=Paper=
{{Paper
|id=Vol-2268/paper16
|storemode=property
|title=Evaluation of Nonlinear Dimensionality Reduction Techniques for
Classification of Hyperspectral Images
|pdfUrl=https://ceur-ws.org/Vol-2268/paper16.pdf
|volume=Vol-2268
|authors=Evgeny Myasnikov
|dblpUrl=https://dblp.org/rec/conf/aist/Myasnikov18
}}
==Evaluation of Nonlinear Dimensionality Reduction Techniques for
Classification of Hyperspectral Images==
https://ceur-ws.org/Vol-2268/paper16.pdf
Evaluation of Nonlinear Dimensionality
Reduction Techniques for Classification
of Hyperspectral Images
Evgeny Myasnikov1
Samara University, Moskovskoe Shosse 34, Samara, Russia, 443086
mevg@geosamara.ru
http://ssau.ru
Abstract. Nonlinear dimensionality reduction techniques are becom-
ing increasingly popular in the analysis of hyperspectral images. In this
work, some such methods are evaluated as a preliminary stage to the
classification of hyperspectral images. The list of methods to be studied
includes Isomap, Locally Linear Embedding, Laplacian Eigenmaps, Non-
linear Mapping, and also the linear principal component analysis tech-
nique. We study the performance of nonlinear methods with both the
Euclidean distance and the Spectral angle mapper (SAM) dissimilarity
measures. Analyzed methods are evaluated in terms of the classification
accuracy and runtime. The experiments are carried out using the well-
known hyperspectral scenes.
Keywords: Hyperspectral image, Dimensionality reduction, Isomap, Lo-
cally Linear Embedding, Laplacian Eigenmaps, Nonlinear Mapping, Prin-
cipal Component Analysis
1 Introduction
Hyperspectral images are three-dimensional arrays with two spatial dimensions
and one spectral dimension. A pixel in a hyperspectral image can be considered
as a vector containing a number (typically, up to a few hundred) of components
corresponding to different wavelengths.
Being compared to color or gray-scale images, hyperspectral images suggest
extended opportunities, for example, to detect materials in a depicted scene or
substantially improve the accuracy of classification. However, the use of hyper-
spectral images is accompanied by increased costs for storage, transmission, and
processing of such images. For this reason, an important task is to eliminate the
redundancy of such images, while maintaining the quality of the solutions to
applied problems.
The most widely used solution to the above task consists in the use of dimen-
sionality reduction techniques. As such a technique, in most cases, the principal
component analysis is used. Nevertheless, nonlinear dimensionality reduction
techniques are becoming increasingly popular in the last years.
In this work, some such methods are evaluated as a preliminary stage to the
classification of hyperspectral images. The list of methods to be studied includes
Isomap [5], Locally Linear Embedding (LLE) [7], Laplacian Eigenmaps (LE) [8],
Nonlinear Mapping (NLM) [3], and also the linear principal component analysis
technique [1].
We study the performance of nonlinear methods with both the Euclidean dis-
tance and the Spectral angle mapper (SAM) dissimilarity measure as they have
been used in hyperspectral image analysis most often. The reason for choosing
the above dimensionality reduction methods was the frequency of application
of these methods in the analysis of hyperspectral images and the possibility of
embedding the SAM measure.
Analyzed methods are evaluated in terms of the classification accuracy and
runtime. The experiments are carried out on the well-known hyperspectral scenes.
The paper has the following structure. Section 2 is devoted to the brief de-
scription of methods used in the paper. Section 3 describes the results of exper-
iments. The paper ends up with the conclusion.
2 Methods
In this study, we use the following dimensionality reduction techniques:
– Principal Component Analysis (PCA) technique [1] is the most well-known
linear dimensionality reduction technique, which is used in the wide range of
applications. This method searches for a linear projection into the subspace
of a smaller dimension that maximizes the variance of data.
– Nonlinear Mapping (NLM) method is based on the principle of preserving
the pairwise distances between datapoints. While the basics of this method
were developed in 1960-s in works by J.B. Kruskal [2] and J.W. Sammon [3],
here we use a different version of the method [10], which differs to the base
method in PCA-based initialization and stochastic gradient descent.
– Isomap method was introduced by J.B. Tenenbaum at al. in papers [4, 5].
The main idea of this method consists in the use of geodesic distances instead
of Euclidean distances in classical metric multidimensional scaling (MDS).
Here we use the Landmark Isomap method [6], which is a faster version of
this algorithm.
– Locally Linear Embedding (LLE) technique was introduced by S.T. Roweis
and L.K. Saul in the paper [7]. This technique is based on the idea that each
particular datapoint and its neighbors lie close to a locally linear patch of
the nonlinear manifold, and can be reconstructed as a linear combination of
its neighbors in both high-dimensional and embedding spaces.
– Laplacian Eigenmaps technique was introduced by M. Belkin and P. Niyogi
in the paper [8]. This technique is based on the eigenvalue decomposition of
the graph Laplacian matrix.
In all the above nonlinear techniques (except PCA), it is assumed that the
Euclidean distance is used as a dissimilarity measure. As we said in the Introduc-
tion, in this paper, we also embed the SAM measure [9] in the above nonlinear
techniques.
