HSI Course

1
Introduction to Hyperspectral Imaging HSI
Feature Extraction Methods
Dr. Richard B. Gomez, Instructor
2
Outline

• What is Hyperspectral Image Data?
• Interpretation of Digital Image Data
• Pixel Classification
• HSI Data Processing Techniques
• Methods and Algorithms (Continued)
• Principal Component Analysis
• Unmixing Pixel Problem
• Spectral Mixing Analysis
• Other
• Feature Extraction Techniques
• N-dimensional Exploitation
• Cluster Analysis

3
What is Hyperspectral Image Data?

• Hyperspectral image data is image data that is
• In digital form, i.e., a picture that a
  computer can read, manipulate, store, and
  display
• Spatially quantized into picture elements
  (pixels)
• Radiometrically quantized into discrete
  brightness levels
• It can be in the form of Radiance, Apparent
  Reflectance, True Reflectance, or Digital Number

4
Difference Between Radiance and Reflectance

• Radiance is the variable directly measured by
  remote sensing instruments
• Radiance has units of watt/steradian/square
  meter
• Reflectance is the ratio of the amount of light
  leaving a target to the amount of light striking
  the target
• Reflectance has no units
• Reflectance is a property of the material being
  observed
• Radiance depends on the illumination (both its
  intensity and direction), the orientation and
  position of the target, and the path of the light
  through the atmosphere
• Atmospheric effects and the solar illumination
  can be compensated for in digital remote sensing
  data. This yields something, which is called
  "apparent reflectance," and it differs from true
  reflectance in that shadows and directional
  effects on reflectance have not been dealt with

5
Interpretation of Digital Image Data

• Qualitative Approach Photointerpretation by a
  human analyst/interpreter
• On a scale large relative to pixel size
• Limited multispectral analysis
• Inaccurate area estimates
• Limited use of brightness levels
• Quantitative Approach Analysis by computer
• At individual pixel level
• Accurate area estimates possible
• Exploits all brightness levels
• Can perform true multidimensional analysis

6
Data Space Representations

• Spectral Signatures - Physical Basis for Response

• N-Dimensional Space - For Use in Pattern Analysis

7
Hyperspectral Imaging Barriers

• Scene - The most complex and dynamic part

• Sensor - Also not under analysts control

• Processing System - Analysts choices

8
Finding Optimal Feature Subspaces
HSI Data Analysis Scheme

• Feature Selection (FS)

• Discriminant Analysis Feature Extraction (DAFE)

• Decision Boundary Feature Extraction (DBFE)

• Projection Pursuit (PP)

Available in MultiSpec via WWW at
http//dynamo.ecn.purdue.edu/biehl/MultiSpec/
Additional documentation via WWW at
http//dynamo.ecn.purdue.edu/landgreb/publicatio
ns.html After David Landgrebe, Purdue
University
9
Dimension Space Reduction
10
Pixel Classification

• Labeling the pixels as belonging to particular
  spectral classes using the spectral data
  available
• The terms classification, allocation,
  categorization, and labeling are generally used
  synonymously
• The two broad classes of classification
  procedure are supervised classification and
  unsupervised classification
• Hybrid Supervised/Unsupervised Methods are
  available

11
Pixel Classification
12
Pixel Classification
13
Classification Techniques

• Unsupervised
• Supervised
• Hybrid

14
Classification
15
Classification (Cont)
16
Classification (Cont)
17
Classification (Cont)
18
Classification (Cont)
19
Classification (Cont)
20
Classification (Cont)
21
Classification (Cont)

22
Classifier Options

• Other types - Nonparametric
• Parzen Window Estimators
• Fuzzy Set - based
• Neural Network implementations
• K Nearest Neighbor - K-NN
• etc.

