I.J. Image, Graphics and Signal Processing, 2015, 5, 42-48
Published Online April 2015 in MECS (http://www.mecs-press.org/)
DOI: 10.5815/ijigsp.2015.05.05
Supervised Classification Approaches to Analyze
Hyperspectral Dataset
Sahar A. El_Rahman
Electrical Department, Faculty of Engineering-Shoubra, Benha University, Cairo, Egypt
sahr_ar@yahoo.com
Wateen A. Aliady
Princess Noura bint AbdAlrahman University, Riyadh, KSA
Wateen.Aliady@gmail.com
Nada I. Alrashed
Princess Noura bint AbdAlrahman University, Riyadh, KSA
nada12.ksa@gmail.com
Abstract—In this paper, Spectral Angle Mapper (SAM)
and Spectral Information Divergence (SID) classification
approaches were used to classify hyperspectral image of
Georgia, USA, using Environment of Visualizing Images
(ENVI). It is a software application used to process and
analyze geospatial imagery. Spatial, spectral subset and
atmospheric correction have been performed for SAM
and SID algorithms. Results showed that classification
accuracy using the SAM approach was 72.67%, and SID
classification accuracy was 73.12%. Whereas, the
accuracy of SID approach is better than SAM approach.
Consequently, the two approaches (SID and SAM) have
proven to be accurately converged in classification of
hyperspectral image of Georgia, USA.
Index Terms—Atmospheric Correction, Hyperspectral
image, Spectral Angle Mapper, Spectral Information
Divergence, Supervised classification.
I. INTRODUCTION
Hyperspectral imaging is spectral imaging technique
that is able to find materials, identify and distinguish
spectrally unique materials. This is done by collecting
and processing hundreds of contiguous narrow
wavebands from the scene, which provide spectral
information [1]. Collecting the information is done by
using an airborne or satellite sensor at a short, medium or
long distance from the scene [2].
The main advantage of hyperspectral image is the
potential to provide more accurate results than any other
type of remotely sensed data, because they commonly
collect more than 200 spectral bands to perform a
detailed information extraction in order to classify,
identify, and detect objects. [2][3][4].
In contrast to traditional multispectral sensors such as
AVHRR (Advanced Very High Resolution Radiometer)
that measures radiation reflected from a scene in three to
six spectral bands of data [4][5]. This small range of
Copyright © 2015 MECS
spectral bands is a primary disadvantage to multispectral
sensors [5].
The main disadvantage of hyperspectral images is the
need for sensitive detectors, fast computers, and
significant data storage capacity for processing
hyperspectral data. This led to an increase of the cost of
acquiring and processing hyperspectral data [6]. Some of
the practical applications for hyperspectral image
classification are: Agricultural, Traffic recognition,
Locate objects in satellite images, Medical.The majority
usage of hyperspectral imaging is for vegetation and
minerals extraction [8]. Classification is identified as the
Information extraction technique that is mostly based on
analyzing the spectral reflectance properties of the study
scene and performing certain algorithms designed for
spectral analysis [9]. It is known by the method that
group pixels with similar characteristics together in an
image and, indeed, the spectral pattern present within the
data for each pixel is used as the numerical basis for
classification. The objective of image classification is to
identify the features occurring in an image in terms of the
object or type of land cover these features actually
represent on the ground as shown in Fig.1.
Image classification is an important part of the remote
sensing, image analyzing and pattern recognition. It
forms a significant tool for digital images examination.
Image classification is perhaps the most important part of
digital image analysis. It is really nice to have a colorful
image, having a magnitude of colors illustrating various
features of the underlying terrain, but it is quite useless,
unless to know what the colors mean. The analyst must
choose a classifier that will accomplish the best for a
certain task. Now a days, it is difficult to state which
classifier is optimum for all situations as the
characteristic of each data set and the circumstances for
each study vary so greatly [10]. There are two main
approaches used in hyperspectral classification:
Supervised and Unsupervised [9].
I.J. Image, Graphics and Signal Processing, 2015, 5, 42-48
Supervised Classification Approaches to Analyze Hyperspectral Dataset
43
tool, sub setting, atmospheric correction, classifying is
given in Section III. The results obtained using this
methodology are presented and discussed in Section VI.
