Digital Image Classification

 Uses the spectral information represented by

the digital numbers in one or more spectral
bands.
 Classifies each individual p
pixel based on this
spectral information. Also known as
spectral pattern recognition.
 The objective: To assign all pixels in the image to
particular classes or themes (e.g., water, coniferous
forest deciduous forest
forest, forest, crops
crops, bare soil
soil, etc
etc.))

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 1 / 72
Digital Image Classification

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 2 / 72
Digital Image Classification
Digital Image
Classification

Residential
Commercial
Transportation
Green crops / grass
Forests
Bare mineral soils
Bare organic soils
Water

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 3 / 72
 Digital Image Classification
 The two approaches:
 Per-pixel classification
• The algorithm categorizes each input pixel into
a spectral feature class based solely on its
individual multispectral vector (signature).
(signature) No
context or neighborhood evaluation is involved.

 Object-oriented classification
• The input pixels are grouped into spectral
features (objects) using an image segmentation
algorithm. These objects are characterized in
both the raster and vector domains. The objects
are classified
l ifi d using
i b both
th spectral
t l and
d spatial
ti l
cues (metrics).
David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 4 / 72
 Digital Image Classification
 The three common per-pixel methods are:
 Supervised classification
 Unsupervised classification
 Rule-based classification

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 5 / 72
 Digital Image Classification
 Supervised classification
 The analyst "supervises” the categorization of a set
of specific classes by providing training statistics
that identify each category.
 Unsupervised classification
 The raw spectral data are grouped first, based
solelyy on the statistical structure of the data. Then
the analyst must label each statistical cluster,
placing them into the appropriate categories (if
possible).
possible)
 Rule-based classification
 Spectrally categorized pixels are classified using
ancillary data in a GIS model.
David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 6 / 72
 Digital Image Classification
 Supervised classification
 In the imagery, the analyst identifies homogeneous,
representative examples of the various surface
cover types (information classes) of interest.

 These samples are referred to as training areas.

 The selection of appropriate training areas is based

on the analyst's familiarity with the geographical
area and their knowledge of the actual surface
co er ttypes
cover pes present in the image
image.

 Good visual interpretation skills are mandatory

f success.
for

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 7 / 72
 Digital Image Classification
 Supervised classification
Spectral
Trainingg Sites Signatures
Conifer

mpiles
The computer sooftware com
Conifers
Water
Used to classifyy
Water all unknown
Deciduous pixels

D d
Deciduous

Digital Image

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 8 / 72
 Digital Image Classification
 Supervised classifiers (algorithms)
 Parallelpiped – based on range or variance of class DNs

 Minimum Distance-to-Means – based on mean class DNs

 Maximum Likelihood – based on probability of class

membership

 Spectral Angle Mapper – class membership based on

minimum difference from the n-dimensional spectral
vectors of the classes

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 9 / 72
 Digital Image Classification
 Supervised classification process
 Step 1: Define training data
 Each training site should appear homogenous and
representative of the legend class

 Delineate several training sites for each legend class

 Make each training site at least 20-25 pixels (i.e., > 5 acres
for 30 meter pixels)

 Each class should be represented by ~ 100 n pixels (where n

= number of spectral bands in the data set)

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 10 / 72
 Digital Image Classification
 Supervised classification process
 Trainingg Sites
 For 6-band TM & ETM imagery, the total number of
training pixels per class should be at least 600
 Try to capture the landscape diversity of the class

100 100
600
pixels 100

100

100

100

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 11 / 72
 Digital Image Classification
 Example of a homogeneous training site

Check all bands

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 12 / 72
 Digital Image Classification
 Example of a heterogeneous training site

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 13 / 72
 Digital Image Classification
 Supervised classification process
 Trainingg data
 Should be normally distributed if the maximum
likelihood classifier is going to be used

Darker DNs Brighter DNs

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 14 / 72
 Digital Image Classification
 Supervised Classification Process
 Co
Collect
ect seve
several
a training
t a g sites
s tes for
o eac
each class
c ass in order
o e to
capture the landscape diversity of the class
Deep, clear water Turbid water

A dark version A bright version

of class 1 of class 1

Darker DNs - 1 +1 - 1 +1 Brighter DNs

 
The advantage of individual signatures:
separate mean DN values and small variances

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 15 / 72
 Digital Image Classification
 Supervised Classification Process
 For variable classes do NOT merge
g signatures
g