23
Classification Algorithms

• Linear Spectral Unmixing (LSU)
• Generates maps of the fraction of each endmember
  in a pixel
• Orthogonal Subspace Projection (OSP)
• Suppresses background signatures and generates
  fraction maps like the LSU algorithm
• Spectral Angle Mapper (SAM)
• Treats a spectrum like a vector Finds angle
  between spectra
• Minimum Distance (MD)
• A simple Gaussian Maximum Likelihood algorithm
  that does not use class probabilities
• Binary Encoding (BE) and Spectral Signature
  Matching (SSM)
• Bit compare simple binary codes calculated from
  spectra

24
Unsupervised Classification

• K-MEANS
• Use of statistical techniques to group
  n-dimensional data into their natural spectral
  classes
• The K-Means unsupervised classifier uses a
  cluster analysis approach that requires the
  analyst to select the number of clusters to be
  located in the data, arbitrarily locates this
  number of cluster centers, then iteratively
  repositions them until optimal spectral
  separability is achieved
• ISODATA (Iterative Self-Organizing Data Analysis
  Technique)
• IsoData unsupervised classification calculates
  class means evenly distributed in the data space
  and then iteratively clusters the remaining
  pixels using minimum distance techniques
• Each iteration recalculates means and
  reclassifies pixels with respect to the new means
• This process continues until the number of
  pixels in each class changes by less than the
  selected pixel change threshold or the maximum
  number of iterations is reached

25
Iterative Self-Organizing Data Analysis Technique
(ISODATA)

• IsoData, unsupervised classification,
  calculates class means evenly distributed in the
  data space and then iteratively clusters the
  remaining pixels using minimum distance
  techniques
• Each iteration recalculates means and
  reclassifies pixels with respect to the new
  means
• This process continues until the number of
  pixels in each class changes by less than the
  selected pixel change threshold or the maximum
  number of iterations is reached

26
Supervised Classification

• Supervised classification requires that the
  user select training areas for use as the basis
  for classification
• Various comparison methods are then used to
  determine if a specific pixel qualifies as a
  class member
• A broad range of different classification
  methods, such as Parallelepiped, Maximum
  Likelihood, Minimum Distance, Mahalanobis
  Distance, Binary Encoding, and Spectral Angle
  Mapper can be used

27
Parallelepiped

• Parallelepiped classification uses a simple
  decision rule to classify multidimensional
  spectral data
• The decision boundaries form an n-dimensional
  parallelepiped in the image data space
• The dimensions of the parallelepiped are
  defined based upon a standard deviation threshold
  from the mean of each selected class

28
Maximum Likelihood

• Maximum likelihood classification assumes that
  the statistics for each class in each band are
  normally distributed
• The probability that a given pixel belongs to a
  specific class is then calculated
• Unless a probability threshold is selected, all
  pixels are classified
• Each pixel is assigned to the class that has
  the highest probability (i.e., the "maximum
  likelihood")

29
Minimum Distance

• The minimum distance classification uses the
  mean vectors of each region of interest (ROI)
• It calculates the Euclidean distance from each
  unknown pixel to the mean vector for each class
• All pixels are classified to the closest ROI
  class unless the user specifies standard
  deviation or distance thresholds, in which case
  some pixels may be unclassified if they do not
  meet the selected criteria

30
Euclidean Distance
31
Mahalanobis Distance

• The Mahalanobis Distance classification is a
  direction sensitive distance classifier that uses
  statistics for each class
• It is similar to the Maximum Likelihood
  classification, but assumes all class covariances
  are equal and, therefore, is a faster method
• All pixels are classified to the closest ROI
  class unless the user specifies a distance
  threshold, in which case some pixels may be
  unclassified if they do not meet the threshold

32
Bhattacharyya Distance
Mean Difference Term
Covariance Term
33
Binary Encoding Classification

• The binary encoding classification technique
  encodes the data and endmember spectra into 0s
  and 1s based on whether a band falls below or
  above the spectrum mean
• An exclusive OR function is used to compare
  each encoded reference spectrum with the encoded
  data spectra and a classification image produced
• All pixels are classified to the endmember with
  the greatest number of bands that match unless
  the user specifies a minimum match threshold, in
  which case some pixels may be unclassified if
  they do not meet the criteria

