IEEE Geoscience and Remote Sensing Magazine - December 2017 - 47

We can also see from Figure 8 that, as the number of features
increases, the classification performance does not always increase-in fact, some decrease. To achieve optimal classification performance, the number of reduced features
needs to be optimized.

SPECTRAL CLASSIFIERS
Based on the availability of training samples (also referred
to as learning with a teacher) for the training stage, classification approaches can be grouped into three categories:
supervised, unsupervised (also known as clustering), and
semisupervised approaches.
Supervised approaches classify input data using training samples. These samples are usually collected in one of
two ways: 1) by manually labeling a small number of pixels in an image or 2) based on some field measurements.
In contrast, unsupervised classification does not consider
training samples. The supervised approach classifies input
data based only on an arbitrary number of initial cluster
centers that may be either user-specified or selected quite
arbitrarily. During processing, each pixel is associated with
one of the cluster centers, usually in an iterative way, based
on a similarity criterion [85], [86]. In semisupervised approaches [87], the training stage is based not only on labeled training samples but also on unlabeled samples.
Because the consideration of training samples leads
to higher classification accuracies than in situations
where there is no class-specific information, supervised

OA (%)

CHALLENGES FOR DIMENSIONALITY REDUCTION
Recent advances in sensor technologies and processing techniques strongly support the use of hyperspectral data. Moreover, global Earth observation missions (e.g., AVIRIS from
NASA, the PROBA series from the European Space Agency,
and the Gaofen series from China) make such data increasingly accessible. Furthermore, at lower altitudes, airplanes and unmanned aerial vehicles can deliver extremely
high-resolution hyperspectral data from targeted locations.
In addition, image processing techniques allow us to extract
multiple-level features from these big hyperspectral data.
Two main challenges remain in hyperspectral DR: 1)
mining complementary features (while reducing the dimension and redundancy) from multiple levels of big hyperspectral data and 2) coupling DR and applications in a
unified framework, ensuring that optimal features for applications are obtained. Most state-of-the-art research has
separated DR and applications into two different steps.
For example, morphological operators were employed
in [80] to extract low-level features (such as the size and
shape of objects) from remote-sensing images. In [81] and
[82], middle-level attribute features were extracted from
HSIs for land cover mapping. High-level features, such
as object-based [83] and so-called deep learning features
[84], have been used for CD and classification.
State-of-the-art DR methods typically deal with either
lower-level or higher-level features, but not with a combination of both. The features extracted at each level have
their own characteristics: high-level -features are usually
more powerful but less robust, while low-level ones are
less informative but more robust.
On the other hand, classification
is taken as one of the most popular
100
applications to validate DR perfor90
mances. Hyperspectral classification
typically consists of two steps: 1) DR
80
(via either feature extraction or fea70
ture selection) and 2) a training procedure for designing the classifier.
60
However, it is difficult to ensure that
the best features from the first step
50
will optimize the classification per40
formance of the -following one.

covers by assigning unknown pixel vectors to one of the
classes (or clusters). The individual classes are commonly
differentiated based on the similarity to a certain class or
by defining decision boundaries constructed in the feature
space. The initial set of features for classification usually encompasses spectral channels [4].
With reference to Figure 1, two types of classification
approaches can be broadly defined: spectral classifiers
and spectral-spatial classifiers [4], where the former consider the HSIs to be a list of spectral measurements with
no spatial organization, while the latter classify the input
data by taking into account the spatial dependencies of
adjacent pixels.

DECEMBER 2017

PCA
SELD

LPP
SEGL

NWFE
OTVCA

SDA

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30

CLASSIFICATION
HSI classification is a fast-growing
and highly active field of research
in the hyperspectral community. A
classification algorithm is used to
distinguish between different land

Raw
SELF

Number of Features
FIGURE 8. The classification performance on the Pavia University data as the number of

features increases. Forty labeled training samples per class were randomly selected from the
training set with an SVM classifier.

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

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