Section V, concludes and summarizes the observations
obtained by using this approach.
II. STUDY AREA AND DATA SET
The study area is located in Georgia, US.
Hyperspectral data was acquired on August 2009, using
hyperspectral data from EO-1 Hyperion system. The test
area covers about 1 km2. This area has a lot of vegetation
scene. It is downloaded from United States Geological
Survey (USGS), which is a scientific agency of the
United States government.
III. METHODOLOGY
First, Hyperion tool is applied on study dataset. Then,
preprocessing of hyperspectral dataset include Spatial,
spectral subset and atmospheric correction have been
performed. Finally, SAM and SID supervised
classification algorithms are applied.
A. Hyperion Tool
Fig. 1. Supervised classification
In supervised techniques, the training areas are used
which are homogeneous representative samples of the
different surface types of interest. All the spectral bands
of the pixels comprising these areas have numerical
information [11]. The algorithm assign some pixels to
information classes based on fieldwork, map analysis,
and personal experience. Then, the algorithm classifies
the rest pixels with unknown identities. The procedure
starts by the user selecting and naming areas on the image,
which correspond to the classes of interest. These classes
correspond to information classes. Then, the algorithm
will evaluate and assign unknown identity pixels to the
class that has the highest likelihood of being a member.
Unlike the unsupervised classification, that depends on
algorithms with statistically determined criteria to
automatically organize pixels into unique groups with
similar spectral characteristics [12].
There are number of supervised approaches that have
been developed to tackle the hyperspectral data
classification
problem.
Each
giving
different
classification
accuracy.
Two
approaches
are
demonstrated in this work to compare their accuracy
results, which are Spectral Angle Mapper (SAM), and
Spatial Information Divergence (SID).
The paper is organized as follows. The study area and
data set used in our work is given in Section II. The steps
followed in our work in sequence are: applying Hyperion
Copyright © 2015 MECS
Hyperion tool is used to convert Geo TIFF datasets
into ENVI format files that contain wavelength, and band
information [12] [13]. The study dataset was in a form of
242 files with .TIF extension each representing a certain
band so these bands has to be collected to form one
image having all bands using this tool.
B.
Preprocessing of Hyperspectral Data (Spatial and
Spectral Subset)
It is often mandatory to perform the preprocessing on
hyperspectral data to extract useful information from
scene. This utilizes the processor by only processing
needed data for the study area and improves the
classification performance in hyperspectral imagery [15].
The data has been subjected to spatial and spectral subset
to extract unwanted information.
Spatial Subset
Performing image spatial sub setting is resizing the
hyperspectral image to any size or aspect ratio by using
ordered cutting that is focused on selecting the area of
study in a square shape.
Spectral Subset
It is based on identifying bad bands the ones that will
not help in the study area. It will only cause over
processing on the processor. ENVI headers may have
associated information for the bad bands list. Mostly first
and last bands are bad in hyperspectral images. The study
area image contains 242 bands, but after the elimination
of bad bands it only have 155.
C. Atmospheric correction
I.J. Image, Graphics and Signal Processing, 2015, 5, 42-48
44
Supervised Classification Approaches to Analyze Hyperspectral Dataset
The atmosphere particles reduces the amount of
incoming energy from the sun reaching Earth’s surface
and further reduce the amount of reflected energy
reaching the sensor. Therefore, the energy reached to the
sensor may be changed due to atmosphere interaction
with incoming and reflected solar energy. So, little
information would be gained from the scene.
Atmospheric correction attempts to minimize these
effects on image spectra, so, it must be applied to correct
the image of the effect of atmospheric gases, and through
the use of ENVI it can correct the captured image of the
effects of atmosphere [15].
QUAC (QUick Atmospheric Correction) is an
approach for a sophisticated atmospheric correction,
whereas, from the information contained within the scene
the parameters directly are determined. The QUAC
method is one of the best atmospheric correction methods,
because it has a user-friendly interface, extremely
accurate, and significantly fast [15].
D. Supervised Classification Approaches
In the processing phase, classification is applied on
corrected image, using two classification approaches
SAM and SID.
