Sie sind auf Seite 1von 26

Heaven’s light is our guide

RAJSHAHI UNIVERSITY OF ENGINEERING AND TECHNOLOGY


DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING

Subspace Detection for Hyperspectral


Image Classification Using Folded
Principal Component Analysis and
Cross Cumulative Residual Entropy

Presented By: Supervised By:


Md. Iftekharul Islam Sadia Zaman Mishu
Roll No. : 133096 Assistant Professor
Dept. of CSE Dept. of CSE
RUET RUET
OUTLINE

1 Introduction 5 Classification Results

2 Objective 6 Conclusion

3 Proposed Method 7 Future Works

4 Methodology 8 References

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 2
INTRODUCTION
𝜆
 Hyperspectral imaging leads to a
vastly improved ability to classify
the objects in the scene based on
their spectral properties.
A pixel vector

Figure 1: Hyperspectral
data cube [1]
 Challenges in Hyperspectral Imaging:

1 2 3 4
Data size Increased Data Hughes effect
processing time redundancy

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 3
OBJECTIVE

Developing effective method to detect the optimal subspace


from the hyperspectral image data for better classification

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 4
PROPOSED METHOD

A combination of feature extraction and feature


selection:

1 2
Feature extraction Feature selection
using Folded Principal using Cross Cumulative
Component Analysis Residual Entropy

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 5
METHODOLOGY
REMOTE SENSING DATA FOR ANALYSIS

Dataset: AVIRIS 92AV3C (Indian Pine)


Wavelength range: 0.4µm – 2.5µm
Bands: 220 (Used: 200)
Spatial Size: 145 x 145
Classes: 16

 The bands covering the water absorption


region (104-108, 150-163,220) are
removed as suggested by researchers. [2]

Figure 2: AVIRIS 92AV3C


Dataset [3]

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 6
METHODOLOGY
RESULT OF PCA

Figure 3: Principal components


Subspace Detection for Hyperspectral Image Classification Using Folded Principal
14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 7
METHODOLOGY
LIMITATIONS OF PCA

 No guarantee that the directions of maximum variance will


contain good features for discrimination.

 Noise is present in some of the principal components.

 So we move to Folded Principal Component Analysis.

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 8
METHODOLOGY
FOLDED PRINCIPAL COMPONENT ANALYSIS

 Mean-adjusted spectral vector of a


pixel:
𝑿𝒏 = [𝑿𝒏𝟏 , 𝑿𝒏𝟐 , … 𝑿𝒏𝑭 ]𝑻

 Converted H x W matrix:
𝑨𝒏 = [𝒂𝒏𝟏 , 𝒂𝒏𝟐 , … 𝒂𝒏𝑯 ]𝑻

 Computed partial covariance matrix:


Xn
𝑪𝒏 = 𝑨𝒏 𝑻 𝑨𝒏 An
Spectral vector
of a pixel
 Overall covariance matrix:
𝑺
𝟏 Figure 4: Construction of a 2D matrix by
𝑪𝑭𝑷𝑪𝑨 = 𝑪𝒏 folding a 1D spectral vector [4]
𝑺
𝒏=𝟏

S = Spatial size of the hypercube

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 9
METHODOLOGY
RESULT OF FPCA

Figure 5: Principal components extracted by FPCA (H = 5, W = 40)


Subspace Detection for Hyperspectral Image Classification Using Folded Principal
14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 10
METHODOLOGY
Comparison between PCA and FPCA
 PCA components are prone to noise whereas FPCA components are
better visualized.

 FCPA extracts both local and global structures by folding the spectral
vectors.

 Only q’ = 4 components are extracted from each of the H = 5 segments


to obtain 20 components from the whole dataset.

Table – 1: Memory requirements of PCA and FPCA for different stages


Method Data Matrix Size Covariance Projection
Matrix Size Matrix Size
PCA S x HW HW x HW HW x q

FPCA HxW WxW W x q/H


2
(Saving factor: 𝑆) (Saving factor: 𝐻 ) (Saving factor: 𝐻2 )

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 11
METHODOLOGY
FEATURE SELECTION

 High variance does not always guarantee more


information.

 Therefore cross cumulative residual entropy is implemented


to select the relevant features.