Darker DNs - 1 +1 Brighter DNs


The disadvantage of merged signatures:
shifted mean DN value (atypical of the class) and a large variance

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 16 / 72
 Supervised Classification
 Training Sites – Tools in ERDAS Imagine
 Drawing tab: Insert Geometry Tools
 Polygon tool
 Region Grow tool

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 17 / 72
 Supervised Classification
Polygon
tool

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 18 / 72
 Supervised Classification
 Drawing Tools  Region Grow tool
Step 1: Set Seed Properties:

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 19 / 72
 Supervised Classification
 Drawing Tools  Region Grow tool
Step 2: Select a seed
pixel with the Region Note excluded areas
Grow Tool:

Selected seed pixel

Grown
region

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 20 / 72
 Supervised Classification
 Drawing Tools  Region Grow tool
Step 3: Evaluate Spectral Euclidean Distance:

Spectral Euclidean Distance = 15 Spectral Euclidean Distance = 5

using 8‐bit imagery
g g y
With 16‐bit imagery, SED ~ 900 to 2000 are appropriate
David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 21 / 72
 Supervised Classification
 Training Sites – Tools in ERDAS Imagine
 Raster tab Classification Group Supervised

 Signature Editor

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 22 / 72
 Supervised Classification
 Training Sites – Tools in ERDAS Imagine
 Signature Editor Create New Signature(s) from AOI(s)

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 23 / 72
 Supervised Classification
 Training Sites – Tools in ERDAS Imagine
 Signature Editor Evaluate each new signature

 Rename the class to

something meaningful
 Change color to Yellow for Image
Alarm evaluation

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 24 / 72
 Supervised Classification
 Training Sites – Tools in ERDAS Imagine
 Signature Editor View  Image Alarm

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 25 / 72
 Supervised Classification
 Signature Editor View  Image Alarm

1.
2.

3.

2 SD eliminates the darkest and brightest 2.3% of the data

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 26 / 72
 Supervised Classification
 Signature Editor View  Image Alarm
 Set Std. Deviation between:
 2.00 (95.4% of the data)
 3.00 (99.7% of the data)
• 1.00 SD = 68.3% of the data

Rationale:
at o a e Eliminate
ate the
t e most
ost atypical
atyp ca pixel
p e values
a ues from
o co
consideration.
s de at o

Image Alarm uses the simple Parallelepiped classifier in which

each class is defined as an n-dimensional volume whose
b
boundaries
d i iin b brightness
i ht space are sett b
by th
these lilimits.
it

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 27 / 72
 Supervised Classification
 Visually evaluate Image Alarm for:
 Errors of Omission
 Errors of Commission

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 28 / 72
 Supervised Classification
 Errors of Omission
 Indicates the need for and location of additional training
sites to adequately capture the spectral variability of the
class in question.

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 29 / 72
 Supervised Classification
 Errors of Commission
 Indicates that the training site is not spectrally “pure”
regarding the assumed land cover/use class class.
 The analyst must go back to “constrict” the training site
(spectrally) by editing the existing training polygon or
drawing a new one.
 General rule: “Do no harm” ((i.e.,, it is better to leave some
pixels unclassified (omission errors) rather than to have
large commission errors).

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 30 / 72
 Digital Image Classification
 Maximum Likelihood Probability classifier
 Without A Priori Probability Information
 Decide unknown measurement vector X is in class i if, and
only if, pi > pj for all i and j out of 1... m possible classes
and
1 
pi  log e | Vi |    X  M i  Vi  X  M i 
1 T 1

2 2 
Where | Vi | is the determinant of the covariance matrix ,  X  M i T is the
transpose
p of the vector  X  M i  and Vi 1 is the inverse of the covariance matrix .
M i is the mean measurement vector for class i, and Vi is the covariance matrix
of class i for bands k through l. Therefore, to assign the measurement vector X
of an unknown pixel to a class, the maximum likelihood decision rule computes
th value
the l pi for
f eachh class
l and d assigns
i the
th pixel i l to
t the
th class
l that
th t has
h the
th largest
l t
(or maximum) value.

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 31 / 72
 Digital Image Classification
 Maximum Likelihood Probability classifier
 Without A Priori Probability Information
 The mean vectors M i and covariance matrix Vi for each
class are estimated from the training data.

