical
ground truth. (c) The color codes. (VV), horizontal-vertical (HV), horizontal-horizontal (HH), and
vertical-horizontal (VH). The amplitude image with four polar-
ization modes, the corresponding optical image from Google
1380×1800 pixels, and the spatial resolution is 8 m. As shown
Earth, and the corresponding ground truth are shown in Fig. 7.
in Fig. 6(b), this region is annotated with 5 terrain categories,
Besides, the main characteristics are also displayed in Table I.
i.e., developed, high-density urban, low-density urban, water,
2) Terrain Category: The scene types of AIR-PolSAR-Seg
and vegetation.
contain the urban areas and the suburbs with different density
The main characteristics of the related PolSAR data men-
levels. Specifically, six typical terrain categories are chosen
tioned above are summarized in Table I. The scale of these
and annotated, including housing areas, industrial areas, natural
PolSAR data is relatively limited, among which the image size
areas, land use areas, water areas, and other areas. The above
almost ranges from 1000 to 2000 pixels. This phenomenon leads
categories are the most typical terrain types, which can cover
to insufficient training samples and test samples, resulting in
common terrain targets on the earth’s surface.
the lack of adequate image patches for deep learning-based
3) Image Annotation: We describe the annotation proce-
algorithms, further limiting their application in large-scale real-
dures in this part. Firstly, the radiation calibration operation is
world scenes. In this case, the deep learning-based approaches
implemented since the speckle noise and the complex scattering
relying on the scale of data cannot achieve promising results.
mechanisms exist in the PolSAR image [42], [43]. Furthermore,
Besides, as shown in Figs. 2–6, the distribution of different
the corresponding optical image with maximum resolution from
terrain categories in these data has regular and uniform shapes,
Google Earth is used as a reference for human annotator [see
and the scene complexity of these images is relatively simple. As
Fig. 7(f)], and its shooting time is the same as that of the full
a result, the terrain segmentation approaches based on these data
polarization SAR image. Then, using the optical image as a
are incapable of handling terrain targets with irregular shapes
reference, we label the full polarization SAR image at pixel
and perform poorly in complex-scene situations. In summary,
level with LabelMe software. Cross-validation is performed
the data mentioned above fall short of providing comprehensive
among annotators in order to reduce category confusion in the
resource support for the subsequent research.
annotation procedures. Besides, the interpretation experts in
SAR field have checked the annotated labels and corrected the
III. AIR-POLSAR-SEG: A NEW DATA SET FOR POLSAR
mislabeled regions.
TERRAIN SEGMENTATION
The ground truth and the color codes are displayed in Fig. 7(c)
To advance the development of PolSAR terrain segmenta- and Fig. 7(g), respectively. As shown in Fig. 7(g), the housing
tion algorithms, we constructed AIR-PolSAR-Seg, a large-scale areas, industrial areas, natural areas, land use areas, water areas,
PolSAR data set with high scene complexity. In Section III-A, and other areas are marked in yellow mask (255, 255, 0), blue
3834 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 15, 2022

Fig. 7. The PolSAR amplitude image in AIR-PolSAR-Seg. (a) Amplitude image of HH channel. (b) Amplitude image of HV channel. (c) The ground truth.
(d) Amplitude image of VH channel. (e) Amplitude image of VV channel. (f) The corresponding optical image form Google Earth. (g) The color codes.

mask (0, 0, 255), green mask (0, 255, 0), red mask (255, 0,
0), cyan mask (0, 255, 255), and white mask (255, 255, 255), B. Why AIR-PolSAR-Seg is Appropriate for PolSAR Terrain
respectively. Moreover, Fig. 8 also displays some examples of Segmentation
the six terrain categories in AIR-PolSAR-Seg. Compared with the related PolSAR data, AIR-PolSAR-Seg
4) Dataset Splits: The PolSAR image in the AIR-PolSAR- presents the following characteristics.
Seg is quite large (a 9082×9805-pixel region) and cannot 1) Large Data Scale: AIR-PolSAR-Seg is a relatively large-
be directly processed by the deep neural networks. To facil- scale data set with a large-size PolSAR image and a large
itate the experimental setups, we crop 500 patches from the quantity of image patches. To be specific, the PolSAR image
original image data with a size of 512×512 [1], and each image in AIR-PolSAR-Seg covers a 9082×9805-pixel region, which
patch contains four polarization modes (HH, HV, VH, and VV). is much larger than that of other data resources listed in Table I.
Therefore, the sum of image patches is 2000. Besides, as mentioned in Section III-A4, the total number of
To keep the data distribution as consistent as possible between the image patches from the original PolSAR image is 2000.
training data and test data, 80% of the image patches are ran- The number of training samples accounts for 1600, and the
domly selected to form the training set, and the remaining 20% number of testing samples is 400. In this case, AIR-PolSAR-Seg
serve as the testing set. In the following experiments, the models can provide relatively adequate terrain category instances to
are trained on the training set and evaluated on the testing set. build robust deep learning-based algorithms. Based on the above
The training set, testing set, and corresponding ground truth are illustrations, AIR-PolSAR-Seg is a relatively large-scale data
released publicly. This data splitting protocol will be used for set, which is suitable to support research work on large-scale
feature research on the AIR-PolSAR-Seg. terrain segmentation with PolSAR images.
WANG et al.: AIR-POLSAR-SEG: A LARGE-SCALE DATA SET FOR TERRAIN SEGMENTATION IN COMPLEX-SCENE POLSAR IMAGES 3835

Fig. 8. Samples of the terrain categories in AIR-PolSAR-Seg. (a) The patch of housing areas and the corresponding optical image. (b) The patch of industrial
areas and the corresponding optical image. (c) The patch of natural areas and the corresponding optical image. (d) The patch of land use areas and the corresponding
optical image. (e) The patch of water areas and the corresponding optical image. (f) The patch of other areas and the corresponding optical image.

