IEEE Geoscience and Remote Sensing Magazine - March 2017 - 36

Equation (1) represents a linear kernel that computes a
dot product in feature space. Equation (2) is a polynomial
kernel, where d 2 0 is a constant that defines the kernel
order. The RBF kernel is represented by (3), where c is the
weight. Finally, (4) shows a particular kind of two-layer sigmoid neural network that essentially serves as a similarity
measure between x i and x j . Note that each kernel has a
dot product term ( x i ·x j ) to measure the similarity between
two vectors x i and x j.
TRAdITIONAL MACHINE-LEARNING ALGORITHMS
fOR IMAGE CLASSIfICATION
The derivation of land-cover information increasingly relies on RS technology due to its ability to measure land
surfaces at various spatial and spectral scales. Classification is one of the major approaches to find land-cover
information from remotely sensed imagery. Many classification algorithms have been devised since the first Landsat
image was acquired in the early 1970s [36], [37]. The most
popular approaches for land-cover mapping are neural
networks, decision trees, and maximum likelihood classifiers (MLCs). The MLC is a parametric classifier based
on statistical theory. Neural networks are nonparametric
techniques with potential discriminating ability. They
avoid some problems faced by MLCs and attract substantial research effort. Among
the many variants of neural
networks developed so far
LEARNING REfERS TO fINd[38], multilayer perceptron
ING AN OpTIMAL dECISION
is the most popular network
bOuNdARY TO SEpARATE
used for RS image classificaTHE TRAINING pATTERNS
tion [39], [40]. Decision trees
ANd THEN TO SEpARATE
are simple parametric classiTEST dATA uNdER THE
fiers that are widely used for
SAME CONfIGuRATION.
classification tasks [16]. It is
expressed as a recursive partition of the instance space,
breaking a very complex classification problem into multiple stages of simpler decision-making processes [41]. Depending on the number of variables used, univariate and
multivariate models are used in decision-making processes. Univariate decision trees were developed for land-cover
mapping at a global scale [42], [43]. Multivariate decision
trees are more compact and accurate than univariate decision trees but involve more algorithmic complexity. For
more complex architectures, like the compound classification of multitemporal data [44], multiclassifier systems
made up of neural algorithms [45], [46] have been developed to address specific problems. All these methods share
the common idea to perform the learning of the classification algorithms according to the minimization of the empirical risk associated with the errors in the training set.
However, SVMs [17] are based on the structural risk minimization principles rather than on empirical risk minimization. They are promising algorithms because of their
lower sensitivity to the curse of dimensionality compared
36

to traditional classification approaches. In the last decade,
SVMs have gained significant success in RS applications. A
review focusing on the assessment of SVMs on land-cover
classification is available in [47]. A recent review focused
on the applications of SVMs for RS data analysis [48].
The objective of our survey is to increase awareness
of the benefits of different SVM-based classification
methods for researchers in the field of RS data analysis.
This review illustrates the most important algorithmic
developments and applications in the domain of RS research, emphasizing the achievements obtained by the
RS research community. However, we would like to point
out that most of the approaches in this article cannot be
grouped into a comparative meta-analysis because each
study has been developed on the basis of different experimental methodologies, and it applies learning algorithms differently.
SVM-bASEd AdVANCEd AppROACHES
fOR THE CLASSIfICATION Of REMOTE
SENSING IMAGERY
In the last decade, many studies have been published in the
RS literature on the application of SVM classifiers in the
analysis of RS imagery [48]. The SVM approach has been
applied first to the classification of hyperspectral data [49]
without reducing the dimensionality. Since the acquisition of labeled samples is expensive and time-consuming,
researchers have developed many SVM-driven techniques
to face complex problems related to the properties of RS
imagery. In this context, it is worth mentioning that the
semisupervised classifiers were developed to address illposed problems that are characterized by a very small ratio
between the number of available training samples and the
number of features, known as the Hughes phenomenon
[50], by reinforcing the learning procedure through the use
of unlabeled samples. Although SVMs provide excellent
generalization capability, they may fail to model the classification problem with very few training samples (ill-posed
classification problems).
In these situations, the exploration of unlabeled samples to enrich the information of the training samples can
result in a significant improvement for model development. The reader can find a comprehensive analysis on
AL and SSL strategies applied to multispectral and hyperspectral image classification in [51]. For a review on
various techniques on multispectral and hyperspectral RS
image classification, the reader is referred to [52]. A good
review on the algorithmic advances in the supervised
learning and SSL methods with kernels for RS image classification is available in [53]. In this survey, we have focused on SVM-based approaches, including active SVMs,
S3VMs, and composite SVM methods, for addressing RS
image classification problems. We have identified 55 research articles in seven leading journals, including some
other sources related to the RS applications from 2004 to
2015 (Table 1). The frequency of articles regarding these
ieee Geoscience and remote sensing magazine

march 2017

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