WASJ

WASJ
WASJ
WASJ

Fuzzy Based Hyperspectral Image Classification in Ridgelet Domain

K. Kavitha, S. Arivazhagan and B. Surya

Mepco Schlenk Engineering College Sivakasi, India- 626005

Abstract: Ridgelet transform is useful to handle the line singularities of the hyperspectral images and thereby
classifying various land-cover classes of hyperspectal images better than other multi resolution transforms such
as wavelet transforms which can handle point singularities alone. By applying ridgelet transform on the
hyperspectral data, it is able to extract both spatial and spectral information. To analyse the textural features,
Grey Level Co-occurrence matrix based feature extraction has been entertained on the ridgelet co-efficients and
classification has been done using Support Vector Machines and classification accuracies have been recorded.
To increase the accuracies, it has been decided to identify the mixed pixels and and to re-classify them.
Hence it has been proposed to use an edge detection to identify the mixed pixels and a fuzzy model, to provide
the proper class membership to the mixed pixel and therefore the mixed pixel can be treated as the pure pixel.
Re-classification is done on the identified mixed pixels with their new memebership and accuracies have been
observed again. The Overall accuracies of the algorithms, before and after applysing fuzzy model has been
compared. The proposed algorithm has been experimented on Airborne Visible Infra-Red Imaging Sensor
(AVIRIS) data of Indian Pine Site and Reflective Optics System Imaging Spectrometer (ROSIS) data of
University of Pavia. The results witness that the overall accuracy of 94.83% and 98.66% for Indian Pines and
University of Pavia, respectively, while using only 5% of the samples in each class were used for training.

Key words: Hyperspectral Images Ridgelet Transform Edge Detection Fuzzification SVM

INTRODUCTION quantization method for remote sensing data analysis [6].

As co-occurrence features provide the spatial information
Hyperspectral image classification algorithms yield about the image, it is decided to combine the advantages
good accuracies by adding spatial features along with the of ridgelet transforms and co-occurrence features and
spectral features. While classifying the sub classes of any hence the co-occurrence features are extracted in the
super class, the spectral features are not sufficient. So, it ridgelet domain. If all the extracted features are used for
is necessary to extract the spatial and spectral features classification, the computation time will be increased.
simultaneously to classify the land cover classes [1]. So, it is necessary to select the best set of features,
Wavelet transform based spectral feature extraction among the extracted features [7]. As Support Vector
algorithms are proposed in various works [2]. But wavelet Machines proved their classification efficiency on
transforms can deal with the point singularities, not the Hyperspectral data and it was reported in many literatures,
lone singularities. But natural images such as it is decided to use SVM for classification [8]. To improve
multispectral and hyperspectral images exhibit line the classification accuracy further, mixed pixels need to be
singularity properties not point singularities. Candès and identified, as they are the pulling agents of the
Donoho developed the type of multi resolution transforms classification rate. To identify the mixed pixels, edge
called as ridgelets; which can overcome the shortcomings detection is incorporated as the mixed pixels are present
of wavelets in higher dimensions and proved their ability at the boundaries between the classes. Pure pixels have
to deal with line singularities [3]. Do and Vetterli proposed single class membership whereas mixed pixels show more
an algorithm for image representation using ridgelet than one class membership due to the reflectance
transforms [4] while Zhang et.al. used the ridgelet characteristics of multiple classes which are closer to each
transforms for noise removal [5]. Grey-level Co-occurrence other at the land covers. But one of the reflectance is
Matrix (GLCM) method is a well-known texture predominant and that class value must be assigned to the

Corresponding Author: K. Kavitha, Mepco Schlenk Engineering College Sivakasi, India- 626005. E-mail: [email protected].
1
 WASJ

