IEEE Geoscience and Remote Sensing Magazine - March 2017 - 26

(a)

(b)

(c)

(d)

(e)

(f)
Thematic Classes
Synthetic Grass

Tree

Soil

Residential

Commercial

Road

Highway

Parking Lot 1

Parking Lot 2

Tennis Court

Running Track

Healthy Grass

Stressed Grass

Water
Railway

FIGURE 13. Scenario 2: classification maps for Houston data using (a) RF, (b) SVM, (c) BP, (d) KELM, (e) MLR, and (f) 1-D CNN.

Regarding the classification accuracy, it can be seen that
the ELM achieves comparable results.
◗ SVM versus KELM: The computational complexity of the
SVM is much bigger than the KELM. It can be seen that
the KELM slightly outperforms the SVM in terms of classification accuracy. Experimental validation shows that
the kernel used in the KELM and SVM is more efficient
than the activation function used in ELM.
◗ BP versus ELM versus KELM: In light of the results, it can be
seen how the three versions of the SLFN provide competitive results in terms of accuracy. However, it should be
noticed that both the ELM and KELM are on the order of
hundreds or even thousands of times faster than the BP.
Actually, the ELM and KELM have a practical complexity of O (L3 + L2 n + (K + d) Ln) and O (2n 3 + (K + d) n 2),
respectively [149].
◗ SVM versus 1-D CNN: The main advantage of 2-D and 3-D
CNNs is that they use local connections to handle spatial dependencies. In this work, however, the 1-D CNN is taken
to have a fair comparison with
other spectral approaches. In
general, the SVM can obtain
DIFFERENT SOLUTIONS
higher classification accuraDEPEND ON THE
cies and work faster than the
COMPLEXITY OF THE
1-D CNN, so the use of SVMs
ANALYSIS SCENARIO AND
over the 1-D CNN is recomON THE CONSIDERED
mended. In terms of central
APPLICATION DOMAIN.
processing unit (CPU) processing time, deep-learning
methods are time consuming
in the training step. Compared to the SVM, the training
time of the 1-D deep CNN is about two or three times
26

longer than the RBF-SVM. On the other hand, the advantage of the deep CNN is that it is extremely fast on the
testing stage.
◗ MLR (executed via LORSAL) versus other methods: Some of
the MLR advantages are as follows: 1) It converges very
fast and is relatively insensitive to parameter settings. In
our experiments, we used the same settings for all data
sets and received very competitive results in comparison
with those obtained by other methods. 2) MLR has a
very low computational cost, with a practical complexity of O (d 2 (K - 1)).
For illustrative purposes, Figure 11 provides a comparison of the different classifiers tested in this work with the
Indian Pines and Pavia University scenes (in terms of OA).
As shown by Figure 11, different classifiers provide different performances for the two considered images, indicating that there is no classifier consistently providing the
best classification results for different scenes. The stability
of the different classifiers with the two considered scenes
is illustrated in Figure 12, which demonstrates how much
a classifier is stable with respect to some changes in the
available training sets. Furthermore, Table 6 gives detailed
information about the classification accuracies obtained by
different approaches in a different application domain, represented by the Houston data set. In this case, the optimized
classifiers also perform similarly in terms of classification
accuracy; so, ultimately, the choice of a given classifier is
more driven by the simplicity of tuning the parameters and
configurations rather than by the obtained classification
results. This is an important observation, as it is felt that
the hyperspectral community has reached a point at which
many classifiers are able to provide very high classification
ieee Geoscience and remote sensing magazine

march 2017

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