Abstract– In Hyperspectral remote sensing, data are

collected in numerous(hundreds to thousands) narrow wavebands in one or more

regions of the electromagnetic spectrum, and large numbers of data are

collected. An important problem in hyperspectral image processing are dimension

reduction, target

detection, target identification, and target classification. In this paper we review the current activity of target

classification, most commonly used methods for dimension reduction, target detection,

target identification methodologies and

techniques. Hyperspectral image processing is a complex process which depends

on various factors. Here we also reviewed problems faced by some methods and to

overcome the problems, discuss the current techniques, problems as well as prospects

of data analysis. The main focus will be on advanced Image data analysis and

classification techniques which are used for improving accuracy. Additionally,

some important issues relating to classification performance are also

discussed.

Keywords— Hyperspectral Image; target

detection; dimensionality reduction; Independent Component Analysis; Principal

Component Analysis; Projection

Pursuit.

I.

INTRODUCTION

Hyperspectral data is classified

as Feature selection/extraction followed by Information extraction, both

feature selection/extraction and Information extraction methods could be either

supervised or unsupervised methods13. The unsupervised methods identify

patterns of interest in an image data. This group of methods do not require

prior knowledge where as supervised methods use prior knowledge on target

characteristics whereas care should be taken while selecting. Although there are numerous unsupervised

methods, only the most commonly used methods are discussed here,

A.

Projection

Pursuit (PP)

Unlike

most of developed target detection algorithms that require statistical models

such as linear mixture, PP is to project a high dimensional data set into a low

dimensional data space while retaining desired information of interest. It

utilizes a projection index to explore projections of interestingness. Since targets

are small compared to their surrounding background, these targets can be viewed

as pixels that cause outliers of the background distribution. In order to find

the optimal projections, a revised Projection Pursuit evolutionary algorithm(PPEA)

is used where a zero-detection thresholding technique is introduced for the

purpose of target detection14. A new technique Legendre index for anomaly

detection is used based on projection pursuit,

the proposed PP technique is able to detect anomalies with a degree of

separation from the normal distribution given by the level of gray of the

pixel. RX algorithm is used in addition to isolated outliers15.

B.

Principal

Component Analysis(PCA)

PCA is a multivariate method

commonly used for reducing data redundancy and dimensionality13. It is an

unsupervised method, if the user is interested in a phenomenon or variable that

causes subtle differences in target reflectance in specific bands, then it is

not the best method for feature selection13. Authors tells that hierarchical

PCA algorithm, which can effectively reduce the hyperspectral data to intrinsic

dimensionality. In this, image are break into various parts and then perform

PCA on each part separately and then combine the results. Therefore

hierarchical PCA provide similar information content as compared to traditional

PCA. On further investigation it was established that the classification

accuracy of hierarchical method is also very close to traditional PCA method

16. Even though PCA is widely used it suffers from high computational cost,

large memory requirement and low efficacy in dealing with high dimensional data

17. The contribution of small target is limited to the variance of the image

frames. In a much higher variance target image frame the smaller targets may

not appear after PCA analysis. A solution for above issue was addressed in 18

using the Independent Component Analysis (ICA) for unsupervised target

detection. The ICA can be used for classification, feature extraction and

target detection in hyperspectral images 18. The goals of PCA are to (1)

extract the most important information from the data table; (2) compress the

size of the data set by keeping only this important information; (3) simplify

the description of the data set; and (4) analyze the structure of the

observations and the variables22.

C.

Independent

Component Analysis (ICA)

Independent-component analysis

_ICA_ is a popular technique for unsupervised classification6. Introduced in

the early 1980’s is a multivariate data analysis method where, given a linear

mixture of statistically independent components, these components are recovered

by solving for an unmixing matrix. Whereas PCA finds the transform of the observed

data that de-correlates the observed variables through the use of second-order

statistics (i.e. A transform based on the eigenvectors of the covariance

matrix), ICA utilizes higher-order statistics to find projections of the data

where the components are independent, a stronger statement than

uncorrelated19. Author tells that Most target detection algorithms uses a

priori available target, target is, seldom available a priori.

Independent component analysis (ICA) is a technique that aims at finding out

components which are statistically independent or as independent as possible.

Since ICA does not require a priori target information. This

technique therefore has the potential of being used for target detection

applications5. The major advantage of using ICA is its ability to classify

objects with unknown spectral signatures in an unknown image scene. But it’s

very high computational complexity impedes its application to high-dimensional

data analysis. The common approach is to use principal component analysis(PCA)

to reduce the data dimensionality before applying the ICA classification6.

II.

CHALLENGES IN HYPERSPECTRAL IMAGE ANALYSIS

The

spectral data in the hyperspectral image can be used to identify known and

unknown objects on the basis of spectral signature. The spectral information measured for the same material differs

due to difference in material composition, atmospheric propagation and sensor

noise. These spectral variations in

the spectral signature for same material make the image analysis challenging1.