SAM Classification Approach
In SAM Approach the spectral similarity between two
spectra is computed. This is by calculating the angle
between each pixel spectrum and each target spectrum.
The smaller the angle is, the more likely to belong to the
reference spectra. It treats the two spectrum as vectors,
not taking into account their magnitude. This technique is
relatively insensitive to changes in pixel illumination
because increasing or decreasing illumination doesn’t
change the direction of the vector, only its magnitude.
Endmember spectra is extracted directly from the study
area image using the library USGS library by selecting
Endmembers of interest. So, it will compare the spectral
signature for each pixel in study dataset to the spectral
signature of selected vegetation Endmember in the library
[6],[7] see Fig. 2.
Spectral Information Divergence
Fig. 2. SAM classification approach
Spectral Information Divergence (SID) is a spectral
classifier that uses a divergence measure to match pixels
to reference spectra. The more likely the pixels are
similar, the smaller the divergence. Pixels are not
classified when they with a measurement greater than the
specified maximum divergence threshold. SID measures
spectral variability of a single mixed pixel from a
probabilistic point of view [16], [17].
Endmember spectra are extracted directly from the
study area image using the USGS library by selecting
Endmembers of interest. The divergence of spectral
signature is calculated for each pixel in study area to the
spectral signature of a selected vegetation Endmember in
the library see Fig. 3.
Copyright © 2015 MECS
I.J. Image, Graphics and Signal Processing, 2015, 5, 42-48
Supervised Classification Approaches to Analyze Hyperspectral Dataset
Fig. 3. SID classification approach
IV. RESULTS AND DISCUSSION
The classification result of study dataset using SAM
approach is shown in Fig.4. Table 1 shows SAM
classification of the entire study area image where Bay
Laurel is the mostly found vegetation in this area of
Georgia, whereas, 4.172% of the image is classified as
Bay Laurel. Also, there were no identification for
Chamise (Flower), Chamise (Green), and Coast Redwood
(Green). The classification result for the study dataset
after applying SID is represented in Fig.5. Table 2 shows
SIM classification. I shows that Jasper Ridge Serpentine
is the mostly found vegetation in this area of Georgia,
whereas, 41.534% of the image is classified as Jasper
Ridge Serpentine. Also, there were no identification for
California Valley Oak, and Coast Redwood (Green). In
order to differentiate statistically between the two
classifications results, accuracy assessments has to be
performed on SAM and SID to differentiate between their
classifications results.
The correctness of classified images is determined by
the accuracy assessments. The correlation between a
standard that is assumed to be correct and an image
classification of unknown quality is considered as the
Copyright © 2015 MECS
45
measure of accuracy. So, at the beginning, the
verification samples must be stated, which are used in
ENVI as a standard for the accuracy assessments of the
classifications performed. Then, generate random
sampling for them. Finally, calculate the accuracy
assessment by the confusion matrix.
The verification samples were chosen by using certain
pixels that has spectral signature with a close match to the
spectral signature of materials used in this work
classification, which are found in the USGS spectral
library, as shown in Fig.5, the spectral signature for
Leather Oak in the USGS Spectral library, which is
presented with the color green. This signature is used as a
reference spectral to find pixels in study area image that
has a spectral signature close to it, where the closest
match was for pixels having the spectral signature
colored in green, as shown in Fig.6.
Also, Red Willow sandstone verification samples were
generated following the same method, where the spectral
signature for Red Willow in the USGS Spectral library, is
shown in Fig.5, and the closest match to it was for pixels
having the spectral signature, as shown in Fig.6. Using
region of interest (ROI) Tool, the verification samples on
the original hyperspectral image are drawn manually.
After that, a random sample is generated, which is used
to find pixels in the image that has a spectral signature
with a close match to the spectral signature of ROI pixels,
because it is helpful and can be valuable in supporting
classification accuracy assessments. The stratified
random sampling is used, also called proportional or
quota random sampling. It involves dividing the
population (all of the ROIs) into homogeneous subgroups
(the individual ROIs) then taking a simple random
sample in each subgroup. The used sampling technique
proportionate, which means the sampling produces
sample sizes that are directly related to the size of the
classes (that is, the larger the class, the more samples will
be drawn from it).