1 
pi  log e | Vi |    X  M i  Vi  X  M i 
1 T 1

2 2 
Where | Vi | is the determinant of the covariance matrix ,  X  M i T is the
transpose of the vector  X  M i  and Vi 1 is the inverse of the covariance matrix .
M i is the mean measurement vector for class i, and Vi is the covariance matrix
of class i for bands k through l. Therefore, to assign the measurement vector X
of an unknown pixel to a class, the maximum likelihood decision rule computes
the value pi for each class and assigns the pixel to the class that has the largest
(or maximum) value.

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 32 / 72
 Digital Image Classification
 Maximum Likelihood Probability classifier
Unknown
Unknown p pixel 1
is assigned to
70% probability
pixels:
class H because
80% probability
•1
90% probability
p1H
1 > p1R
1 >> p1J
•2
•3 1

Unknown pixel 2
is assigned to
class R because
p2R > p2J >> p2H
Red

Unknown pixel 3
n-dimensional is assigned to
class means class R because
p3R > p3H >> p3J

NIR
David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 33 / 72
 Digital Image Classification
 Maximum Likelihood Probability classifier
90

90
80

80 70

70 60

50
60

40
50

30
40
20

30
10

20

10

Adapted from Jensen, 2005

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 34 / 72
 Digital Image Classification
 Maximum Likelihood Probability classifier
 Advantages
 Sensitive to both variance and covariance in the training data

 Disadvantages
 Training data must be normally distributed
 The per-pixel, rather than pixel neighborhood approach,
remains a problem

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 35 / 72
 Digital Image Classification
 Maximum Likelihood Probability classifier
 Multi-band Variance,
Variance Covariance and Correlation

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 36 / 72
 Digital Image Classification
 Maximum Likelihood Probability classifier
 Variance – Covariance Matrix
211 = 5192.16 212 = 3866.65 213 = 2722.83 214 = 1094.98
222 = 3781.67 223 = 2520.50 224 = 1462.00
233 = 4806.52 234 = 3652.66
244 = 3927.78
Sum of variances = 17708.13

4-band data set (range 0 – 255)

Mean band 1 = 129.29

129 29
Mean band 2 = 106.63
Mean band 3 = 104.71
Mean band 4 = 121.80

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 37 / 72
 Supervised Classification
 Maximum Likelihood Classifier
 Raster tab Classification Group Supervised

2 3

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 38 / 72
 Supervised Classification
 Maximum Likelihood Classifier
 All pixels are classified initially, irrespective of how low the
probabilities of class membership are are.
 E.g., assume the probabilities of pixel i belonging to one
of eight classes is:
 p1 = 0.0125
 p2 = 0.0237
 p3 = 0.0105
Pixel i is assigned to Class 5 even though its
 p4 = 0.0095 probability of class membership is only
 p5 = 0.0275
0 0275 2.75%
2 75% (i(i.e.,
e itit’ss 97.25%
97 25% likely to NOT be a
member of Class 5 (and even less likely to be
 p6 = 0.0078 a member of any other class).
 p7 = 0.0083
 p8 = 0.0195
David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 39 / 72
 Supervised Classification
 Maximum Likelihood Classifier
 Threshold tool (Raster tab Supervised Threshold)
 Thresholding is a statistical method used to refine a supervised
classification by determining which pixels in the new thematic
raster layer are most likely to be incorrectly classified (using
the distance file that was created during the supervised
classification). These distances occur in probability space, so
larger distances are associated with lower probabilities of class
membership. p
 The histogram of the distance file usually approximates a chi-
square distribution. Normally, there are many pixels with a
distance image value at or near 0, indicating that these pixels
are likely to be classified correctly. A histogram is generated
for each class in the classified image file.
 Analyst displays
displays, inspects and edits (thresholds) each histogram.
histogram
The rejected pixels become “unclassified.”

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 40 / 72
 Supervised Classification
 Maximum Likelihood Classifier
 Threshold tool (Raster tab Supervised Threshold)

Keep

Reject

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 41 / 72
 Digital Image Classification
 Spectral Angle Mapper
 n-dimensional vectors Hyperangle  < hyperangle ,
S the
So th Unknown
U k pixel
i l iis
classified as a member of
class “A”

255 Unknown (165,155,127)

Known A


155
Band B

 255
Known B

127

0
165 255
Band A

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 42 / 72
 Digital Image Classification
 Spectral Angle Mapper

 nb 
 
1 
 tr

  cos  1 /
i
2 1 / 2 
  t 2   r 2  
nb nb

 i  i 
 i   i  
Where nb is the number of bands in the image, t is the unknown
pixel spectrum, and r is reference spectrum.