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 12
METHODOLOGY
CROSS CUMULATIVE RESIDUAL ENTROPY

The CCRE of two images I and R is Here,


given by [5],[6]:
L = maximum pixel value of
𝐿 𝐿 the images
𝐺(𝑢, 𝑣)
𝐶= 𝐺 𝑢, 𝑣 log( )
𝐺𝐼 (𝑢)𝑃𝑅 (𝑣) 𝑮 𝒖, 𝒗 = Joint residual
𝑢=1 𝑣=1
entropy distribution

𝑮𝑰 𝒖 = Cumulative residual
The joint and cumulative residual
entropy distribution of I
entropy distribution is given as:
𝑷𝑹 (𝒗) = Marginal probability
𝐺 𝑢, 𝑣 = 𝑝(𝑢 < 𝐼, 𝑅 = 𝑣) of R
𝐿

𝐺𝐼 𝑢 = 𝐺 𝑢, 𝑣
𝑣=1

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 13
METHODOLOGY
RESULT OF CCRE

Figure 6: Ground truth of Indian Pines dataset

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 14
METHODOLOGY
RESULT OF CCRE

Figure 7: CCRE between the principal components and class label


Subspace Detection for Hyperspectral Image Classification Using Folded Principal
14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 15
METHODOLOGY
RESULT OF IMPLEMENTING CCRE

Figure 8: F-PC 13 contains more information than F-PC 9

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 16
CLASSIFICATION RESULTS

 Support vector machine (SVM) classifier is used to classify


in to the selected classes [7].

 Training and test samples are selected randomly based on


the supplied ground-truth image.

 The SVM is trained with RBF kernel and 10 fold cross-


validation to build the model for classification.

 The cost parameter C and kernel width 𝛾 are determined


through cross-validation for measuring the test accuracy.

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 17
CLASSIFICATION RESULTS

Table – 2: Train and Test Samples


Class Name Train Test
Corn-notill 28 28
Woods 24 24
Soybeans-min 24 24
Soybeans-notill 30 30
Hay-widrowed 25 24

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 18
CLASSIFICATION RESULTS

(a) (b)

Figure 9: Comparison of class separation between (a) PCA and (b) FPCA-CCRE

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 19
CLASSIFICATION RESULTS

Table – 3: Classification Accuracy

Method H x W Fold Classification


Accuracy (%)
PCA - 98.2579

2 x 100 96.1752
FPCA-CCRE 4 x 50 97.7030
5 x 40 98.4437
8 x 25 96.9801
10 x 20 99.5833

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 20
CLASSIFICATION RESULTS

100

Classification Accuracy (%) 99

98

97

96

95

94
2x100 4x50 5x40 8x25 10x20
Folding Options

FPCA-CCRE PCA

Figure 10: Classification accuracy between


PCA and FPCA-CCRE

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 21
CONCLUSION

 PCA suffers from drawbacks such as extremely high


computational cost, large memory requirement and low
efficacy.

 Therefore, Folded-PCA is implemented that considers global


as well as local structures.

 Using CCRE as feature selection criteria on the folded-PC


with appropriate folding option yields significant
improvement in classification accuracy compared to PCA.

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 22
FUTURE WORKS

 Testing the method on a different dataset

 Trying Mutual information as feature selection along with


FPCA

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 23
REFERENCES
[1] “Hyperspectral Data Processing : University of Dayton, Ohio”. Available:
https://udayton.edu/engineering/centers/vision_lab/was_data_analysis_and_pr
ocessing/hyperspectral_data_processing.php. [Accessed: 22-Mar-2018].

[2] H. Huang, J. Liu, Y. Pan, "Semi-supervised marginal fisher analysis for


hyperspectral image classification", ISPRS Annals of the Photogrammetry Remote
Sensing and Spatial Information Sciences, vol. 1, no. 3, pp. 377-382, 2012.

[3] M. F. Baumgardner, L. L. Biehl, and D. A. Landgrebe, “220 Band AVIRIS


Hyperspectral Image Data Set: June 12, 1992 Indian Pine Test Site 3,” PURR.
[Online]. Available: https://purr.purdue.edu/publications/1947/1. [Accessed: 22-
Mar-2018].

[4] J. Zabalza, J. Ren, M. Yang, Y. Zhang, J. Wang, S. Marshall, J. Han, "Novel


folded-PCA for improved feature extraction and data reduction with
hyperspectral imaging and SAR in remote sensing", ELSEVIER ISPRS Journal of
Photogrammetry and Remote Sensing, vol. 93, pp. 112-122, 2014.

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 24
REFERENCES
[5] M. A. Hossain, B. Ahmed, S. S. Ghosh and M. N. I. Mondal, “Effective subspace
detection based on cross cumulative residual entropy for hyperspectral image
classification”, 2017 International Conference on Electrical, Computer and
Communication Engineering (ECCE), Cox's Bazar, pp. 548-551, 2017.

[6] M. Rao, Y. Chen, B. C. Vemuri, and F. Wang, “Cumulative residual entropy: a


new measure of information”, IEEE Transactions on Information Theory, vol. 50,
no. 6, pp. 1220- 1228, 2004.

[7] R. Archibald, G. Fann, "Feature Selection and Classification of Hyperspectral


Images With Support Vector Machines", IEEE Geoscience and Remote Sensing
Letters, vol. 4, pp. 674-677, 2007.

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 25
THANK YOU

Subspace Detection for Hyperspectral Image Classification Using Folded Principal


14/7/2018
Component Analysis & Cross Cumulative Residual Entropy 26

Das könnte Ihnen auch gefallen