It is critical to consider spectral variations in target identification problem

for accurate identification of the targets.

Some problems of hyperspectral

remote sensing are listed below,

A.

Processing and

Visualization Problem

Hyperspectral images contain far

more spectral bands than can be displayed with a standard red, green and blue

(RGB) display. Here, Color Matching functions(CMF) is one of the method

specifies how much of these three primary colors must be mixed to create the

color sensation of a monochromatic light at the particular wavelength to

produce original spectrum7. A disadvantage of the CMF is that there might be

a decrease in sensitivity of human vision at the edge of the visible

spectrum7. Principal component analysis (PCA) is also used to reduce

hyperspectral data dimensionality by assigning the first three principle

components to RGB9. And also using wavelets to de-noise the spectra before

applying PCA could improve Visualization10. The disadvantage of PCA include

the difficulty to interpret the displayed image because the displayed colors

represents that do not typically represent natural colors of the features. The

colors change drastically, depending on the data, and they do not correlate

strongly with data variation. The standard saturation used in PCA display leads

to simultaneous contrast problems and the computational complexity is high7.

B.

Data Handling Issues

The users of hyperspectral

should have the capability to store and handle large size data sets. they would

require high performance computer with large storage capacity13.

C.

Data Redundancy

Problem

If often refers to the fact that

the information contained in each band of the hyperspectral image is not

unique. On the contrary, many bands are very similar or redundant.

Hyperspectral data redundancy can be visualized through covariance or

correlation between bands13.

D.

The Curse of Dimensionality

As the number of bands in an

image increases, the number of observations required to train a classifier

increases exponentially to maintain the classification accuracies13. I.

TARGET

DETECTION Target

detection in hyperspectral images is important in many applications including

search and rescue operations, defense systems, mineral exploration, border

security, agricultural crops and several other anthropogenic and natural object/phenomenon. For these

purpose, several target detection algorithms have been proposed over the years,

classification of target detection and review of them is done in1. Authors, describe the fundamental structure

of the hyperspectral data and explain how these data influence the signal

models used for the development and theoretical analysis of detection

algorithms4. However,

it is not clear which of these algorithms perform best on real data targets,

and moreover, which of these algorithms have complementary information and

should be fused together. Apart from1, For this purpose, eight signature-based

hyperspectral target detection algorithms, namely the Generalized Likelihood

Ratio Test(GLRT), Adaptive Coherence Estimator(ACE), Signed Adaptive Coherence

Estimator(SACE), Adaptive Matched Subspace Detector(AMSD)( The use of adaptive

algorithms deals quite effectively with the problem of unknown backgrounds4),

Constrained Energy Minimization(CEM), Matched Filter(MF), Orthogonal Subspace

Projection(OSP) and Hybrid Unstructured Detector(HUD), and three anomaly

detectors, namely RX, Maxmin and Diffdet, were tested and compared. Among the

signature-based target detectors, the three best performing algorithms that

have complementary information were identified. Finally these algorithms were

fused together using four different fusion algorithms11. SACE,

CEM and AMSD were found to be the better-performance algorithms, and AMSD

showed the best performance. In fact, AMSD showed a good performance especially

if the sub-pixel target area was close to at least half of the pixel area.

However, AMSD requires to model the background endmembers( is a pure spectra or

pure materials and pure pixels are often referred as endmembers each having a

characteristic spectral signature) , which increases the computational

complexity11. With

this study, it was shown that AMSD, SACE and CEM showed success and weaknesses

on different regions, and complemented each other. Hence, these algorithms were

fused with the sum, product, MFF and hybrid fusion methods11. It

is seen that the effect of CEM detection

is relatively poor3, To solve the low detection efficiency problem of

Constrained Energy Minimization (CEM) method used for hyperspectral remote

sensing imagery, Author firstly presents two improved detection methods:

principal component CEM (PCCEM) and matrix taper CEM (MTCEM). Then, based on

these two methods, a more optimized Two-Time detection (TTD) method is

proposed. Primarily, the targets of interest in the hyperspectral image are

detected by using the PCCEM and MTCEM method. show that the detection

performance of PCCEM and MTCEM algorithms varies with the image data. These

methods are not robust detector. No matter what kind of image is used, the TTD

method is able to get the good target detection result that is superior to the

above methods, and has a robust performance at target detection20. Author21

showed that there is no “best hyperspectral detection algorithm” for all images

and targets. We noted the significant effect spatial distribution has on

detector performances, and we showed that the RBTA can be used to select the

proper detectors from among several detectors but without any need for ground

truth. However, point targets can influence their neighboring pixels, due

either to the PSF or to the target spreading across more than one pixel. To

account for this potential source of inaccuracy, therefore, we introduced the

improved RBTA (IRBTA), whose exact method of use depended on the target size.