2) High Scene Complexity: Scene complexity is another es-

sential point that needs to be taken into account for the es-
tablishment of a data set. Fig. 8 displays some examples of
the six terrain categories in AIR-PolSAR-Seg. As shown in
Fig. 8, the housing areas (yellow mask) are usually distributed in
L-shape, and presented dense arrangements in city regions. The
industrial areas (blue mask) are generally distributed as separate
rectangular areas, which are divided by the regions of other
categories. The natural areas (green mask) contain regions of
multiple textures that present similar scattering characteristics,
including farmland, forest, and meadow. The land use areas (red
mask) consist of several components, such as bare rock, soil,
and sand. The water areas (cyan masks) have the appearance
of narrow curves and have many bends and twists. In this
case, a promising algorithm is supposed to estimate the accu-
rate segmentation results at the region boundary and decrease Fig. 9. The proportion of pixels in each category with respect to the whole
PolSAR image.
error predictions over the surrounding heterogeneous region.
These terrain categories mentioned above have their own unique
characteristics, and the category combination makes the terrain proportion of pixels labeled as natural is 41%. On the surface of
segmentation a great challenge. In addition, we also calculate the earth, it is a common phenomenon that the distribution of ter-
the proportion of pixels in each category with respect to the rain categories is imbalanced. The category imbalance in AIR-
whole PolSAR image, and the results are visualized in Fig. 9. PolSAR-Seg is consistent with practical applications, making
The proportion of pixels belonging to land use accounts for 1%, it complex and posing significant challenges in PolSAR terrain
the percentage of pixels annotated as water body is 12%, and the segmentation.
3836 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 15, 2022

TABLE II
SCENE COMPLEXITY ANALYSIS

Local Binary Pattern (LBP) [47]. The classifiers include Support

Vector Machine (SVM) [48] and Random Forest (RF) [49].
In addition, the evaluation also includes several represen-
tative convolutional neural network (CNN) based approaches:
Fully Convolutional Network (FCN) [50], Pyramid Scene Pars-
ing Network (PSPNet) [51], Encoder-Decoder with Atrous
Separable Convolution Network (DeepLab V3+) [52], Asym-
metric Non-local Neural Network (ANN) [53], Criss-Cross
Attention Network (CCNet) [54], Dual Attention Network
(DANet) [55], Non-Local Neural Network (NonLocal) [56],
Expectation-Maximization Attention Network (EmaNet) [57],
Context Encoding Network (EncNet) [58], Point-wise Spatial
Fig. 10. The distribution of boundary points between different categories on
Attention Network (PSANet) [59], and GCNet [60].
AIR-PolSAR-Seg.
B. Implementation Settings
Furthermore, to quantitatively analyze the scene complexity In the following experiments, all the algorithms are imple-
of AIR-PolSAR-Seg, the boundary points between different mented on the Pytorch and MMSegmentation codebase [61].
categories are calculated on the PolSAR image. If the regions of ResNet-50 [36] serves as the backbone of the segmentation
different categories have an uneven distribution and a complex network. The stochastic gradient descent optimizer (SGD) [62]
arrangement, the PolSAR image will have a large number of is used to tune the parameters in the training state. There are
boundary points. The ground truth annotation is processed with 50 K iterations in the entire training process. The learning rate is
the Sobel operator [44] to retrieve the boundary points. Fig. 10 initially set to 0.001 and decays by a factor of 10 at the iteration
demonstrates the visualization results of the boundaries on AIR- of 40 K. The momentum of the optimizer is set to 0.95. The
PolSAR-Seg. For intuitive comparison, we also calculated the training samples and the validation samples are prorated at a
boundary points on the related PolSAR data, and the results ratio of 4:1. The sliding window is adopted in the traditional
are shown in Table II. The number of boundary points on algorithms, and the kernel size of the sliding window is set to
AIR-PolSAR-Seg is far exceeds that of other data, which is 16×16 grid with a spacing step of 8 [63].
2 863 635. This result demonstrates that the terrain categories
in AIR-PolSAR-Seg are more irregular and diverse. As a con- C. Evaluation Metrics
sequence, AIR-PolSAR-Seg presents high scene complexity. The segmentation performances are evaluated on Overall
In summary, AIR-PolSAR-Seg presents the property of large Accuracy (OA), Average Accuracy (AA), Kappa Coefficient
data scale and high scene complexity. It can provide a better (Kappa), and Mean Intersection over Union (mIoU ). The
benchmark to evaluate and develop advanced PolSAR segmen- formulations of each evaluation metric are described in the
tation approaches. following contents.