Finally, , the Confusion Matrix is used to show the
accuracy of a classification by comparing a classification
result with ground truth information. A confusion matrix
is calculated using ground truth ROIs previously
determined. Table 3 shows the Confusion Matrix for
SAM, and Table 4 shows the confusion matrix of SID.
The overall accuracy for SAM is 72.67%, and 73.12
for the SID. So, the SID has given a better classification
for the study area image.
V. CONCLUSIONS
In this paper, the potential use of SAM, and SID
classifiers combined with the EO-1 Hyperion imagery
analysis for deriving total vegetation is achieved. They
are applied in a test site representative in study area in
Georgia, USA, as that is one of the famous vegetation
areas. SID and SAM approaches use the same set of
training and validation points selected over the acquired
EO-1 Hyperion imagery, which allowed a direct
comparison of their performance. The overall accuracy
was reported as 72.67% for the SAM classification
I.J. Image, Graphics and Signal Processing, 2015, 5, 42-48
46
Supervised Classification Approaches to Analyze Hyperspectral Dataset
approach, and 73.12% for the SID classification approach.
SID has a better result on study area image, because
SAM approach is insensitive since it depends on the
spectrum direction and not the length of the spectra
unlike the SID that measures the discrepancy between
each pixel spectrum and a reference spectrum.
Fig. 6. Spectral signature for leather oak, and Red Willow in the USGS
Spectral library
Fig. 4. Classified image using SAM classification approach.
Fig. 7. Spectral signature for pixels in study area image that has a close
spectral signature to leather oak, and Red Willow signature in USGS
Spectral library.
ACKNOWLEDGMENT
Fig. 5. Classified image using SID classification approach.
Copyright © 2015 MECS
We would like to thank our families for encouraging
and supporting us. Also thank the people who have been
instrumental in the successful completion of this work.
The editing and comments of the reviewers is gratefully
appreciated.
I.J. Image, Graphics and Signal Processing, 2015, 5, 42-48
47
Supervised Classification Approaches to Analyze Hyperspectral Dataset
Table 2. Classification Of The Entire Study Area Image By SID
Table 1. Classification Of The Entire Study Area Image By SAM
Class
Points
Percent
Area
Unclassified
7,883
3.9%
7,085,522.2 m2
Arroyo Willow
Bay Laurel
Blue Oak
1,169
20,418
198
0.593%
10.350%
0.100%
1,050,738.98 m2
8,352,428.30 m2
177,969.4781 m2
0
0.000%
0.0000 m2
C. Buckeye
4,315
2.187%
3,878,476.2 m2
Chamise
(Flower)
1,070
0.542%
961,754.2502 m2
0.0000 m2
Chamise (Green)
2,559
1.297%
2,300,120.6 m2
0.00%
0.0000 m2
Coast Redwood
(Dry)
9,024
4.574%
8,111,093.7 m2
3
0.002%
2,696.5072 m2
Coast Redwood
(Green)
0
0.000%
0.0000 m2
0
0.000%
0.0000 m2
9,229
4.678%
8,295,355.1 m2
2
0.001%
1,797.67 m2
Common Buck
Bush
363
0.184%
326,277.3 m2
Common Buck
Bush
3
0.002%
2,696.5072 m2
Coyote Bush 1
200
0.101%
179,767.1496 m2
481
0.244%
432,339.9947 m2
61
0.031%
54,828.98 m2
Coyote Bush 2
Coyote Bush 1
1,011
0.512%
908,722.9411 m2
184
0.09%
165,385.7 m2
Dove Weed
Coyote Bush 2
Dove Weed
681
0.345%
612,107.14 m2
Dry Grass
9,095
4.610%
8,174,911.1 m2
Dry Grass
840
0.426%
755,022.02 m2
Leather Oak
22,022
11.163%
19,794,160.8 m2