In particular, for the Nonlinear Mapping technique, we replace the calcula-
tion of Euclidean distances in hyperspectral space with the calculation of SAM
measures. According to [10], it means approximation of spectral angles by the
Euclidean distances in the embedding space (SAED technique). In the ISOMAP
technique we use SAM measures to construct the neighborhood graph, that is
to find neighbor points, and to initialize weights of edges. In the Locally linear
embedding method, we use SAM measures only to find neighbor points. In the
Laplacian Eigenmaps technique, we use SAM measures both to find neighbor
points, and to define the heat kernel.
3 Experiments
Datasets For the reported study, we used several well-known hyperspectral im-
age scenes [14], which supplied with groundtruth segmentation: Salinas, Indian
pines, Botswana, and Kennedy space center. In this paper we describe the exper-
imental results for two well-known scenes, namely, Salinas and Kennedy space
center (Figure 1).
Both hyperspectral image scenes were acquired using the AVIRIS sensor. The
first scene contains 512×217 pixels, and 224 spectral bands. In our experiments
we used the image containing 204 spectral bands, in which some spectral bands
were discarded due to a high noise and water absorption. As the Salinas scene
contains more than 100 thousand pixels, and it was necessary to perform a lot
of runs of nonlinear dimensionality reduction techniques, for our experiments
we used regularly sampled test image, which was masked with the provided
groundtruth image. The classified pixels of the groundtruth image are divided
into 16 classes.
Kennedy space center scene contains 512×614 pixels. The version containing
176 spectral bands was used in the experiments. The groundtruth image contains
information only on a small amount of pixels, so we applied groundtruth mask,
and did not use any sampling. The classified pixels of the groundtruth image are
divided into 13 classes.
Experimental setup. To perform the experiments we used PCA implementa-
tion provided with Matlab, C++ implementation of Nonlinear Mapping method,
and for LLE, Laplacian Eigenmaps and Isomap, we used Matlab Toolbox for Di-
mensionality Reduction [11].
A laptop based on Intel Core i7-6500U CPU 2.5 GHz, 12 Gb RAM was used
to perform experimental studies.
Evaluation. The k-Nearest Neighbor (k-NN) classifier and the Support Vector
Machine (SVM) were used in this study. To measure the quality of classifica-
tion we used the overall classification accuracy, defined as the proportion of the
correctly classified pixels of the test set.
Fig. 1. False grayscale images for test hyperspectral image scenes (contrasted): Salinas
(left) and Kennedy space center (right).
The whole set of ground truth samples was divided into a training (60 per-
cents) and a test (40 percents) subsets in our experiments. The dimensionality
of the reduced space ranged from 3 to 30.
Experimental results. The results of the experimental study for the NN clas-
sifier are shown in Figure 2. As it can be seen from the figure, the use of the
SAM measure was preferable for the Salinas and Kennedy space center hyper-
spectral scenes, and for almost all the considered nonlinear techniques, as SAM
provided a better quality of classification compared to Euclidean distance. This
observation is also confirmed for two other scenes involved in the experiments.
In all the considered cases, for the NN classifier, the best results were ob-
tained using the Nonlinear Mapping technique. The linear PCA technique pro-
vided similar or slightly worse results than the Nonlinear Mapping combined
with Euclidean distances. PCA outperformed LLE, LE, and Landmark Isomap
methods on Salinas and two other test hyperspectral scenes, except the Kennedy
space center scene.
In KSC scene, the linear PCA technique performed much worse than non-
linear techniques for the dimensionality of the reduced space up to 25. In this
scene, we can also observe that the Nonlinear Mapping technique in combina-
tion with Euclidean distances loses advantages over other nonlinear techniques
based on SAM measures for the dimensionality of the reduced space up to 20.
Thus, on the one hand, we see a significant advantage of using the SAM measure
over the Euclidean distance for this scene. On the other hand, we can explain
obtained results by substantially nonlinear properties of this dataset. It noted
earlier that the discrimination of land cover for the KSC scene is difficult due to
the similarity of spectral signatures for certain vegetation types [14].
Fig. 2. Dependence of the classification accuracy Acc for the 1-NN classifier on the di-
mensionality m for the Salinas (top) and Kennedy space center (bottom) hyperspectral
scenes.
The results of the experimental study for the SVM classifier are shown in
Figure 3. The experimental results showed that in many considered cases (some
results are not shown in the figures) the PCA was a preferable choice. The
Nonlinear Mapping performed a bit worse for the SVM classifier. But again,
the performance of the PCA technique was drastically reduced on the Kennedy
space center scene for the dimensionality of the reduced space up to 25. This
indicates the importance of the careful selection of the output dimensionality.
Fig. 3. Dependence of the classification accuracy Acc for the SVM classifier on the di-
mensionality m for the Salinas (top) and Kennedy space center (bottom) hyperspectral
scenes.
