IEEE Geoscience and Remote Sensing Magazine - December 2020 - 62

MACHINE LEARNING AND FE: AN EVER-GROWING
RELATION
Figure 2(a) illustrates the basic idea of a machine-learning
approach, which consists of FE and classification. In machine learning, users are requested to provide guidelines for
the machine (algorithm). This is usually done by applying
handcrafted FE approaches to produce informative features
for the subsequent classifier. At the very beginning, each
image pixel is regarded as a pattern, and its spectrum (i.e.,
a vector of different values of a pixel in different spectral
channels) is considered the initial set of features. This set
of features, also known as spectral features, suffers from two
important downsides: 1) the features are often redundant,
and 2) they do not consider the spatial dependencies of the
adjacent pixels.
To address the first issue, a feature-reduction step (through
FE or selection) can be applied to reduce the dimensionality of the input data (from p 1 dimensions in the original
data to p 2 dimensions in a new feature space, p 2 1 p 1).
This step, also called spectral FE, tries to preserve the key
spectral information of the data by reducing the dimensionality and maximizing separability between classes. It
is interesting that the second issue can also be addressed
using FE approaches. Note that, here, the FE step, also
known as spatial FE, is not aimed at reducing the dimensionality; instead, it is intended to model (extract) spatial
contextual information suitable for the subsequent classification or object-detection step, and it usually leads to
an increase in the number of features. The simultaneous

[8]. This finding went against the general understanding
of hyperspectral data, which wrongly maintained that full
dimensionality is always better than subspace in terms of
classification accuracies.
The unique characteristics of high-dimensional data, as
discussed, have a pronounced impact on the performances
of supervised classifiers [9], as they demand an adequate
number of training samples, which is almost impossible
to obtain in practice since the collection of such training
samples is either expensive or time demanding. To address
this issue, FE-based dimensionality reduction is found to
be effective.
STORAGE SYSTEMS AND PROCESSING TIMES
We are now in the era of massive data acquisition. Statistics demonstrate that the cumulative volume of existing big
data has increased tremendously, from 4.4 ZB in 2013 to
44 ZB in 2020 [130]. The Earth observation (EO) community has also faced a similar trend because of the enormous
volume and variety of data being generated by EO missions.
For example, EnMAP, a hyperspectral satellite mission, is
planning to capture hyperspectral data with a maximum
ground coverage of 5,000 # 30 km per day and a target revisit time of four days ^! 30c h with 512-Gb onboard mass
memory [10]. FE-based dimensionality reduction helps in
data compression, which leads to faster transmission time,
removal of redundant features, smaller storage space requirements, and decreasing the required time for performing the same computations.

x2
w

∗
w

x
∗

-
x

b
-

=
b

1
=

∗

w


-1

w


(a)

(b)
FIGURE 2. FE via (a) machine learning and (b) DL.

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

DECEMBER 2020

IEEE Geoscience and Remote Sensing Magazine - December 2020

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2020