4,305
2.182%
3,869,487.8 m2
Live Oak
9
0.005%
8,089.5217 m2
Live Oak
26
0.013%
23,369.72 m2
Madrone
2
0.001%
1,797.6715 m2
Madrone
394
0.200%
354,141.2 m m2
Red Willow
953
0.483%
856,590.46 m2
Red Willow
610
0.309%
548,289.8 m2
Toyon
20
0.010%
17,976.71 m2
Toyon
5
0.003%
4,494.1 m2
2,414
1.224%
2,169,789.495m2
Tarweed
0
0.000%
0 m2
10,597
5.372%
9,524,962.4 m2
17
0.009%
15,280.2 m2
4,021
2.038%
3,614,218.5 m2
8,288
4.201%
7,449,550.6 m2
81,939
41.534%
73,649,702.3 m2
Class
Points
Percent
Area
169,176
85.7%
152,061,436.4 m2
Arroyo Willow
1,256
0.6%
1,128,937.69 m2
Bay Laurel
8,230
4.172%
482
0.244%
433,238.8 m2
75
0.038%
67,412.6 m2
357
0.181%
320,884.3 m2
Chamise
(Flower)
0
0.00%
Chamise (Green)
0
Coast Redwood
(Dry)
Unclassified
Blue Oak
California
Valley Oak
C. Buckeye
Coast Redwood
(Green)
Coast Sage
Leather Oak
Jasper Ridge
Butano andstone
Jasper Ridge
Grassland Soil
Jasper Ridge
Gravel
Jasper Ridge
Serpentine
1,353
0.686%
California Valley
Oak
Coast Sage
Tarweed
2
1,216,124.7 m
2
2,827
1.433%
2,541,008.6 m
6,393
3.241%
5,746,256.9 m2
Jasper Ridge
Butano
Sandstone
Jasper Ridge
Grassland Soil
Jasper Ridge
Gravel
Jasper Ridge
Serpentine
Table 4. Confusion Matrix For Sid Classification Approach
Table 3. Confusion Matrix For Sam Classification Approach
Reference data
Reference data
Classified data
leather oak
Red Willow
Classified data
leather oak
Red Willow
leather oak
100
89.28
leather oak
17.11
0.75
Red Willow
0.0
10.31
Red Willow
0.01
2.14
Overall Kappa Statistic:
Copyright © 2015 MECS
Overall Kappa Statistic:
0.5575
0.1354
I.J. Image, Graphics and Signal Processing, 2015, 5, 42-48
48
Supervised Classification Approaches to Analyze Hyperspectral Dataset
REFERENCES
[1]
Lamyaa G. Taha , Atia A. Shahin, "Assessment of
Cartographic potential of airborne hyperspectral data for
large scale mapping", Recent Advances in Image, Audio
and Signal Processing, wseas2013, PP 143-153. EGYPT,
2013. ISBN: 978-960-474-350-6. http://www.wseas.us/elibrary/conferences/2013/Budapest/IPASRE/IPASRE19.pdf.
[2] U. Heiden, W. Heldens, S. Roessner, K. Segl, T. Esch and
A. Mueller, "Landscape and Urban Planning in Urban
structure type characterization using hyperspectral remote
sensing and height information", Elsevier, pp 361-375,
2012.
[3] Stefan A. Robila, Andrew Gershman, " Spectral Matching
Accuracy in Processing Hyperspectral Data", IEEE, 07803-9029-6/05, 2005. PP 163-166.
[4] Schurmer,
"Air
Force
Research
Laboratories
Technology," J.H., U.K., 2003.
[5] S. J. Purkis and V. V. Klemas, "Remote Sensing and
Global Environmental Change," Remote Sensing and
Global Environmental Change, vol. 3, no. 10, 2011.
[6] Chein-I Chang, “ An Information-Theoretic Approach to
Spectral Variability, Similarity, and Discrimination for
Hyperspectral Image Analysis”, IEEE TRANSACTIONS
ON INFORMATION THEORY, VOL. 46, NO. 5,
AUGUST 2000, 1927-1932.
[7] S. Rashmi, A. Swapna, Venkat and S. Ravikiran,
“Spectral Angle Mapper Algorithm for Remote Sensing
Image Classification,” Spectral Angle Mapper Algorithm
for Remote Sensing Image Classification, vol. 1, no. 4, pp.
201-205, 2014.
[8] Belkacem Baassou, Mingyi He,Shaohui Mei, Yifan Zhang,
"Unsupervised Hyperspectral Image Classification
Algorithm By Integrating Spatial-Spectral Information",
ICALIP 2012, 978-1-4673-0174-9/12, ©2012 IEEE, PP
610-615.