34
Spectral Angle Mapper Classification

• The Spectral Angle Mapper (SAM) is a
  physically-based spectral classification that
  uses the n-dimensional angle to match pixels to
  reference spectra
• The SAM algorithm determines the spectral
  similarity between two spectra by calculating the
  angle between the spectra, treating them as
  vectors in a space with dimensionality equal to
  the number of bands

35
Spectral Angle Mapper (SAM) Classification

• The Spectral Angle Mapper (SAM) is a physically
  based spectral classification that uses the
  n-dimensional angle to match pixels to reference
  spectra
• The algorithm determines the spectral
  similarity between two spectra by calculating the
  angle between the spectra, treating them as
  vectors in a space with dimensionality equal to
  the number of bands
• The SAM algorithm assumes that hyperspectral
  image data have been reduced to "apparent
  reflectance", with all dark current and path
  radiance biases removed

36
Spectral Angle Mapper (SAM) Algorithm
The SAM algorithm uses a reference spectra, r,
and the spectra found at each pixel, t.  The
basic comparison algorithm to find the angle ?
is (where nb number of bands in the image)
OR
37
Minimum Noise Fraction (MNF) Transformation

• The minimum noise fraction (MNF) transformation
  is used to determine the inherent dimensionality
  of image data, to segregate noise in the data,
  and to reduce the computational requirements for
  subsequent processing
• The MNF transformation consists essentially of
  two-cascaded Principal Components transformations
• The first transformation, based on an estimated
  noise covariance matrix, decorrelates and
  rescales the noise in the data. This first step
  results in transformed data in which the noise
  has unit variance and no band-to-band
  correlations
• The second step is a standard Principal
  Components transformation of the noise-whitened
  data.
• For further spectral processing, the inherent
  dimensionality of the data is determined by
  examination of the final eigenvalues and the
  associated images
• The data space can be divided into two parts
  one part associated with large eigenvalues and
  coherent eigenimages, and a complementary part
  with near-unity eigenvalues and noise-dominated
  images. By using only the coherent portions, the
  noise is separated from the data, thus improving
  spectral processing results.

38
N - Dimensional Visualization

• Spectra can be thought of as points in an n
  -dimensional scatterplot, where n is the number
  of bands
• The coordinates of the points in n -space
  consist of "n" values that are simply the
  spectral radiance or reflectance values in each
  band for a given pixel
• The distribution of these points in n - space
  can be used to estimate the number of spectral
  endmembers and their pure spectral signatures

39
Pixel Purity Index (PPI)

• The "Pixel-Purity-Index" (PPI) is a means of
  finding the most "spectrally pure," or extreme
  pixels in multispectral and hyperspectral images
• PPI is computed by repeatedly projecting
  n-dimensional scatterplots onto a random unit
  vector
• The extreme pixels in each projection are
  recorded and the total number of times each pixel
  is marked as extreme is noted
• A PPI image is created in which the DN of each
  pixel corresponds to the number of times that
  pixel was recorded as extreme

40
Matched Filter Technique

• Matched filtering maximizes the response of a
  known endmember and suppresses the response of
  the composite unknown background, thus "matching"
  the known signature
• Provides a rapid means of detecting specific
  minerals based on matches to specific library or
  image endmember spectra
• Produces images similar to the unmixing
  technique, but with significantly less
  computation
• Results (values from 0 to 1), provide a means
  of estimating relative degree of match to the
  reference spectrum where 1 is a perfect match

41
Spectral Mixing

• Natural surfaces are rarely composed of a
  single uniform material
• Spectral mixing occurs when materials with
  different spectral properties are represented by
  a single image pixel
• Researchers who have investigated mixing scales
  and linearity have found that, if the scale of
  the mixing is large (macroscopic), mixing occurs
  in a linear fashion
• For microscopic or intimate mixtures, the
  mixing is generally nonlinear

42
Mixed Spectra Models

• Mixed spectra effects can be formalized in three
  ways
• A physical model
• A mathematical model
• A geometric model

