IEEE Geoscience and Remote Sensing Magazine - September 2022 - 138

used g1
to detect oil spills [216], [217]. Shirvany [90],
[211] and Staples [75] used the DoP for oil spill and
ship detection. Cao demonstrated that the CPR, relative
phase, and | all have a certain degree of ship detection
capability [88]. Salbeg proposed three new parameters
based on pseudo-FP data and the X-Bragg scattering
model for oil spill detection [213].
◗ Multiple features used: Atteia and Dabboor directly used
RCM parameters for ship [40], [212] and oil spill [53]
detection. Bouno employed the consistency coefficient,
the DoP, and entropy for oil spill detection [86]. Denbina
harnessed
| , m , and } for iceberg detection [218].
Singh used eigen values for snow detection [87]. Zhang
utilized
m , d , | , and span for internal solitary wave detection
[204]. Atteia used Stokes vector parameters and
subparameters for ship detection [40]. Yin [55] and Kumar
[214], [215] carried out ship detection based on the
decomposition results of m |and
m .dMore
often, PSFs are combined with other detection
methods to obtain higher detection accuracy and lower
false alarm probability, as follows:
◗ PSF + CFAR: Gao combined the proposed new feature
HESA with CFAR based on Gaussian distribution to
achieve adaptive ship detection [62]. Atteia combined
Stokes parameters with the CFAR algorithm and LRT algorithm,
and the proposed hybrid detection algorithm
can be effectively used for ship detection in mediumand
low-resolution CP SAR images [40]. Denbina combined
polarization ellipse parameters ,} ,| and m with
the LRT algorithm to improve iceberg detection performance
[218].
◗ PSF + machine learning: Ji obtained candidate parameters
based on H adecomposition,
combined them
with a weighted support vector machine (SVM) classifier
for ship detection, and used the intensity difference of
the scattering mechanism based on m |decomposition
to remove false alarms [54]. Dabboor first obtained
decorrelated RCM CP parameters for the oil spill detection
task and then used the random forest algorithm for
classification [67]. Zhang chose ,1
g ,0
m H ,w Lambda , span ,
and g3 features with good internal solitary wave
detection ability and combined them with the k-means
clustering algorithm, and the detection results are better
than the traditional Wishart polarization clustering
algorithm [204].
Trisasongko combined the three amplitudes of the
CP covariance matrix (i.e., ,, and C )22
polarimetric features derived from m |CC
11
12
and the
decomposition
with the random forest, SVM, and other algorithms
for waterlogged rice field detection, respectively, and the
results show that the polarimetric features from m |decomposition
have a slight advantage over FP backscattering
coefficients [146].
◗ PSF + CFAR + machine learning: Fan combined the decomposition
results of m |and
m d138
method
and the SVM classification method to achieve
with the CFAR
MACHINE LEARNING BASED
By considering target detection as a binary classification
problem of backgrounds and targets, machine learning can
be used to learn the distinction between the two and thus
achieve the target detection task. Fan introduced the SVM
algorithm in the final step of a proposed ship detection algorithm
to achieve pixel-level classification [66]. Ji implemented
ship detection using a weighted SVM classifier directly
after feature selection and feature enhancement [54].
Dabboor also implemented oil spill detection using the
random forest algorithm after the decorrelation of RCM CP
parameters [67]. Fan combined a U-Net network structure
with CP SAR data for ship detection. Experimental results
show that CP SAR outperforms single-polarized and DP
SAR, and the proposed method outperformed the CFAR
as well as the Faster Region-Based Convolutional Neural
Network (RCNN) [70]. Trisasongko used the random forest
and SVM to detect waterlogged rice fields, based on the polarimetric
features of the CP covariance matrix and m |decomposition
[146]. Zhang implemented internal solitary
wave detection using the k-means clustering algorithm after
feature selection [204].
A comparison of current CP SAR target detection methods
appears in Table 13. Table 14 summarizes these CP SAR
target detection methods in detail. In it, two types of methods,
" pseudo-FP + PSF " and " pseudo-FP + CFAR, " are also
counted in addition to the four mixed methods. As shown
in Tables 11 and 14 and Figure 15, currently, CP SAR target
detection is mainly carried out in marine scenarios, which
reflects people's concern for marine targets. However, attention
for other types of targets in other complex scenes
also is needed, and in these cases, questions like, " What
can CP SAR do? " " What is the advantage of CP SAR? " and
" How can we use CP SAR theories and methods? " wait for
answers. In addition, as shown in the " Data Source " column
of Table 14, there has been no open and available CP
SAR data set so far, and most relevant research has been
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE SEPTEMBER 2022
pixel-level ship detection with good false alarm rejection
[66].
◗ PSF + CFAR + visual attention: Xu used the pulsed cosine
transform (PCT) visual attention model to suppress
background clutter and highlight suspected ship targets,
used the m ddecomposition
results to eliminate
false alarms and correct the saliency map, and finally
employed the CFAR algorithm based on the lognormal
distribution for ship detection [52].
VISUAL ATTENTION BASED
The visual attention mechanism enables the fast and accurate
extraction of regions of interest and suppression of
background regions from the external complex environment.
Xu used the PCT visual attention model to achieve
prediscrimination between background clutter and suspected
ship targets, and the generated saliency map was
used for subsequent ship detection [52].
IEEE Geoscience and Remote Sensing Magazine - September 2022

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - September 2022