IV. EXPERIMENTAL STUDIES M

OA = (1)
N
A. Baseline Methods
where M denotes the sum of pixels that are segmented correctly,
Two types of algorithms are evaluated as the benchmark on and N calculates the sum of pixels in the whole image. OA is
AIR-PolSAR-Seg. For each type, a set of typical and repre- an intuitive measure of revealing the classification performance
sentative algorithms are chosen as the baselines for evaluation. across the entire image.
Specifically, the experiments involve some traditional algo-
1 T Mi
rithms, which are a combination of several feature extraction AA = (2)
algorithms and early machine-learning classifiers. The feature T i=1 Ni

extraction algorithms include Scale Invariant Feature Transform where T represents the sum of terrain categories, i is indexing
(SIFT) [45], Histogram of Oriented Gradient (HOG) [46], and of category, Mi denotes the sum of pixels that are correctly
WANG et al.: AIR-POLSAR-SEG: A LARGE-SCALE DATA SET FOR TERRAIN SEGMENTATION IN COMPLEX-SCENE POLSAR IMAGES 3837

recognized as the category of i, and Ni means the sum of pixels purpose of this setup is to promote more anti-interference seg-
belonging to the category of i. mentation methods on building targets. The research achieve-
OA − p ments of building segmentation have vital roles in urban
Kappa = (3) planning.
1−p
Water body and building segmentation are aimed at specific
1 T and fine terrain categories, and can provide some method-
p= H (i, :) ∗ H (:, i) (4)
N2 i=1 ological support for future research. For water body segmen-
where H represents the confusion matrix. H(i, :) is the sum of tation and building segmentation, other categories in AIR-
elements in i-th row of H, and H(:, i) is the sum of elements PolSAR-Seg are considered as background, and the labels
in i-th column of H. The operator ∗ represents the dot product of other categories are encoded to be consistent with the
operation in matrix. background.
In pixel classification problems, OA is the commonly used
evaluation metric, which can directly reflect the correct propor-
tion of scores. However, considering the imbalance of pixels in E. Experimental Results
each category, the calculation of Kappa can well compensate for
the simplicity of OA, since it is based on the confusion matrix. In this section, we evaluate different approaches on AIR-
PolSAR-Seg. The results of the experiments are shown by the
1 T Pii tasks they were used for: multi-category terrain segmentation,
mIoU = k k (5)
T +1 i=0 P
j=0 ij + j=0 Pji − Pii water body segmentation, and building segmentation.
1) Multi-Category Terrain Segmentation: Table III illus-
where Pij means the number of pixels which belong to class i
trates the multi-category terrain segmentation results on AIR-
but are recognized as class j.
PolSAR-Seg. The approaches involve six traditional methods,
which are a combination of several feature extraction algo-
D. Experimental Protocols on AIR-PolSAR-Seg
rithms and classifiers (e.g., SIFT+SVM, SIFT+RF, HOG+SVM,
In order to further facilitate the research development of ter- HOG+RF, LBP+SVM, and LBP+RF). Besides, the experimen-
rain segmentation in PolSAR images, three typical segmentation tal approaches include several segmentation methods based
tasks are introduced in this section: multi-category terrain seg- on deep learning. Although different methods show different
mentation, water body segmentation, and building segmentation. performances on AIR-PolSAR-Seg, we can observe that the
Details about the experimental protocols are listed below. methods based on deep learning outperform the traditional
1) Multi-Category Terrain Segmentation: In this setting, all methods. These deep learning-based algorithms show a per-
the categories (mentioned in Section III-A2) are involved in formance of 44.23%∼52.58% on mIoU , 75.53%∼77.46%
the experiments and evaluations. This protocol challenges the on OA, 53.90%∼64.97% on AA, and 62.06%∼65.73% on
robustness and generalization ability of algorithms to recog- Kappa. Compared with the traditional algorithms, these deep
nize multiple categories comprehensively, which encourages learning-based algorithms gain a boost of 10.81%∼16.09% on
research development on multi-category terrain segmentation. mIoU , 10.01∼13.13% on OA, 4.31%∼9.14% on AA, and
Multi-category terrain segmentation has great value in practical 13.84%∼16.82% on Kappa. The above results demonstrate the
applications, such as urban planning, agriculture assessment, superiority of these deep learning-based methods. In addition,
and environmental monitoring. Table III also contains the evaluation metric of IoU for each
2) Water Body Segmentation: The second setting focuses on terrain category. Among the traditional algorithms, the land use
the water body, aiming to encourage research work on water category shows a performance of 0.10%∼0.29% on IoU . For
body segmentation. The purpose of establishing the water body deep learning-based algorithms, this terrain category shows a
segmentation is that the water body usually has a narrow shape performance of 0.49%∼2.47% on IoU . This is because there
and a long size, which poses great challenges to the algorithms. are few pixels belonging to land use, accounting for only 1%.
To advance this task, the algorithms should have a stronger The category imbalance in AIR-PolSAR-Seg is consistent with
capability of identifying the irregularly distributed water bodies the practical applications, which brings great challenges for the
and countering the interference from adjacent categories. This segmentation algorithms.
task helps researchers to explore advanced methods of water 2) Water Body Segmentation: Table IV illustrates the wa-
body segmentation, which plays important roles in practical ter body segmentation results. Specifically, the deep learning-
applications, such as disaster monitoring, hydraulic engineering, based algorithms show a performance of 85.80%∼89.29% on
and water transportation. mIoU , 98.48%∼98.83% on OA, 88.11%∼91.56% on AA,
3) Building Segmentation: The third setting pays more at- and 83.70%∼86.90% on Kappa. Compared with these tra-
tention to the category of building. In practical applications, ditional methods, the deep learning-based methods gain a
the buildings in PolSAR images are usually interfered with the boost of 14.41%∼16.15% on mIoU , 1.58%∼1.75% on OA,
surrounding environments, leading to incomplete segmentation 11.56%∼13.17% on AA, and 22.29%∼22.37% on Kappa. The
and inaccurate recognition of buildings [64]. Under this proto- above results show that deep learning-based methods outper-
col, the terrain segmentation method is required to accurately form traditional methods in feature construction and water body
retrieve the building target regions from PolSAR images. The segmentation.
3838 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 15, 2022