mixed pixel. By this way, mixed pixel tend to be identified size 256 x 256. After three level decomposition of ridgelet
as the pure pixel. To calculate such class membership of transformation, four sub-bands are obtained as shown in
the mixed pixel and to assign the single membership to Figure 2. In the decomposed image the sub-band 1
mixed pixel, fuzzy min-max rule based algorithm is used. contain the approximation coefficients while the sub-
For these identified and membership changed pixels, bands 2, 3, 4 contain the detailed coefficients.
re-classification is done and accuracies are recorded.
Indian Pines test site and University of Pavia dataset were Feature Extraction: As features are the significant
used to evaluate the algorithm. Experiment results show representatives of an image, they can able to differentiate
that the proposed method produces better performance various land cover classes present in the image. The
even while only 5% of the samples are used for training. statistical and Co-occurrence features such as Mean,
Rest of the paper is organised as follows: Section-2 deals Standard Deviation, Energy, Contrast, Homogeinity,
with the Proposed Work. Section-3 is dedicated to Maximum Probability were extracted for all the pixels in the
Experiment Setup followed by Section-4 which elaborates individual sub-bands and they exhibit the inter pixel
Results and Discussions. Section–5 gives the Conclusion relationship. The features are calculated using the
about the work. formulas given in Table 1. The extracted features provide
the characteristics of the input pixels to the classifier
Proposed Method: Hyperspectral data cube is which is necessary for any classifier to make a decision.
decomposed by ridgelet transform and ridgelet co- For each spatial location (i, j), ridgelet co-efficients are
efficients were stored. Using these co-efficients, Grey obtained from the bands of the hyperspectral image and
level Co-occurrence Matrix was formed and textural features listed in Table.1 were extracted. Eight statistical
features were calculated. Extracted features were and co-occurence features for four sub-bands were
concatenated and are used for classification and extracted. All these extracted thirty two features were
accuracies are recorded. Then mixed pixels are identified concatenated together and classification was done using
and re-classification is done and accuracies are recorded. it.
Figure 1 gives the method to get Ridgelet co-efficients.
Re-Classification: To reduce the computational
Ridgelet Transforms: Ridgelet transform provides the complexity, it is decide to re-classify the identified mixed
structural information of an image in the frequency pixels alone. For identifying the mixed pixels, conventional
domain using multiple radial directions, whereas the edge detection is used, as mixed pixels are present in
traditional wavelet transform provides limited directional edges. Canny edge detector is used and mixed pixels are
information only. The process of Digital Ridgelet identified.
Transformation is shown in Figure 2. Input hyperspectral
data is transformed into 3D Fourier Domain and hence the Fuzzy Based Membership Assignment: For any
resulting transformed function is in the frequency domain. hyperspectral image, the number of land cover classes
These transformed frequency values are converted in a present is assumed as the fuzzy subsets while the number
polar grid, which pass through the origin and spread of bands are assumed as the fuzzy sets. As reported in [8],
uniformly in angle. The gridding process is the Fuzzy sets and subsets can be defined by its membership
conversion of the Fourier transformed image from function fb,c(xb) where xb is the grey level of X in band b.
Cartesian to polar grid, that is obtained by interpolation. The pixel vector X, in B-dimensional space is;
Then the Radon projections are obtained by applying one
dimensional inverse Fourier transform. The polar grid is X = [ x1, x2 ,....xb ....xB ]T (1)
replaced with a pseudo-polar one. Concentric squares of
linear growing sides are used to replace the concentric The fuzzy requirement is;
circles of linearly growing radius. The rays are spread
N
uniformly in slope but not in angle. This appears like the
polar grid and for this grid a direct Fast Fourier Transform
∀xb =∈ [0,255], ∑ i =1 fb,i ( xb ) =
1 (2)

can be performed with no interpolation. One-dimensional N is the number of fuzzy subsets (Land Cover Classes)
wavelet transform is applied along the radial variable in The membership function of class ‘c’ in band ‘b’ is
Radon space to complete the Ridgelet transform. Figure 2.
shows the band structure of ridgelet transformed image of
(
fb,c ( xb ) = exp −( xb − b ,c )
2
|2 2
b,c ) (3)

2
 WASJ

Fig. 1: Block diagram for implementation of Digital Ridgelet Transform

Fig. 2: Band structure of ridgelet transformed image

Table 1: Extracted Statistical and Co-occurrence Features

where µb,c is the mean of class c in band b. The The overall reclassification process algorithms has
fuzzification process computes for a given pixel the the following steps:
membership degrees for each class c and band b from the
membership functions fully defined using the Gaussian Identify the Mixed pixels
distribution and the results of the statistics extraction. Decide Fuzzy Sets and Fuzzy Sub Sets
We obtain a matrix of fuzzy inputs Fi,p of order B* N. Record the grey level of the identified mixed pixel ‘x’
where, N is the number of classes and B the number in band ‘b’.
of bands; Calculate mean grey level value and standard
 f 1,1( x1 )  f 1, N ( x1 )  deviation b,c of class ‘C’ in band ‘b’
  Calculate membership function for mixed pixel using
fi , p =     
 f B ,1 ( xB )  f B, N ( xB )  (3)
 
(4) Identify the class with high membership value.

3
 WASJ

Replace the original grey level value of mixed mean the propose method for Indian Pines Site and University
value of the identified class of Pavia respectively. For the above mentioned datasets
Repeat the steps iv to ix for all mixed pixels. the classification maps before and after selecting the
Re-Classify. features are shown in Figure 3 and Figure 4 respectively.

Experiment Design: The proposed method was evaluated Table 2: Obtained Accuracies (in %) of Indian Pines dataset

