IEEE Geoscience and Remote Sensing Magazine - December 2020 - 83

CONCLUSIONS AND SUMMARY
In the past decade, HSI FE has considerably evolved, leading to three main research lines (i.e., shallow UFE, shallow
SFE, and deep FE approaches) that include the majority of
FE techniques presented in this article. We systematically
provided a technical overview of the state-of-the-art techniques proposed in the literature by categorizing the aforementioned three focuses into subcategories. To make this
research article easy to follow for researchers at different
levels (i.e., students, researchers, and senior researchers),
we aimed to show the evolution of each category over the
decades rather than including many techniques with an exhaustive reference list.
The experimental section was designed to compare the
performances of the techniques in two ways: 1) between
all of the categories (i.e., shallow UFE, shallow SFE, and
deep FE approaches) and 2) within each category by analyzing the corresponding subcategories. In this manner, a
various subcategories were investigated, detailing the evolution of the shallow UFE (i.e., conventional data-projection schemes, band clustering/splitting techniques, lowrank reconstruction techniques, and manifold-learning
DECEMBER 2020

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

0.9
0.8
0.7
OA

RF on the spectral bands (labeled by HSI) and the features extracted by OTVCA and JPlay along with the OAs
obtained by CNN and PCNN. The results are mean values
over 10 experiments based on selecting the samples randomly. (The standard deviations are shown by the error
bars.) The outcomes of the experiment can be summarized as follows:
◗◗ The SFE technique (i.e., JPlay) improves the OAs compared to the spectral classifier. However, it provides a
much lower OA compared with UFE and deep FE for all
cases. Two aspects might explain this point. One is that
JPlay fails to model spatial and contextual information;
another is that, although JPlay attempts to enhance the
reorientation ability of the features via multilayered linear mapping, it is still incomparable to the nonlinear
deep-FE-based techniques, particularly when the number of samples is increased.
◗◗ In this experiment, the UFE technique (i.e. OTVCA) and
the deep FE method, CNN, performed similarly in terms
of classification accuracies. Compared with the results
given in Table 7, it can be observed that the random selection of the training samples over the entire class of regions from the ground reference considerably improves
the performance of RF applied on the features extracted
by OTVCA. This is often due to the lack of a parameter
selection technique to choose the optimum parameter
for the OTVCA algorithm, which could lead to oversmoothing on the features.
◗◗ The DL technique (i.e., PCNN), after using the reduction (i.e., PCA), provides very high OA for all the cases.
Comparing the results with CNN (i.e., without using the
PCA reduction) confirms the advantage of using the reduction stage prior to DL techniques.

0.6
0.5
0.4
0.3

20
40
60
80
Number of Training Samples
HSI

OTVCA

JPlay

CNN

100
PCNN

FIGURE 13. The classification accuracies with respect to the
number of the training samples on the Houston University 2018
data set. The results shown are means over 10 experiments, and
standard deviations are shown by the error bars.

techniques), shallow SFE (i.e., class-separation discriminant analysis, graph-embedding discriminant analysis,
regression-based representation learning, and JPlay), and
deep FE approaches (i.e., AE, CNN, RNN, and integrative
approaches). Three recent hyperspectral data sets were
studied, and the results were evaluated in terms of classification accuracies and the quality of the classification maps.
The experiments carried out in this study showed the
following, in terms of classification accuracies: 1) DL FE
techniques (i.e., CNN and PCNN) can outperform the
shallow methods, particularly when a sufficient amount
of training data are available; 2) applying a dimensionality
reduction step (such as PCA) prior to the DL techniques
considerably improves their performances; and 3) shallow
UFE techniques not only outperform the SFE methods but
also are very competitive compared with deep FE methods. However, we should mention that the conclusions are
limited by the experiments carried out on the three HSI
data sets. In addition, this article provides an impressive
amount of code and libraries, mostly written in Python
and MATLAB, to ease the task of researchers in this vibrant
field of research.
ACKNOWLEDGMENTS
We would like to thank Prof. Melba Crawford for providing the Indian Pines 2010 Data and the National Center for
Airborne Laser Mapping, the Hyperspectral Image Analysis Laboratory at the University of Houston, and the IEEE
GRSS Image Analysis and Data Fusion Technical Committee. This work is partially supported by an Alexander von
Humboldt research grant. We also would like to thank the
AXA Research Fund for supporting the work of Prof. Jocelyn Chanussot and the corresponding author of this paper,
Dr. Danfeng Hong.
83

IEEE Geoscience and Remote Sensing Magazine - December 2020

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