TABLE III
RESULTS ON MULTI-CATEGORY TERRAIN SEGMENTATION

TABLE IV
RESULTS ON WATER BODY SEGMENTATION

Fig. 11. Visualization results on water body. (a) The PolSAR image patch.
(b) The annotation. (c) Segmentation result (DANet).

technology in PolSAR image interpretation. Meanwhile, it also

demonstrates that convolutional networks have better perfor-
mance on building segmentation compared with hand-crafted
The PolSAR image patch, the ground truth, and the water
methods.
body segmentation result are visualized in Fig. 11(a)–(c), re-
The PolSAR image patch, the ground truth, and the building
spectively. Compared with the ground truth, the segmentation
segmentation result are visualized in Fig. 12(a)–(c), respec-
result on the narrow water body suffers from fragmentation
tively. The segmentation results on densely distributed buildings
and incompleteness, which is shown in the red rectangles. In
present inaccurate boundaries, and the details are shown in the
practical applications, this phenomenon also exists and limits
red rectangles. This phenomenon exists in practical applications,
the recognition accuracy of the water body segmentation system.
which limits the recognition accuracy of building segmentation
Therefore, conducting more work on water body segmentation
systems. To find promising solutions to the inaccurate problem,
tasks based on AIR-PolSAR-Seg has great research value.
we encourage researchers to conduct more studies on AIR-
3) Building Segmentation: Table V illustrates the build-
PolSAR-Seg and building segmentation tasks.
ing segmentation results. Specifically, the deep learning-
based algorithms show a performance of 69.88%∼71.99% on
mIoU , 82.43%∼83.84% on OA, 82.15%∼83.65% on AA, F. Discussion
and 64.50%∼67.39% on Kappa. Compared with these tra- Section IV presents the experimental results of different meth-
ditional methods, the deep learning-based methods gain a ods on the AIR-PolSAR-Seg and we can get several discussions
boost of 13.25%∼13.35% on mIoU , 9.76%∼10.03% on OA, as follows. First, for the multi-category terrain segmentation,the
9.53%∼9.99% on AA, and 19.37%∼20.15% on Kappa. The segmentation methods perform poorly in categories with fewer
performance gap between the deep learning-based methods and pixels, while these methods pay more attention to categories
traditional methods highlights the advantage of deep learning with more pixels. The category imbalance in AIR-PolSAR-Seg
WANG et al.: AIR-POLSAR-SEG: A LARGE-SCALE DATA SET FOR TERRAIN SEGMENTATION IN COMPLEX-SCENE POLSAR IMAGES 3839

TABLE V that the article is reproducible in all its terms for at least five years
RESULTS ON BUILDING SEGMENTATION
and that the materials concerning this article will be publicly
available permanently [65]. All materials concerning this arti-
cle can be downloaded from https://github.com/AICyberTeam/
AIR-PolSAR-Seg.

V. CONCLUSION
In this article, we construct an appropriate data set, i.e.,
AIR-PolSAR-Seg, for PolSAR terrain segmentation. First, AIR-
PolSAR-Seg is empowered with a large scale both in terms
of the image size and the number of image patches, which
can provide sufficient training examples and reliable validation
results to support the scientific research. Second, the PolSAR
image has high scene complexity. This property encourages
researchers to promote robust approaches for complex-scene
PolSAR image interpretation. Based on AIR-PolSAR-Seg, three
segmentation tasks are introduced. Moreover, a performance
analysis of traditional methods and deep learning-based ap-
proaches was conducted on AIR-PolSAR-Seg, which can be
served as the baseline. The data set has been released, which can
provide some references for the subsequent research on PolSAR
image analysis.