It is worth noting that three graph-based dimensionality reduction tech-
niques, namely Isomap, LLE, and Laplacian Eigenmaps have the mataparameter,
which defines the number of neighbors. In the above experiments, we used the
default value equal to 12. We found that these techniques are very sensitive to
the choice of this parameter. This was especially evident for the LLE method. In
some cases, the classification accuracy could be substantially (by some percent)
improved over the reported above values by the good choice of the considered
metaparameter. But for other cases, the same value could provide worse results.
In any case, we were not able to outperform the Nonlinear Mapping by varying
this parameter in the reasonable range from 10 to 100 with the step equal to 10.
The example dependency of the classification accuracy on the metaparameter k
is shown in Figure 4.
The same Figure 4 shows the runtime of the considered techniques. The
timing for the PCA technique is not shown as it was about 0.1 sec. that is negli-
gible compared to the considered nonlinear methods. The run time of Landmark
Isomap, LLE and Laplacian Eigenmaps is less than the run time of the Non-
linear Mapping technique for the minimum considered value of k, but it raises
fast with the growth of k. So the Nonlinear Mapping becomes faster for k > 30.
Moreover, there are approaches, which allow to speed-up this technique [12, 13].
It is worth noting, however, that timings depend hardly on the hardware and
implementation.
Fig. 4. Dependence of the classification accuracy Acc for the NN classifier on the
metaparameter k (left), and the dependence of the dimensionality reduction run time
on the metaparameter k (right) for the Kennedy space center hyperspectral scene.
4 Conclusion
In this paper, we studied several popular nonlinear dimensionality reduction
techniques in the task of per-pixel hyperspectral image classification. We showed
that the Nonlinear Mapping technique could be considered as a reasonable choice
when the nearest neighbor classifier is used.
While in many cases the combination of the PCA technique with SVM classi-
fier provides nice results, for complex hyperspectral scenes containing substantial
nonlinear effects the traditional PCA technique could be a bad choice. In such
cases, it is necessary to carefully choose the output dimensionality, and consider
the possibility of using the nonlinear dimensionality reduction techniques.
The main drawback of the nonlinear methods is their high computational
complexity, which is expressed by their long run time, which exceeds the runtime
of the PCA technique by orders of magnitude.
Acknowledgments The reported study was funded by RFBR according to the
research project no. 18-07-01312-a.
References
1. K. Fukunaga: Introduction to Statistical Pattern Recognition. London: Academic
Press, 2nd edn. (2003)
2. Kruskal J. B.: Multidimensional scaling by optimizing goodness of fit to a non-
metric hypothesis. Psychometrika, 29, P. 1-27. (1964)
3. Sammon, J.W.: A nonlinear mapping for data structure analysis. Transactions on
Computers. Vol. 18. P. 401-409. (1969)
4. J. Tenenbaum: in Advances in Neural Information Processing, 10. M. Jordan, M.
Kearns, S. Solla (Eds.). MIT Press, Cambridge, MA. P. 682-688. (1998)
5. J. B. Tenenbaum, V. de Silva, J. C. Langford: A Global Geometric Framework for
Nonlinear Dimensionality Reduction. Science, P. 2319-2323. (2000)
6. J. B. Tenenbaum, V. de Silva: Global versus local methods in nonlinear dimen-
sionality reduction. Advances in Neural Information Processing Systems, Vol. 15,
P. 705-712. (2003)
7. S. T. Roweis and L. K. Saul: Nonlinear Dimensionality Reduction by Locally Linear
Embedding. Science, Vol. 290. P. 2323-2326. (2000)
8. M.Belkin, P.Niyogi: Laplacian eigenmaps and spectral techniques for embedding
and clustering. Advances in Neural Information Processing Systems, 14. P. 585-
591.(2001)
9. Kruse, F.A., Boardman, J.W., Lefkoff, A.B., Heidebrecht, K.B., Shapiro, A.T., Bar-
loon, P.J., and Goetz, A.F.H.: The Spectral Image Processing System (SIPS) Inter-
active Visualization and Analysis of Imaging Spectrometer Data. Remote Sensing
of Environment, 44. P. 145-163. (1993)
10. Myasnikov, E.: Nonlinear Mapping Based on Spectral Angle Preserving Principle
for Hyperspectral Image Analysis. In: Felsberg M., Heyden A., Krger N. (eds)
Computer Analysis of Images and Patterns. LNCS 10425, pp. 416-427. (2017)
11. L.J.P. van der Maaten, E.O. Postma, and H.J. van den Herik. Dimensionality
Reduction: A Comparative Review. Tilburg University Technical Report, TiCC-
TR 2009-005. (2009)
12. Myasnikov, E.: The use of Interpolation Methods for Nonlinear Mapping. In:
Chmielewski L., Datta A., Kozera R., and Wojciechowski K. (eds) Computer Vi-
sion and Graphics. LNCS 9972, pp. 649-655. (2016)
13. Myasnikov, E.V.: Fast Techniques for Nonlinear Mapping of Hyperspectral Data.
Proc. SPIE 10341, Ninth International Conference on Machine Vision (ICMV
2016), 103411D. (2017)
14. Hyperspectral Remote Sensing Scenes. Available from: http://www.ehu.eus/
ccwintco/index.php?title=Hyperspectral Remote Sensing Scenes