IEEE Geoscience and Remote Sensing Magazine - March 2017 - 40

and the output scores of unlabeled samples are estimated to
find the most uncertain points belonging to the low-density regions of the feature space. Subsequently, a K-means
clustering algorithm is applied to these points to minimize
redundancy and maintain diversity. However, this clustering approach suffers from the lack of correspondence between the clusters found and the classes desired by the user.
This limits the use of the cluster assumption in defining the
query functions in AL.
An interactive-domain-adaptation technique based on
AL is proposed in [65] for RS image classification. In this approach, the SVM is first trained using the source-domain
training samples. Next, the MCLU-ECBD query function
proposed in [63] is used at all subsequent iterations to find
the most informative samples of the target image to annotate
by human expert. The newly labeled samples are merged with
the training set to retrain the
supervised classification algorithm. The classifier is trained
MuLTIVIEW dISAGREEMENTconsidering different weights
bASEd AL IS VERY EffECTIVE
for instances of the source and
IN HYpERSpECTRAL IMAGE
target domains. Another quCLASSIfICATION.
ery function removes sourcedomain samples that are inconsistent for the target-domain
problem. The proposed method yields significantly higher
accuracy than the standard methods, including random selection, the standard MCLU AL method [63], and a method that
combines the MCLU query technique with the reweighting
procedure proposed in [66] on VHR and hyperspectral data.
A similar technique based on MLC is proposed in [67] for the
classification of VHR and hyperspectral images.
Multiview disagreement-based AL is very effective in
hyperspectral image classification. Using the idea of QBC,
the multiview information constructs a committee of
classifiers from the feature space, and their disagreement
is used to measure the uncertainty for a given unlabeled
sample [68]. This study investigated different view generation from hyperspectral data using the clustering, random
selection, and uniform subset slicing methods. In addition, this technique incorporates a quantitative disagreement measure to compensate for the risk of view insufficiency and avoids the inflation of the contention pool.
The experimental results show its effectiveness compared
to random sampling and SVMs with MS. This multiview
method, which was originally developed for the SSL paradigm, can also be used to develop AL methods. In [69], regularization has been integrated with AL for hyperspectral
data classification. The first regularizer measures the disagreement level among the highly uncertain samples and
builds a small contention pool of unlabeled samples. The
second regularizer measures consistency based on the spatial- or spectral-based manifold space. It further focuses on
the most informative samples within the contention pool
by penalizing rapid changes in the classification function
close to sample points. This proposed method performed
40

very well with hyperspectral data compared to random
sampling and the state-of-the-art SVM-based simple margin sampling.
The mechanism of combining spectral and spatial criteria of pixels through an iterative process for active sample
selection was investigated in [70]. MS and breaking-of-ties
(BT) heuristics are used to incorporate spectral information. For estimating spatial information, three criteria are
proposed. The first criterion refers to the computation of
Euclidean distances from the training samples, while the
second criterion is based on the Parzen window method.
The last criterion is the idea of spatial entropy. The spectral and spatial information are then combined through
multiobjective optimization. Experimentation is conducted on the two VHR images using the proposed approach
and three spatial criteria. Results indicate that the methods using spatial information provide statistically significant advantages with respect to the techniques involving
spectral information.
Persello and Bruzzone [51] investigated different approaches for integrating SSL and AL in real classification
problems by leveraging the merits of both strategies. In
this context, the proposed progressive S3VM with diversity
(PS3VM-D) framed with traditional AL concepts improved
the classification performance. The merits and drawbacks
of both strategies were analytically studied to determine
possible ways to improve RS applications characterized by
different kinds of data and classification problems. A series
of experiments has been conducted both on synthetic and
real multispectral and hyperspectral RS data using SVM
classification methods.
SEMISUPERVISED SUPPORT VECTOR MACHINES
The S3VM is an alternative route to supervised SVMs,
which is entirely built on the available ground truth samples. S3VMs leverage both labeled and unlabeled samples
in the training phase of the learner, and they often outperform SVMs under proper problem definitions when few
labeled samples are available. Recently, the RS community
has focused on developing semisupervised approaches to
address the problems of RS data classification. In [71], the
authors investigated a classification problem for spatially distributed RS imagery using a biased SVM (BSVM) as
proposed in [72]. A BSVM assumes all unlabeled samples
belong to one class (Cothers) and tries to minimize the
number of those unlabeled samples that are classified as
a particular class of interest (Cint) while maintaining the
labeled samples (i.e., Cint) to be correctly classified. Each
of the unlabeled samples is assigned a weighting factor to
indicate the likelihood of this assumption. The BSVM was
shown to be superior to other SVM-based techniques like
positive example-based learning (PEBL) [73] and Rocchio
algorithm-SVM [74].
TSVMs have been proposed to handle ill-posed classification problems (i.e., the available training set does
not represent the actual data distribution) [75], [76].
ieee Geoscience and remote sensing magazine

march 2017

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