Fully Polarimetric Synthetic Aperture Radar (SAR)
Processing for Crop Type Identification
Gang Hong, Shusen Wang, Junhua Li, and Jingfeng Huang
Abstract
The target or polarimetric decomposition is widely used to
process multi-polarization
SAR
imagery to establish a cor-
respondence between physical characteristics of interested
objects and observed scattering mechanisms. Polarimetric
decomposition parameters are used as the basis for devel-
oping new classification methods for analyzing polarimetric
SAR
data. This study proposes to combine two polarimetric
decomposition parameters (entropy (H) and angle (
α
)) derived
from the Cloude and Pottier decomposition method and total
scattered power (Span) in crop type identification. Support
vector machine (
SVM
) classification algorithm was selected
as a classifier to resolve limitations of classifications based
on polarimetric decomposition parameters. The advantag-
es of the proposed method are determined by comparing
with other commonly used methods based on polarimetric
features and the results produced from the coherency ma-
trix, i.e., without target decomposition. Results show that
the proposed method is about 10 percent better than other
methods based on polarimetric features without Span, and
it outperforms the result from the coherency matrix with
about 4 percent improvement in the overall accuracy.
Introduction
Microwave
SAR
data is complementary to optical images
in crop identification since it is weather independent and
sensitive to crop structure, the moisture contents of crops
and the soil dielectric under crop canopy.
RADARSAT-
2, a new
generation
SAR
sensor, can acquire multiple polarizations
simultaneously, and the full polarimetric information for a
target can be represented in the form of a scattering matrix.
The capabilities of
SAR
for crop classification may be assessed
in various configurations which include single polarization,
dual polarization, and fully polarimetric modes (Skriver,
2012). The reason for the use of multi-polarization
SAR
in crop
classification is that polarization is sensitive to orientation
and makes it possible to differentiate crops with different
canopy architectures (Skriver
et al
., 2005).
Cloude and Pottier’s polarimetric decomposition parame-
ters (Cloude and Pottier, 1996) have been used for target char-
acterization and as the basis for developing new classification
methods using polarimetric data. The incoherent target de-
composition derives a two-dimensional
H
and
α
classification
plane (
H
/
α
), where the entropy (
H
) is a measure of randomness
of scattering mechanisms, and angle (
α
) characterizes the scat-
tering mechanism. The
H
/
α
plane is subdivided into eight ba-
sic zones with different scattering mechanisms. Different class
boundaries are determined to discriminate scattering type (sur-
face scattering, volume diffusion, and double-bounce scatter-
ing) along the
α
axis and degrees of randomness (low, medium,
and high) along the entropy (
H
) axis. The physical scattering
characteristic associated with each zone provides information
for land cover assignment. However, the problem is caused by
those preset zone boundaries in
H
/
α
plane: classes might fall
on a boundary or more than one class may be included in a
zone (Lee
et al
., 2004). Especially, agricultural crops with sim-
ilar scattering mechanisms overlap from one class to another
in
H
/
α
plane (Tan
et al
., 2011; Atwood
et al
., 2012). The
H
/
α
classification was improved by using additional information
provided by anisotropy (
A
), which is particularly useful for
discriminating scattering mechanisms with different eigenval-
ue distributions but with similar intermediate entropy values
(Touzi
et al
., 2004). However,
H
/
α
/
A
only contains the infor-
mation of the three scattering mechanisms with relative values
and ignores the information of backscatter intensities with
absolute values (Cao and Hong, 2005); the absolute magnitude
of eigenvalues and other parameters are not used in
H
/
α
/
A
classification (Lee
et al
., 1999). Chen
et al
. (2007) reported that
crop types’ discrimination accuracy was very low when only
entropy, alpha, and anisotropy were supplied to maximum
likelihood classifier or a spatial spectral based classifier.
The objective of this study is to propose a new methodolo-
gy which uses
H
,
α
, and power (Span) (Simply called
HaSpan
)
as inputs to
SVM
classification. Among various supervised/
unsupervised classification methods, support vector machine
(
SVM
) classification is one of the most reliable approaches for
crop classifications with
SAR
data (Li
et al
., 2012).
SVM
is to
determine an optimal hyperplane of the input data space to
maximize the distance separating the training classes. Train-
ing support vectors are projected into a higher-dimension
feature space to search for an optimal hyperplane if such a hy-
perplane cannot be found at the current feature space.
SVM
is
widely used in remote sensing mainly because: it can handle
the high dimensionality problems existed in statistical pattern
classifications (Chen, 1999); it can produce high classification
accuracy with a small training data set (Shao and Lunetta,
2012; Foody and Mathur, 2006); it is robust to any overfitting
problem due to maximizing margin from the training samples
instead of finding a decision boundary (Hsu
et al
., 2003);
its structure is less complex compared with that of Neural
Gang Hong is with the Department of Geography, York Uni-
versity, Toronto, ON, M3J 1P3, Canada.
Shusen Wang and Junhua Li are with the Canada Center for
Mapping and Earth Observation, Natural Resources Canada,
560 Rochester St., Ottawa, ON K1A 0E4, Canada
(
).
Jingfeng Huang is with the College of Environment and Re-sources
Science, Zhejiang University, Hangzhou, 310058, China.
Photogrammetric Engineering & Remote Sensing
Vol. 81, No. 2, February 2015, pp. 109–117.
0099-1112/15/812–109
© 2014 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.81.2.109
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
February 2015
109
