IEEE Geoscience and Remote Sensing Magazine - June 2019 - 162

than 100 samples for alfalfa or oats). Water absorption
bands (104 → 108, 150 → 163, and 220) are usually removed before processing.
In spite of its limited size, it is one of the main reference
data sets of the community, even though rare classes are
usually not taken into account when evaluating classification algorithms.

HYPERSPECTRAL IMAGE ANALYSIS ISSUES AND
STANDARD APPROACHES
This section briefly recalls standard issues and current approaches for hyperspectral data processing. In particular,
supervised statistical learning approaches for classification
are detailed because they are obvious reference baselines
and inspiration for methods based on deep learning.

DATA SET SUMMARY
Various data sets are compiled in Table 1. This highlights
that the main issue with applying machine-learning approaches to hyperspectral data is in the small number of
available samples. Existing data sets are small with respect
to standard optical imagery. Moreover, because of sensor
diversity and postprocessing methods, it is not possible to
train algorithms simultaneously on different data sets.
Nevertheless, these data sets are available and have
been shared among researchers for years thanks to the
good will of some research groups throughout the world
(for example, from http://www.ehu.eus/ccwintco/index
.php?title=Hyperspectral_Remote_ Sensing _ Scenes).
Moreover, the IEEE Geoscience and Remote Sensing Society (GRSS) is providing the community with the GRSS
Data and Algorithm Standard Evaluation (DASE) website
(see http://dase.grss-ieee.org/). Through DASE, researchers
can access the data for Indian Pines, Pavia, and DFC2018,
and they can submit classification maps, which are evaluated online. For each data set, a leaderboard allows participants to compare the state-of-the-art methods as soon as
they are tested.

PREPROCESSING AND NORMALIZATION
Working with hyperspectral images often implies preprocessing the data. Besides the aforementioned atmospheric
and geometric corrections, band selection and normalization are often also applied. Those normalizations will impact how classifiers are able to separate spectral features in
the ways described next.

(a)

(b)

(c)

(d)

FIGURE 4. Examples of train and test splits on the Indian Pines data

set. (a) Train (random), (b) test (random), (c) train (disjointed), and
(d) test (disjointed).
162

BAND SELECTION
Depending on the sensor, some spectral bands might be
difficult to process or may contain outliers that modify the
spectrum dynamics. For example, we often remove bands
related to water absorption, bands with a low signal-tonoise ratio, and saturated values. This not only improves
the robustness of the classifiers by alleviating the noise
present in the data, but it also helps fight against the wellknown curse of dimensionality that provokes decreasing
performances of statistical classification models when the
dimensions of the data increase. Band selection can also be
used by dropping uninformative bands, e.g., using principal component analysis (PCA) [14] or mutual information
[15]. However, band selection should be done carefully.
Unsupervised dimension reduction can sometimes lead
to worse performance than using the raw data because it
might remove information that is not useful for compression but was discriminant for classification [16].
STATISTICAL NORMALIZATION
It is a common practice in the machine-learning community
to normalize the data beforehand to rely on the common assumptions for which classifiers are known to behave well,
such as zero mean and unit variance. The standard strategies
often allow for a significant improvement of processing with
statistical approaches. We denote the individual spectra by X i
and the whole image by I. These strategies then are as follows:
◗ Using the spectral angle, the normalized variant of the
spectrum with a unit Euclidean norm: X = X/ X . The
angle between two spectra is a common similarity measure used for classification, notably in the popular spectral angle mapper classifier [17].
◗ Normalizing first- and second-order moments (to obtain a zero mean and unit variance). This can be done for
each band independently, which works especially well
with classifiers that expect all features to have similar
amplitudes, such as support vector machines (SVMs).
However, this squashes the dynamics in the spectral dimension. Alternatively, the normalization can be done
globally on the whole image: I = ^I - m I h v I .
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

JUNE 2019

http://www.ehu.eus/ccwintco/index.php?title=Hyperspectral_Remote_Sensing_Scenes http://www.ehu.eus/ccwintco/index.php?title=Hyperspectral_Remote_Sensing_Scenes http://dase.grss-ieee.org/

IEEE Geoscience and Remote Sensing Magazine - June 2019

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - June 2019