43
Mixed Spectra Physical Model
44
Mixed Spectra Mathematical Model
45
Mixed Spectra Geometric Model
46
Mixture Tuned Matched Filtering (MTMF)

• MTMF constrains the Matched Filtering as
  mixtures of the composite unknown background and
  the known target
• MTMF produces the standard Matched Filter score
  images plus an additional set of images for each
  endmember infeasibility images
• The best match to a target is obtained when the
  Matched Filter score is high (near 1) and the
  infeasibility score is low (near 0)

47
Principal Component Analysis (PCA)

• Calculation of new transformed variables
  (components) by a coordinate rotation
• Components are uncorrelated and ordered by
  decreasing variance
• First component axis aligned in the direction
  of the highest percentage of the total variance
  in the data
• Component axes are mutually orthogonal
• Maximum SNR and largest percentage of total
  variance in the first component

48
Principal Component Analysis (PCA)
49
Principal Component Analysis (PCA) (Cont)

• The mean of the original data is the origin of
  the transformed system with the transformed axes
  of each component mutually orthogonal
• To begin the transformation, the covariance
  matrix, C, is found. Using the covariance matrix,
  the eigenvalues, ?i, are obtained from
  C ?iI 0
• where i 1,2,...,n (n is the total number of
  original images and I is an identity matrix)

50
Principal Component Analysis (PCA) (Cont)

• The eigenvalues, ?i,, are equal to the variance
  of each corresponding component image
• The eigenvectors, ei , define the axes of the
  components and are obtained from
  (C ?iI) ei 0
• The principal components are then given as
• PC T DN
• where DN is the digital number matrix of the
  original data and T is the (n x n) transformation
  matrix with
• matrix elements given by eij , i, j
  1,2,3,...n

51
A Matrix Equation
Problem Find the value of vector x from
measurement of a different vector y, where they
are related by the matrix equation given
by y Axor yi ?aijxj sum over j
Note1 If both A and x are known, it is
trivial to find y Note2 In our problem, y is
the measurement, and A is determined from the
physics of the problem, and we want to retrieve
the value of x from y
52
Mean and Variance

• Mean ?x? (1/N)? xk
• Variance
• var(x) (1/N) ?(xk - ?x?)2 ?x2 where k
  1,2,,N

53
Covariance
cov(x,y) (1/N) ?(xk ? ?x?)(yk ? ?y?)
(1/N) ? xk yk ? ?x? ?y? Note1 cov(x,x)
var(x) Note2 If the mean values of x and y are
zero, then cov(x,y) (1/N) ? xk yk Note3 Sums
are over k 1,2,., N
54
Covariance Matrix

• Let x (x1, x2, ,xn) be a random vector with
  n components
• The covariance matrix of x is defined to
  be C ?(x ? ?)(x ? ?)T?
• where ? (?1, ?2, ?k)T
• and ?k (1/N)?xmk
• Summation is over m 1,2,, N

55
Gaussian Probability Distributions

• Many physical processes are well represented
  withGuassian distributions given by
• P(x) (1/?2??x)e(x?ltxgt)?2 /2 ?x ?2
• Given the mean and variance of a Guassian
  random variable,it is possible to evaluate all of
  the higher moments
• The form of the Gaussian is analytically simple

56
Normal (Gaussian) Distribution
57
Scatterplots
58
Spectral Signatures
Laboratory Data Two classes of vegetation
59
Discrete (Feature) Space
60
Hughes Effect
G.F. Hughes, "On the mean accuracy of statistical
pattern recognizers," IEEE Trans. Inform.
Theory., Vol IT-14, pp. 55-63, 1968.
61
Higher Dimensional Space Implications

• High dimensional space is mostly empty. Data in
  high dimensional space is mostly in a lower
  dimensional structure.

• Normally distributed data will have a tendency to
  concentrate in the tails Uniformly distributed
  data will concentrate in the corners.