[9] S. Pignatti, M. R. Cavalli, V. Cuomo, L. Fusilli, M.
Poscolieri and San, "Remote Sensing of Environment,"
Evaluating Hyperion capability for land cover mapping in
a fragmented ecosystem: Pollino National Park, Italy,, vol.
3, no. 12, pp. 622-634, 2009.
[10] Khamael Abbas, Mustafa Rydh, “Satellite Image
Classification and Segmentation by Using JSEG
Segmentation Algorithm”, I.J. Image, Graphics and Signal
Processing, Copyright © 2012 MECS, 10, pp 48-53.
http://www.mecs-press.org/.
[11] J. Senthilnath, Nitin Karnwal, D. Sai Teja, “Crop Type
Classification Based on Clonal Selection Algorithm for
High Resolution Satellite Image”, Image, Graphics and
Signal Processing, Copyright © 2014 MECS, 9, pp 11-19.
http://www.mecs-press.org/.
[12] Samuel Rosario Torres, "Implementation of the SVDSS in
the ENVI/IDL Environment”, vol. 15, no. 12, 2002.
[13] D. White, “Hyperion Tools 2.0 Installation and User
Guide, 2013.
[14] Jinguo Yuan, Zheng Niu, “Classification Using EO-1
Hyperion Hyperspectral and ETM+ Data”, Fourth
International Conference on Fuzzy Systems and
Knowledge Discovery (FSKD 2007), IEEE, 0-7695-28740/07, 2007.
[15] P. Shipper, “Introduction to Hyperspectral Image
Analysis,” An International Electronic Journal, 2003.
http://spacejournal.ohio.edu/pdf/shippert.pdf.
[16] Chein-I Chang, “ Spectral Information Divergence for
Hyperspectral Image Analysis”, IEEE, 0-7803-5207-6/99,
1999,pp 509-511.
[17] E. Zhang, X. Zhang, Y. Shuyuan and W. Shuang,
“Improving Hyperspectral Image Classification Using
Spectral
Information
Divergence,”
Improving
Hyperspectral Image Classification Using Spectral
Information Divergence, vol. 11, no. 1, pp. 249-253, 2013.
Authors’ profiles
Sahar Abd El_Rahman was born Cairo, Egypt, B.Sc.
Electronics & communication, Electrical Engineering
Department. Benha University, Shoubra Faculty of Engineering,
Cairo-Egypt. M.Sc. in an AI Technique Applied to Machine
Aided Translation, Electronic Engineering, Electrical
Engineering Department, Benha University, Shoubra Faculty of
Engineering, Cairo-Egypt, May2003. PHD. in Reconstruction
of High-Resolution Image from a Set of Low-Resolution
Images, Electronic Engineering, Electrical Engineering
Department, Benha University, Shoubra Faculty of Engineering,
Cairo-Egypt in Jan2008.
She is assistant professor from 2008 till now at Electrical
Engineering Department, Faculty of Engineering, Shoubra,,
Benha University, Cairo, Egypt. She was a lecture in the same
location from 2003 and instructor in the same location in 1998.
Her research interests include computer vision, digital image
processing, Signal processing, Robotics and Networks.
Wateen A. Aliady received B.Sc. degree in Computer Science
from College of Computer Science and Information, Princess
Noura Bint Abdul Rahman University in 2014, Saudi Arabia.
Nada I. Alrashed received B.Sc. degree in Computer Science
from College of Computer Science and Information at Princess
Noura bint AbdulRahman University in 2014,Saudi Arabia, and
now I’m working as teacher in Computer Sciences department
in Princess Noura bint AbdulRahman University.
How to cite this paper: Sahar A. El_Rahman, Wateen A. Aliady, Nada I. Alrashed,"Supervised Classification
Approaches to Analyze Hyperspectral Dataset", IJIGSP, vol.7, no.5, pp.42-48, 2015.DOI: 10.5815/ijigsp.2015.05.05
Copyright © 2015 MECS
I.J. Image, Graphics and Signal Processing, 2015, 5, 42-48