REFERENCES
[1] J. Fu, X. Sun, Z. Wang, and K. Fu, “An anchor-free method based on feature
balancing and refinement network for multiscale ship detection in SAR
images,” IEEE Trans. Geosci. Remote Sens., vol. 59, no. 2, pp. 1331–1344,
Feb. 2021.
[2] B. Zou, X. Xu, L. Zhang, and C. Song, “High-resolution PolSAR image
interpretation based on human image cognition mechanism,” IEEE J. Sel.
Topics Appl. Earth Observ. Remote Sens., vol. 11, no. 11, pp. 4259–4269,
Nov. 2018.
[3] W. Xie, G. Ma, F. Zhao, H. Liu, and L. Zhang, “PolSAR image
classification via a novel semi-supervised recurrent complex-valued
Fig. 12. Visualization results on building. (a) The PolSAR image patch. convolution neural network,” Neurocomputing, vol. 388, pp. 255–268,
(b) The annotation. (c) Segmentation result (DeepLab V3+). 2020. [Online]. Available: https://www.sciencedirect.com/science/article/
pii/S0925231220300497
[4] L. Li, J. Zeng, L. Jiao, P. Liang, F. Liu, and S. Yang, “Online active extreme
learning machine with discrepancy sampling for PolSAR classification,”
is consistent with the practical applications, which brings great IEEE Trans. Geosci. Remote Sens., vol. 58, no. 3, pp. 2027–2041, Mar.
challenges in PolSAR terrain segmentation. Second, for water 2020.
[5] X. Liu, L. Jiao, X. Tang, Q. Sun, and D. Zhang, “Polarimetric convolutional
body segmentation, the segmentation results suffer from frag- network for PolSAR image classification,” IEEE Trans. Geosci. Remote
mentation and incompleteness. In this case, conducting more Sens., vol. 57, no. 5, pp. 3040–3054, May 2019.
effective water body segmentation methods to avoid the frag- [6] F. Liu et al., “Hierarchical semantic model and scattering mechanism based
PolSAR image classification,” Pattern Recognit., vol. 59, pp. 325–342,
mented and incomplete phenomenon has great research value. 2016. [Online]. Available: https://www.sciencedirect.com/science/article/
Third, for building segmentation, the results present inaccu- pii/S0031320316000911
rate boundaries. To find promising solutions to the inaccurate [7] Q. Wu, B. Hou, Z. Wen, Z. Ren, and L. Jiao, “Cost-sensitive latent
space learning for imbalanced PolSAR image classification,” IEEE Trans.
problem, we encourage researchers to conduct more studies on Geosci. Remote Sens., vol. 59, no. 6, pp. 4802–4817, Jun. 2021.
building segmentation tasks. [8] G. D. De Grandi, J. Lee, and D. L. Schuler, “Target detection and tex-
In this article, the AIR-PolSAR-Seg contains six terrain cat- ture segmentation in polarimetric SAR images using a wavelet frame:
Theoretical aspects,” IEEE Trans. Geosci. Remote Sens., vol. 45, no. 11,
egories. Terrain segmentation with finer category settings can pp. 3437–3453, Nov. 2007.
improve PolSAR image interpolation and empower practical [9] B. Liu, H. Hu, H. Wang, K. Wang, X. Liu, and W. Yu, “Superpixel-based
applications. We are actively exploring adding finer category classification with an adaptive number of classes for polarimetric SAR
images,” IEEE Trans. Geosci. Remote Sens., vol. 51, no. 2, pp. 907–924,
annotation to AIR-PolSAR-Seg and constantly presenting better Feb. 2013.
data resources to the research community. Finally, all materials [10] H. Bi, J. Sun, and Z. Xu, “Unsupervised PolSAR image classification using
concerning this article have been provided in an open format, discriminative clustering,” IEEE Trans. Geosci. Remote Sens., vol. 55,
no. 6, pp. 3531–3544, Jun. 2017.
along with a link to the code for reading and using. The details [11] F. Liu, L. Jiao, and X. Tang, “Task-oriented GAN for PolSAR image
of the experiments are given. Readers can reproduce the exper- classification and clustering,” IEEE Trans. Neural Netw. Learn. Syst.,
iments according to the materials we offered. We acknowledge vol. 30, no. 9, pp. 2707–2719, Sep. 2019.
3840 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 15, 2022

