IEEE Geoscience and Remote Sensing Magazine - March 2017 - 42

reinforce training samples and demonstrates its capability to improve the performance on hyperspectral images.
The classification performance of kernel-based methods using spectral information can be further augmented
by incorporating spatial information into the classifier. To
this end, a family of composite-kernels has been proposed
in the context of SVMs that combine spectral, spatial, and
local cross-information in a hyperspectral image [94]. The
idea of composite-kernels is introduced in [95] for hypertext classification. For this purpose, spatial feature vectors
are derived from either a computed mean or the mean and
standard deviation taken together from a certain neighborhood window of the corresponding spectral feature vector. Subsequently, kernel matrices are computed for both
spectral and spatial vectors and combined using different
strategies. These spectral-spatial classifiers have noticeably
better classification performance compared to methods
using only spectral signatures. In another work, Marconcini
et al. [96] proposed a PS3VM
that uses unlabeled samples
to increase the reliability when
ANOTHER IMpORTANT CATfew training samples are availEGORY Of CLASSIfICATION
able and uses composite kerpRObLEMS IN THE CONTExT
nels to take into account spectral and spatial information
Of RS dATA IS THE AuTOembedded in the image. The
MATIC updATE Of LANdapproach was tested using a
COVER MApS.
hyperspectral image acquired
by the Airborne Visible/InfraRed Imaging Spectrometer
(AVIRIS) sensor, and PS3VM with composite kernels was
shown to outperform both S3VM and PS3VM with single
kernels as well as supervised SVMs with composite kernels.
In some classification problems, the available training set
is not fully reliable. For instance, a novel context-sensitive
S3VM classifier was devised in [97] to overcome this problem. The proposed approach updates the small-size training data by gathering contextual knowledge from the adjacent pixels of each training point. Experimental analysis
of the proposed scheme on the VHR multispectral image demonstrates its superiority over the supervised and
semisupervised classification algorithms including conventional SVMs, PS3VMs, maximum likelihood, and K-nearest
neighbor in terms of effectiveness when the training set is
not fully reliable.
The issue of kernel predetermination was investigated
by proposing a regularization strategy that identified kernel
structure through the analysis of unlabeled pixels [98]. In
this method, a base kernel is deformed with a likelihood of
a kernel encoding the geometrical relationships between labeled and unlabeled samples. This kernel is encoded by running a clustering algorithm several times over the unlabeled
points. The resulting kernel is known as a cluster kernel. In
this way, a spatial-spectral regularization is performed in the
input space. The final kernel is obtained by taking the sum or
product of RBF and the bagged kernel. The method has been
42

successfully tested in multispectral and hyperspectral image
classification tasks. S3VM based on cluster kernels for urban
image classification was also tested [99]. In a recent letter
[100], an S3VM with a deformed kernel function is proposed.
This method is suggested to increase the similarity between
two pixels in the same cluster with respect to both spatial
and spectral assessment. Spectral consistency is achieved by
designing a bagged kernel, while a spatial kernel is obtained
by running a hierarchical segmentation algorithm on the
entire image. The deformed kernel is a linear combination
of an RBF kernel, a spatial kernel, and a spectral kernel. The
SVM learner is then constructed using this deformed kernel.
The proposed method incorporating spatial consistency outperforms other traditional semisupervised methodologies
in terms of accuracy and smoothness on multispectral and
hyperspectral data. An AL approach is engineered in [101]
under a self-learning framework for searching new, useful
unlabeled samples using machine-machine interaction instead of the human supervision that is commonly used in
state-of-the-art AL. The proposed semisupervised approach
integrates the spatial and spectral information to find highly
informative samples using two probabilistic classifiers, i.e.,
multinomial logistic regression and a probabilistic pixelwise SVMs. The methods are demonstrated using different
selection algorithms including BT, MS, random sampling,
modified BT, and normalized entropy querying-by-bagging
(nEQB) on AVIRIS and reflective optic system imaging
spectrometer (ROSIS) data.
Another important category of classification problems
in the context of RS data is the automatic update of landcover maps, which is necessary due to the increased number of images acquired by the satellite sensors on the same
region at different times. This problem is addressed in [102]
under the hypothesis that reliable ground truth samples
are not always helpful for all considered acquisitions. The
classification model is devised in the domain-adaptation
framework by introducing domain-adaptation SVMs. Furthermore, a novel circular accuracy-measuring strategy is
proposed for the validation of the results provided by the
classifiers when no ground truth information of the image
is available. Experimental results over a multitemporal and
multispectral data set confirmed the usefulness and the reliability of the proposed scheme.
In the RS domain, an important machine-learning
approach is the semisupervised one-class classification. In
the one-class classification, one attempts to detect pixels
belonging to one of the classes in the image and ignores
the others. In RS, researchers investigated a semisupervised
single-class SVM to detect an oil-slick in a synthetic aperture
radar (SAR) image for detecting ocean pollution [103]. A specific kernel is designed to perform accurate segmentation of
the local sea-surface wave spectrum using both radiometric
and texture information. The SAR imagery is decomposed
into a multiscale representation using a wavelet transform.
A region of interest is then defined from a nonpolluted
sea area to generate the training set, which improves the
ieee Geoscience and remote sensing magazine

march 2017

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - March 2017