by making use of Indian Pine Dataset and University of Sl.No Class Without Fuzzy With Fuzzy
1 Alfalfa 85.19 94.44
Pavia Dataset. Indian Pines dataset has 220 bands.
2 Corn Notill 94.91 91.00
Among 220 bands, forty bands were excluded as they are
3 Corn Mintill 92.21 84.05
suffered by water absorption and noise and the remaining 4 Corn 73.08 92.74
180 bands were used to test the algorithm. University of 5 Grass pastures 87.32 85.31
Pavia dataset contains 9 Land cover classes captured with 6 Grass Trees 98.13 89.83
the geometrical resolution of 1.3m. The dataset has 103 7 Grass pasture Mowed 100 84.63
8 Hay windowed 99.59 91.62
bands with the size of 610x610 pixels each. Twenty bands
9 Oats 55.00 75.00
are discarded, as they are not having information and 10 Soybean Notill 97.52 91.01
others are used for the experiment. Eight statistical and 11 Soybean Mintill 92.67 91.33
co-occurrence features were extracted from the four 12 Soybean clean 85.83 90.55
different sub-bands of the transformed ridgelet domain. 13 Wheat 94.34 89.62
Hence the feature set contains 32 features. All these 14 Woods 94.20 86.17
15 BGTD 83.42 90.26
features were concatenated and accuracies were recorded.
16 Stone-Steel-Towers 51.58 92.63
The concatenated features were used for classification Overall Accuracy (%) 92.33 94.83
and results were observed. Then, fuzzy based algorithm
re-classification is done. The performance of the algorithm Table 3: Obtained Accuracies (in %) of University of Pavia Dataset
was tested with randomly selected 5% of the labelled Sl. No. Class With out Fuzzy With Fuzzy
training samples and all the samples were used for testing. 1 Asphalt 94.60 96.73
The selected best features have the potential to show the 2 Meadows 98.63 94.23
3 Gravel 97.90 90.14
changes in space, spectrum and joint spatial-spectral
4 Trees 95.50 80.68
domains of hyperspectral data. Support Vector Machine 5 Paintmetalsheets 98.51 64.31
(SVM) was used as the classifier. MATLAB package is 6 Basesoil 86.70 85.92
used for conducting the experiment. 7 Bitumen 96.92 97.89
8 Selfblockingbricks 96.58 95.60
9 Shadows 98.52 97.47
RESULTS AND DISCUSSION
Overall accuracy(%) 94.08 98.66

The observed Overall Accuracies using ridgelet

features are 92.33% and 94.08% for Indian Pines and
University of Pavia data sets respectively. The small
classes like Oats and Grass Pasture Mowed victim low
pixelwise accuracies in case of Indian Pines. To improve
the overall accuracies further, fuzzy based re-classification
is entertained. The overall accuracies are increased while
mixed pixels are re-classified. For Indian Pines and
University of Pavia data sets, the overall accuracies are
94.83% and 98.66% respectively while using 8x8
neighborhood. As, Indian Pines Site contains, several
classes which are present very nearby 4x4 neighborhood
produced poor result. For most of the classes, the Fig. 3: Classification Maps of Indian Pines using ridgelet
piecewise accuracies are also lower. But while using transform (a) without fuzzy process (b) with fuzzy
Table 1 and Table 2 presents the accuracies, produced by process

4
 WASJ

REFERENCES

1. Li, C.H., B.C. Kuo, C.T. Lin and C.S. Huang, 2012.
A Spatial–Contextual Support Vector Machine for
Remotely Sensed Image Classification. IEEE Trans on
Geosci and Remote Sens, 50(7): 784-99.
2. Jia, S., Y. Qian, J. Li, W. Liu and Z. Ji, 2010. Feature
extraction and selection hybrid algorithm for
hyperspectral imagery classification in Proc IGARSS,
pp: 72-75.
3. Daugman, G.J., 1988. Complete discrete 2-D Gabor
transforms by neural networks for image analysis and
compression. IEEE Trans. Acoust. Speech Signal
Process, 36: 1169-1179.
4. Linlin, S. and S. Jia, 2011. Three-dimensional Gabor
wavelets for pixel-based hyperspectral imagery
Fig. 4: Classification Maps of University of Pavia using classification. IEEE Trans. Geosci. Remote Sens,
ridgelet transform (a) without fuzzy process (b) 49(12): 5039-5046.
with fuzzy process 5. Field, J.D., 1987. Relations between the statistics of
natural images and the response properties of cortical
CONCLUSION cells. J. Opt. Soc. Am. A, 4: 2379-2394.
6. Arróspide, J. and L Salgado, 2013. Log-Gabor filters
Feature extraction in ridgelet domain and fuzzy based for image-based vehicle verification. IEEE Trans.
re-classification have been presented. Mixed pixel problem Image Process, 22: 2286-2295.
has been alleviated using Fuzzy based process. 7. Shijin, L., W. Hao, W. Dingsheng and Z. Jiali, 2011.
Experiments were conducted on AVIRIS, ROSIS An Effective Feature Selection Method for
hyperspectral data the observed Overall Accuracy were Hyperspectral Image Classification based on Genetic
94.83% and 98.66% respectively for the training samples Algorithm and Support Vector Machine. Journal of
of 5%. The obtained results are compared with other Knowledge-Based Systems, 24(11): 40-48.
existing algorithms. The proposed method was producing 8. Melgani, F., B. Al Hashemy and M.R.S. Taha, 2000.
promising results for some classes. Our next immediate An Explicit Fuzzy Supervised Classification Method
work is to apply other evolutionary computing algorithms for Multispectral Remote Sensing Images, IEEE
to select the best features. Trans. Geosci. Remote Sens., 38: 287-295.