62
Higher Dimensional Space Geometry

• The number of labeled samples needed for
  supervised classification increases rapidly with
  dimensionality

In a specific instance, it has been shown that
the samples required for a linear classifier
increases linearly, as the square for a quadratic
classifier. It has been estimated that the number
increases exponentially for a non-parametric
classifier.

• For most high dimensional data sets, lower
  dimensional linear projections tend to be normal
  or a combination of normals.

63
HSI Data Analysis Scheme
After David Landgrebe, Purdue University
200 Dimensional Data
Class Conditional Feature Extraction
Feature Selection
Classifier/Analyzer
Class-Specific Information
64
Define Desired Classes
HSI Image of Washington DC Mall
Training areas designated by polygons outlined in
white
After David Landgrebe, Purdue University
65
Thematic Map of Washington DC Mall
Legend
Operation CPU Time (sec.) Analyst Time Display
Image 18 Define Classes lt 20 min. Feature
Extraction 12 Reformat 67 Initial
Classification 34 Inspect and Mod. Training
5 min. Final Classification 33 Total 164 sec
2.7 min. 25 min.
Roofs Streets Grass Trees Paths Water Shadows
(No preprocessing involved)
After David Landgrebe, Purdue University
66
Hyperspectral Imaging Barriers
Scene - Varies from hour to hour and sq. km to
sq. km
Sensor - Spatial Resolution, Spectral bands, S/N
Processing System -

• Classes to be labeled

• Number of samples to define the classes

• Features to be used

• Complexity of the Classifier

67
Operating Scenario

• Remote sensing by airborne or spaceborne
  hyperspectral sensors
• Finite flux reaching sensor causes
  spatial-spectral resolution trade-off
• Hyperspectral data has hundreds of bands of
  spectral information
• Spectrum characterization allows subpixel
  analysis and material identification

68
Spectral Mixture Analysis

• Assumes reflectance from each pixel is caused by
• a linear mixture of subpixel materials

69
Mixed Pixels and Material Maps
Input Image
PURE
PURE
PURE
MIXED
70
Traditional Linear Unmixing
i 1 k

• Unconstrained
• Partially Constrained
• Fully Constrained

• Constraint Conditions

71
Hierarchical Linear Unmixing Method

• Unmixes broad material classes first
• Proceeds to a groups constituents only if the
  unmixed fraction is greater than a given threshold

72
Stepwise Unmixing Method

• Employs linear unmixing to find fractions
• Uses iterative regressions to accept only the
  endmembers that improve a statistics-based model
• Shown to be superior to classic linear method
• Has better accuracy
• Can handle more endmembers
• Quantitatively tested only on synthetic data

73
Performance Evaluation
Error Metric

• Compare squared error from traditional, stepwise
  and hierarchical methods
• Visually assess fraction maps for accuracy

74
Endmember Selection

• Endmembers are simply material types
• Broad classification road, grass, trees
• Fine classification dry soil, moist soil...
• Use image-derived endmembers to produce spectral
  library
• Average reference spectra from pure sample
  pixels
• Chose specific number of distinct endmembers

75
Materials Hierarchy

• Grouped similar materials into 3-level hierarchy

• Level 1

• Level 2

• Level 3

76
Squared Error Results
77
Stepwise Unmixing Comparisons

• Linear unmixing does poorly, forcing fractions
  for all materials
• Hierarchical approach performs better but
  requires extensive user involvement
• Stepwise routine succeeds using adaptive
  endmember selection without extra preparation

78
HSI Image of Washington DC Mall
HYDICE Airborne System 1208 Scan Lines, 307
Pixels/Scan Line 210 Spectral Bands in 0.4-2.4
µm Region 155 Megabytes of Data (Not yet
Geometrically Corrected)
79
Hyperspectral Imaging Potential

• Assume 10 bit data in a 100 dimensional space

• That is (1024)100 10300 discrete locations

• Even for a data set of 106 pixels, the
  probability

of any two pixels lying in the same discrete
location
is extremely small