In addition, we showed that when detectors calculated the mean for estimating

the pixel signature value, we did not need ground truth to find the best

estimate. We tested our concept through the selection of the best detectors

from among stochastic algorithms for target detection, that is, the constrained

energy minimization (CEM), generalized likelihood ratio test (GLRT), and

adaptive coherence estimator (ACE) algorithms, using the dataset and scoring

methodology of the Rochester Institute of Technology (RIT) Target Detection

Blind Test project. The results showed that our concepts predicted the best

algorithms for the particular images and targets provided by the website21. The author made the comparative

study of target detection algorithms in HSI applied to crops scenarios in

Colombia from images acquired by hyperspectral satellite sensor Hyperion. The

tested algorithms were ACE, CEM, MF, SAM and OSP. The results show that the

ACE algorithm has a better performance with probabilities detection PD > 90% for diverse

HSI and agricultural targets, in both synthetic and real images, followed by

CEM and MF algorithms that exhibit acceptable performance with averages

detection probabilities PD =

80%. In contrast, the OSP and SAM algorithms are able to detect targets

with average PD =

45% however, the number of false alarms (FA) is high and their performance

decreases12. Target

detection from hyperspectral images using ICA and other algorithms based on

spectral modeling may be of immense interest5. Author

compares ICA with four spectral matching algorithms namely Orthogonal Subspace

Projection (OSP), Constrained Energy Minimization (CEM)3, Spectral Angle

Mapper (SAM) and Spectral Correlation Mapper (SCM), and four anomaly detection

algorithms namely OSP anomaly detector (OSPAD), Reed–Xiaoli anomaly detector

(RXD), Uniform Target Detector (UTD) and a combination of Reed–Xiaoli anomaly

detector and Uniform Target Detector (RXD–UTD) were considered. The experiments

were conducted using a set of synthetic and AVIRIS hyperspectral

images containing aircrafts as military targets. A comparison of true positive

and false positive rates of target detections obtained from ICA and other

algorithms indicates the superior performance of the ICA5 over other

algorithms3.

II.

PROBLEM IDENTIFICATION Identifying

theoretical computational problem in Hyperspectral image processing itself is a

big problem. The theoretical computational

problem are : Finding / putting data in

a certain format , dig into data , perform analysis in High Performance

Computing(HPC) way , remove sensor noise and atmospheric noise, identifying end-member

/ target. All this problems can be solved using parallel computing aspect while

doing HPC. A. Methodology·

Independent

Component Analysis(ICA) is used instead of Principal Component Analysis(PCA)

24-26.·

Noise

Reduction (removing high frequency noise by cloud cover, Atmospheric

correction is an important pre-requisite step to enhance and improve

identification of spectral signatures of different objects or materials and

their compositions8).·

Data

Spaces(Dimensionality reduction).How much of

dimensionality reduction is needed? data-mining/data-cube

concepts are used then no need to do data reduction. This is where

computational skills come into picture to enter any part of the cube. A variety

of dimension reduction techniques exist: principal component analysis (PCA) ,

minimum noise fraction, locally linear embedding independent component analysis

, discrete wavelet transform etc17.·

Data

Mining(data reduction).

III.

LITERATURE GAP AND CONCLUSION All

the Hyperspectral Image (HSI) Processing algorithms discussed were built on

their own assumptions and thus have limitations. The basic theory and the most

canonical works are discussed along with the most recent advances in each

aspect of hyperspectral image processing, both the most classic and advanced

work were introduced. From this work, directions in current research trends are

revealed2. An

idealized universal Hyperspectral Image Processing system has not been

developed yet. The method used for HSI processing should be a function problem

type, Target Detection, Material Mapping, Material Identification, Mapping

details and surface properties, Material Classification, Atmospheric correction

etc. from remote sensing data. To

identify the target first we need to reduce the band size, from 224 band to

suitable size of 10-20 bands, to perform this Independent Component Analysis is used instead of Principal Component Analysis(PCA),

Next the Noise has to be Reduced( removing high frequency noise by cloud cover

). Data Spaces(Dimensionality reduction) and with proper selection of a suite

of data mining(data reduction) techniques, it is possible to reduce data

dimensionality and data redundancy, and extract unique information from HS

images13. After all the

operation the target

Detection methods which are used are not always reliable especially for

hyperspectral data, since they depend on image statistics only. hence we need

to find the better target detectors. But it requires much effort. Fusion

methods produce precise results. It

is also found that hybrid methods like Genetic algorithms, swarm techniques,

Bayesian formulations23 despite the promising research results, are not often

used in the processing of hyperspectral data. Hybrid methods23 can be tried

with physical models and other scene dependent properties of the area. New

approaches like hyperspectral imageries using neural network, machine learning

techniques, open possibilities in the improvement of the topic.

The primary disadvantages are cost and complexity.

Fast computers, sensitive detectors, and large data storage capacities are

needed for analyzing hyperspectral data. Significant data storage capacity is

necessary since hyperspectral cubes are large, multidimensional datasets,

potentially exceeding hundreds of megabytes.

All of these factors greatly increase the cost of acquiring and processing

hyperspectral data.