[12] B. Hou, J. Guan, Q. Wu, and L. Jiao, “Semisupervised classifica- [32] P. Wang, X. Sun, W. Diao, and K. Fu, “FMSSD: Feature-merged single-
tion of PolSAR image incorporating labels’ semantic priors,” IEEE shot detection for multiscale objects in large-scale remote sensing im-
Geosci. Remote Sens. Lett., vol. 17, no. 10, pp. 1737–1741, Oct. agery,” IEEE Trans. Geosci. Remote Sens., vol. 58, no. 5, pp. 3377–3390,
2020. May 2020.
[13] J. Cheng, F. Zhang, D. Xiang, Q. Yin, and Y. Zhou, “PolSAR image clas- [33] X. Sun, Y. Liu, Z. Yan, P. Wang, W. Diao, and K. Fu, “SRAF-Net: Shape
sification with multiscale superpixel-based graph convolutional network,” robust anchor-free network for garbage dumps in remote sensing imagery,”
IEEE Trans. Geosci. Remote Sens., vol. 60, 2022, Art. no. 5209314. IEEE Trans. Geosci. Remote Sens., vol. 59, no. 7, pp. 6154–6168, Jul.
[14] J. Cheng, F. Zhang, D. Xiang, Q. Yin, Y. Zhou, and W. Wang, “PolSAR 2020.
image land cover classification based on hierarchical capsule network,” [34] X. Sun, P. Wang, C. Wang, Y. Liu, and K. Fu, “PBNet: Part-based
Remote Sens., vol. 13, no. 16, 2021. [Online]. Available: https://www. convolutional neural network for complex composite object detection
mdpi.com/2072-4292/13/16/3132 in remote sensing imagery,” ISPRS J. Photogrammetry Remote Sens.,
[15] G. Akbarizadeh, “A new statistical-based kurtosis wavelet energy feature vol. 173, pp. 50–65, 2021. [Online]. Available: https://www.sciencedirect.
for texture recognition of SAR images,” IEEE Trans. Geosci. Remote Sens., com/science/article/pii/S0924271620303555
vol. 50, no. 11, pp. 4358–4368, Nov. 2012. [35] L. Ding et al., “MP-ResNet: Multipath residual network for the semantic
[16] Z. Tirandaz and G. Akbarizadeh, “A two-phase algorithm based on kurtosis segmentation of high-resolution PolSAR images,” IEEE Geosci. Remote
curvelet energy and unsupervised spectral regression for segmentation of Sens. Lett., vol. 19, 2022, Art. no. 4014205.
SAR images,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 9, [36] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
no. 3, pp. 1244–1264, Mar. 2016. recognition,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2016,
[17] F. Mohammadimanesh, B. Salehi, M. Mahdianpari, E. Gill, and M. pp. 770–778.
Molinier, “A new fully convolutional neural network for semantic seg- [37] K. Fu, Z. Chang, Y. Zhang, and X. Sun, “Point-based estima-
mentation of polarimetric SAR imagery in complex land cover ecosys- tor for arbitrary-oriented object detection in aerial images,” IEEE
tem,” ISPRS J. Photogrammetry Remote Sens., vol. 151, pp. 223–236, Trans. Geosci. Remote Sens., vol. 59, no. 5, pp. 4370–4387, May
2019. [Online]. Available: https://www.sciencedirect.com/science/article/ 2021.
pii/S092427161930084X [38] J. Kang, D. Hong, J. Liu, G. Baier, N. Yokoya, and B. Demir, “Learn-
[18] W. Wu, H. Li, X. Li, H. Guo, and L. Zhang, “PolSAR image semantic seg- ing convolutional sparse coding on complex domain for interferometric
mentation based on deep transfer learning-realizing smooth classification phase restoration,” IEEE Trans. Neural Netw. Learn. Syst., vol. 32, no. 2,
with small training sets,” IEEE Geosci. Remote Sens. Lett., vol. 16, no. 6, pp. 826–840, Feb. 2021.
pp. 977–981, Jun. 2019. [39] J. Kang, M. Körner, Y. Wang, H. Taubenböck, and X. X. Zhu, “Building in-
[19] H. Bi, L. Xu, X. Cao, Y. Xue, and Z. Xu, “Polarimetric SAR image semantic stance classification using street view images,” ISPRS J. Photogrammetry
segmentation with 3D discrete wavelet transform and Markov random Remote Sens., vol. 145, pp. 44–59, 2018.
field,” IEEE Trans. Image Process., vol. 29, pp. 6601–6614, 2020. [40] X. Liu, L. Jiao, D. Zhang, and F. Liu, “Polsar image classification based
[20] B. Hou, C. Yang, B. Ren, and L. Jiao, “Decomposition-feature-Iterative- on polarimetric scattering coding and sparse support matrix machine,” in
Clustering-Based superpixel segmentation for PolSAR image classifica- Proc. IEEE Int. Geosci. Remote Sens. Symp., 2019, pp. 3181–3184.
tion,” IEEE Geosci. Remote Sens. Lett., vol. 15, no. 8, pp. 1239–1243, [41] W. Hua, S. Wang, H. Liu, K. Liu, Y. Guo, and L. Jiao, “Semisupervised
Aug. 2018. PolSAR image classification based on improved cotraining,” IEEE J. Sel.
[21] A. P. Doulgeris, “An automatic U -Distribution and markov random field Topics Appl. Earth Observ. Remote Sens., vol. 10, no. 11, pp. 4971–4986,
segmentation algorithm for PolSAR images,” IEEE Trans. Geosci. Remote Nov. 2017.
