IEEE Geoscience and Remote Sensing Magazine - March 2017 - 38

TaBLE 1. aDVaNcED cLaSSIFIErS USING SVms IN rS aPPLIcaTIONS. (CONTINUED)
machINELEarNING mEThODS

cLaSSIFIErS

rS DaTa

rEFErENcES

sVms integrated with
other approaches

integrated Bayesian rule
and sVm

atm, Landsat-5 tm, and Landsat-7
enhanced tm+

[129]

Ga and sVm

Hyperspectral

[133]

Pso and sVm

multispectral

[137]

moG and sVm

multitemporal

[139]

Ga and sVm

multisource rs data

[140]

multiobjective fuzzy clustering
and sVm

multispectral

[141]

sVm and selective Bayesian
thresholding

multispectral

[142]

cluster-based ensemble algorithm

rosis data

[143]

Generative and discriminative model
combined

rosis hyperspectral

[146]

Fcm and sVm

multispectral

[147]

GKclust fuzzy clustering and sVm

multispectral

[148]

Fuzzy topology integrated sVm

thematic mapper image

[150]

ASTER: Advanced Spaceborne Thermal Emission and Reflection Radiometer; TM: thematic mapper, ATM: airborne topographic mapper, ERS:
European remote sensing.

9
6

2014

2013

2012

1
2011

2010

2009

2008

2007

2006

2005

2015

2004

ACTIVE SVMs FOR REMOTELY SENSED
IMAGE CLASSIFICATION
Conceptually separate from the SSL paradigm, AL enriches
both the information given as input to the supervised classifier as well as improves the classifier's performance by iteratively expanding the ground truth according to a process
that involves an interaction between the user and automatic
classification system. AL usually utilizes three strategies, i.e.,
the geometrical features of SVMs, the estimation of the posterior probability distribution function (pdf) of the classes,
and a query-by-committee (QBC) paradigm to obtain an optimal training set. In this framework, the model returns the
sample that has a highly uncertain classification result. The
user must label the unspecified
point accurately, which is then
merged with the training set
THE MOST pOpuLAR
to increase the generalization
AppROACHES fOR LANdperformance of the model.
COVER MAppING ARE NEuWhen the AL process is over,
RAL NETWORkS, dECISION
the classifier is trained with the
TREES, ANd MAxIMuM
final training set, and classifiLIkELIHOOd CLASSIfIERS.
cation of the considered image
is performed. This optimizes
the model for the challenging
problem considered. However, the complexity of the problem, financial resources, and overlapping class boundaries
may reduce the efficiency of an algorithm. Although AL
methods have theoretical and empirical advantages, their

Frequency of Papers

techniques published through January 2015 is illustrated
in Figure 1.

January
fIGuRE 1. The number of articles on RS studies reviewed from
2004 on a yearly basis through January 2015.

use in RS image classification is limited. A review of AL algorithms can be found in [54].
Some promising SVM-driven AL methods to generate
optimal training sets can be found in the RS literature. If
the training set is biased, active SVMs converge slowly. An
AL model was developed using SVMs to query the labels of
pixels from the unlabeled data [55]. The method was applied
successfully to a multispectral image of the Indian RS satellite with a spatial resolution of 36.25 m. The most effective unlabeled point is searched at each step and labeled by
a human annotator. The objective is to reduce the number of
labeled points used by the SVM classifier by several orders
compared to standard SVMs. This query-based SVM resulted
in better performance compared to conventional SVMs on
multispectral imagery. However, the addition of a single
ieee Geoscience and remote sensing magazine

march 2017

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