Sens., vol. 53, no. 4, pp. 1819–1827, Apr. 2015. [42] D. Xiao, Z. Wang, Y. Wu, X. Gao, and X. Sun, “Terrain segmentation
[22] H. Bi, F. Xu, Z. Wei, Y. Xue, and Z. Xu, “An active deep learning in polarimetric SAR images using dual-attention fusion network,” IEEE
approach for minimally supervised PolSAR image classification,” IEEE Geosci. Remote Sens. Lett., vol. 19, 2020, Art. no. 4006005.
Trans. Geosci. Remote Sens., vol. 57, no. 11, pp. 9378–9395, Nov. 2019. [43] F. Sharifzadeh, G. Akbarizadeh, and Y. S. Kavian, “Ship classification
[23] Z. Wen, Q. Wu, Z. Liu, and Q. Pan, “Polar-spatial feature fusion learning in SAR images using a new hybrid CNN-MLP classifier,” J. Indian Soc.
with variational generative-discriminative network for PolSAR classifica- Remote Sens., vol. 47, no. 4, pp. 551–562, 2019.
tion,” IEEE Trans. Geosci. Remote Sens., vol. 57, no. 11, pp. 8914–8927, [44] Z. Jin-Yu, C. Yan, and H. Xian-Xiang, “Edge detection of images based
Nov. 2019. on improved Sobel operator and genetic algorithms,” in Proc. Int. Conf.
[24] H. Jing, Z. Wang, X. Sun, D. Xiao, and K. Fu, “PSRN: Polarimetric space Image Anal. Signal Process., 2009, pp. 31–35.
reconstruction network for PolSAR image semantic segmentation,” IEEE [45] D. G. Lowe, “Distinctive image features from scale-invariant keypoints,”
J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 14, pp. 10716–10732, Int. J. Comput. Vis., vol. 60, no. 2, pp. 91–110, 2004. [Online]. Available:
2021. https://doi.org/10.1023/B:VISI.0000029664.99615.94
[25] X. Zeng, Z. Wang, X. Sun, Z. Chang, and X. Gao, “DENet: Double-encoder [46] P. E. Rybski, D. Huber, D. D. Morris, and R. Hoffman, “Visual classifi-
network with feature refinement and region adaption for terrain segmenta- cation of coarse vehicle orientation using histogram of oriented gradients
tion in PolSAR images,” IEEE Trans. Geosci. Remote Sens., vol. 60, 2022, features,” in Proc. IEEE Intell. Veh. Symp., 2010, pp. 921–928.
Art. no. 5217419. [47] T. Ojala, M. Pietikäinen, and T. Mäenpää, “Gray scale and rotation invari-
[26] D. Xiang et al., “Adaptive statistical superpixel merging with edge penalty ant texture classification with local binary patterns,” in Proc. Eur. Conf.
for PolSAR image segmentation,” IEEE Trans. Geosci. Remote Sens., Comput. Vis., 2000, Berlin, Heidelberg, Germany, 2000, pp. 404–420.
vol. 58, no. 4, pp. 2412–2429, Apr. 2020. [48] J. A. K. Suykens and J. Vandewalle, “Least squares support vector ma-
[27] L. Yu et al., “A lightweight complex-valued DeepLabv3 for semantic chine classifiers,” Neural Process. Lett., vol. 9, no. 3, pp. 293–300, 1999.
segmentation of PolSAR image,” IEEE J. Sel. Topics Appl. Earth Observ. [Online]. Available: https://doi.org/10.1023/A:1018628609742
Remote Sens., vol. 15, pp. 930–943, 2022. [49] M. Pal, “Random forest classifier for remote sensing classification,” Int.
[28] X. Shi, S. Fu, J. Chen, F. Wang, and F. Xu, “Object-level semantic seg- J. Remote Sens., vol. 26, no. 1, pp. 217–222, 2005. [Online]. Available:
mentation on the high-resolution Gaofen-3 FUSAR-Map dataset,” IEEE https://doi.org/10.1080/01431160412331269698
J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 14, pp. 3107–3119, [50] J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks
2021. for semantic segmentation,” in Proc. IEEE Conf. Comput. Vis. Pattern
[29] X. Liu, L. Jiao, F. Liu, X. Hou, D. Zhang, and X. Tang, “PolSF: PolSAR Recognit., 2015, pp. 3431–3440.
image dataset on San Francisco,” 2019, arXiv:1912.07259. [51] H. Zhao, J. Shi, X. Qi, X. Wang, and J. Jia, “Pyramid scene parsing
[30] X. Sun et al., “FAIR1M: A benchmark dataset for fine-grained object network,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2017,
recognition in high-resolution remote sensing imagery,” ISPRS J. Pho- pp. 6230–6239.
togrammetry Remote Sens., vol. 184, 2022, Art. no. 116130. [Online]. [52] L. Chen, Y. Zhu, G. Papandreou, F. Schroff, and H. Adam, “Encoder-
Available: https://www.sciencedirect.com/science/article/pii/S09242716 decoder with atrous separable convolution for semantic image seg-
21003269 mentation,” in Proc. Eur. Conf. Comput. Vis., V. Ferrari, M. Hebert,
[31] B. Hou, Q. Wu, Z. Wen, and L. Jiao, “Robust semisupervised classification C. Sminchisescu, and Y. Weiss, Eds., vol. 11211, Munich, Germany,
for PolSAR image with Noisy labels,” IEEE Trans. Geosci. Remote Sens., 2018, pp. 833–851. [Online]. Available: https://doi.org/10.1007/978-3-
vol. 55, no. 11, pp. 6440–6455, Nov. 2017. 030-01234-2_49
WANG et al.: AIR-POLSAR-SEG: A LARGE-SCALE DATA SET FOR TERRAIN SEGMENTATION IN COMPLEX-SCENE POLSAR IMAGES 3841

[53] Z. Zhu, M. Xu, S. Bai, T. Huang, and X. Bai, “Asymmetric non-local Zhiyuan Yan (Member, IEEE) received the B.Sc.
neural networks for semantic segmentation,” in Proc. IEEE/CVF Int. Conf. degree in microelectronics from Xiamen University,
Comput. Vis., 2019, pp. 593–602. Xiamen, China, in 2016, and the M.Sc. degree in sig-
[54] Z. Huang, X. Wang, L. Huang, C. Huang, Y. Wei, and W. Liu, “CCNet: nal and information processing from the Aerospace
Criss-cross attention for semantic segmentation,” in Proc. IEEE/CVF Int. Information Research Institute, Chinese Academy of
Conf. Comput. Vis., 2019, pp. 603–612. Sciences, Beijing, China, in 2019.
[55] J. Fu et al., “Dual attention network for scene segmentation,” in Proc. She is currently an Assistant Engineer with the
IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2019, pp. 3141–3149. Aerospace Information Research Institute, Chinese
[56] X. Wang, R. Girshick, A. Gupta, and K. He, “Non-local neural net- Academy of Sciences. Her research interests include
works,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2018, computer vision and remote sensing image analysis.
pp. 7794–7803.
[57] X. Li, Z. Zhong, J. Wu, Y. Yang, Z. Lin, and H. Liu, “Expectation-
maximization attention networks for semantic segmentation,” in Proc.
IEEE/CVF Int. Conf. Comput. Vis., 2019, pp. 9166–9175.
[58] H. Zhang et al., “Context encoding for semantic segmentation,” in Proc.
IEEE Conf. Comput. Vis. Pattern Recognit., 2018, pp. 7151–7160.
[59] H. Zhao et al., “PSANet: Point-wise spatial attention network for scene
parsing,” in Proc. Eur. Conf. Comput. Vis., 2018, pp. 267–283. Jian Kang (Member, IEEE) received the B.S. and
[60] Y. Cao, J. Xu, S. Lin, F. Wei, and H. Hu, “GCNet: Non-local networks M.E. degrees in electronic engineering majoring in
meet squeeze-excitation networks and beyond,” in Proc. IEEE/CVF Int. electronic information engineering from the Harbin
Conf. Comput. Vis. Workshops, 2019, pp. 1971–1980. Institute of Technology, Harbin, China, in 2013 and
[61] M. Contributors, “MMSegmentation: OpenMMLab semantic segmenta- 2015, respectively, and the Dr.-Ing. degree in signal
tion toolbox and benchmark,” 2020. [Online]. Available: https://github. processing in Earth observation from Signal Process-
com/open-mmlab/mmsegmentation ing in Earth Observation (SiPEO), Technical Univer-
[62] Y. Le Cun et al., “Backpropagation applied to handwritten zip code sity of Munich, Munich, Germany, in 2019.
recognition,” Neural Comput., vol. 1, no. 4, pp. 541–551, Dec. 1989. In August of 2018, he was a Guest Researcher
[63] G.-S. Xia et al., “AID: A benchmark data set for performance evaluation with the Institute of Computer Graphics and Vision,
of aerial scene classification,” IEEE Trans. Geosci. Remote Sens., vol. 55, TU Graz, Graz, Austria. From 2019 to 2020, he was
no. 7, pp. 3965–3981, Jul. 2017. with the Faculty of Electrical Engineering and Computer Science, Technische
[64] H. Jing, X. Sun, Z. Wang, K. Chen, W. Diao, and K. Fu, “Fine building Universität Berlin, Berlin, Germany. He is currently with the School of Electronic
segmentation in high-resolution SAR images via selective pyramid dilated and Information Engineering, Soochow University, Suzhou, China. His research
network,” IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 14, interests include signal processing and machine learning techniques, and their
pp. 6608–6623, 2021. applications in remote sensing, with particular focus on intelligent SAR/InSAR
[65] A. C. Frery, L. Gomez, and A. C. Medeiros, “A badging system for data processing, and deep learning-based techniques for remote sensing image
reproducibility and replicability in remote sensing research,” IEEE J. Sel. analysis.
Topics Appl. Earth Observ. Remote Sens., vol. 13, pp. 4988–4995, 2020. Dr. Kang was a Guest Editor of the IEEE JOURNAL OF SELECTED TOPICS IN
APPLIED EARTH OBSERVATIONS, AND REMOTE SENSING. He was the recipient
of the first place of the Best Student Paper Award in EUSAR 2018, Aachen,
Germany. His joint work was selected as one of the 10 Student Paper Competition
Zhirui Wang (Member, IEEE) received the B.Sc. Finalists in IGARSS 2020.
degree in electronic information engineering from
the Harbin Institute of Technology, Harbin, China, in
2013, and the Ph.D. degree in information and com-
munications engineering from Tsinghua University,
Beijing, China, in 2018.
He is currently an Assistant Researcher with Xian Sun (Senior Member, IEEE) received the B.Sc.
Aerospace Information Research Institute, Chinese degree in electronic information engineering from the
Academy of Sciences, Beijing. His research interests Beijing University of Aeronautics and Astronautics,
include synthetic aperture radar (SAR) terrain classi- Beijing, China, in 2004, and the M.Sc. and Ph.D.
fication, and SAR target detection and recognition. degrees in signal and information processing from
the Institute of Electronics, Chinese Academy of Sci-
ences (CAS), Beijing, in 2009.
He was a Visiting Scholar, Karlsruher Institut für
Xuan Zeng received the B.Sc. degree in communica- Technologie, Karlsruhe, Germany, in 2013. He is
tion engineering from the Minzu University of China, currently a Professor with the Aerospace Informa-
Beijing, China, in 2019. She is currently working tion Research Institute, CAS. His research interests
toward the Ph.D. degree in signal and information include computer vision, geospatial data mining, and remote sensing image
processing with the University of Chinese Academy understanding.
of Sciences, Beijing, China, and the Aerospace In- Dr. Sun was the recipient of the Outstanding Science and Technology Achieve-
formation Research Institute, Chinese Academy of ment Prize of the CAS in 2016 and the First Prize for The State Scientific
Sciences, Beijing, China. and Technological Progress of China in 2019. He is an Associate Editor for
Her research interests include computer vision and IEEE ACCESS and a Guest Editor for the Special Issue of the IEEE JOURNAL
deep learning, especially on PolSAR image segmen- OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING
tation and SAR object detection. and other journals.