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 43 / 72
 Digital Image Classification
 Spectral Angle Mapper
 Given a set of reference signatures
g collected from
an image with m bands, the SAM classifier performs
the following two steps:
– Step 1: Determine the spectral angle  i,r,
between a pixel i in the image and every
reference
f class
l r:  nb 
 
1 
 ti,m rr ,m 
 i ,r  cos 
nb
m
nb 
 
  i ,m  r ,m
t 2
r 2

 m m 
David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 44 / 72
 Digital Image Classification
 Spectral Angle Mapper
– Step 2: Assign each pixel to the reference class r
that has the smallest spectral angular distance
between pixel i and reference class r. For each
pixel i = 1 to n, find the reference class r such that
the  i,r, is the minimum of all reference classes.

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 45 / 72
 Digital Image Classification
 Spectral Angle Mapper Classifier
 Advantages
g
 Often more accurate than maximum likelihood

 Less sensitive to illumination differences

 Disadvantages
 Numerous training spectra strongly

Band A
required to describe the illuminated
“best” mean vector for moderately
each class illuminated
dimly
illuminated
Band B

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 46 / 72
 Supervised Classification
 Spectral Angle Mapper
 Raster tab Classification Group Supervised

2 3

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 47 / 72
 Supervised Classification
 Image Classification
 Classificationscan always be improved - perfection is
rarely
l possible,
ibl especially
i ll using
i precise
i LULC categories
t i

 Per pixel classification works best with moderate spatial

resolution
l ti iimagery (GSD >= 20 meters)t )

The classification process is iterative – it’s rare to have

high classification
f accuracy after
f the first
f attempt

 Consider a hybrid classification – part supervised, part

unsupervised (especially in areas where your field
experience is limited)

 Manual Classification and/or Editing is not cheating, but

can be extremely time consuming
David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 48 / 72
 Digital Image Classification
 Supervised classification
 The analyst "supervises” the categorization of a set
of specific classes by providing training statistics
that identify each category.
 Unsupervised classification
 The raw spectral data are grouped first, based
solelyy on the statistical structure of the data. Then
the analyst must label each statistical cluster,
placing them in the appropriate categories (if
possible).
possible)
 Rule-based classification
 Spectrally categorized pixels are classified using
ancillary data in a model.
David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 49 / 72
 Unsupervised Classification
 Requires only minimal initial input from the
analyst.
 However, the analyst has the task of interpreting
the many classes that are generated by the
unsupervised classification algorithm.
 Uns
Unsupervised
per ised classification is also called
clustering, because it is based on the natural
groupings of pixels in image data when they are
plotted in feature space.

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 50 / 72
 Unsupervised Classification
 Clusters are defined with a clustering algorithm.
 The clusteringg algorithm
g has no regard
g for the
contiguity of the pixels that define each cluster
unlike the training areas for supervised
classification.
l ifi i
 The Iterative Self-Organizing Data Analysis –
ISODATA clustering method uses spectral
distance and iteratively classifies the pixels.

 After each iteration, ISODATA redefines the criteria for

each class, and classifies again, gradually “discovering”
the spectral distance patterns (i.e., the clusters) in the
data.
David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 51 / 72
 Unsupervised Classification
 The initial cluster means are evenly distributed
across the feature space alongg a vector that
runs from the
 multidimensional -1  (1-1, 2-2, 3-3, ... n-n)
to the
 multidimensional +1  (1+1, 2+2, 3+3, ... n+n)

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 52 / 72
 Unsupervised Classification
 ISODATA Initialization

User specified
U ifi d
number of
clusters

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 53 / 72
 Unsupervised Classification
 ISODATA uses minimum spectral distance to
assign a cluster for each candidate pixel.
 The process begins with a user-specified number of
arbitrary cluster means (or the means of existing
signatures).
signatures)
 With each iteration, the means shift to those of the clusters
in the data
data.
 Because ISODATA is iterative, it is not biased to the top of
the data file
file, as are the one-pass
one pass clustering algorithms
algorithms.
 The spectral distance between any candidate pixel and
each cluster mean is calculated and the pixel is assigned
to the cluster whose mean is the closest.
David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 54 / 72
 Unsupervised Classification
 Multispectral Feature Space

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 55 / 72
 Unsupervised Classification
ISODATA first pass

Initial arbitrary
multi-band means

Initial cluster
boundaries

Adapted from Jensen,

Jensen 2005

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 56 / 72
 Unsupervised Classification
ISODATA second pass

New multi-band
cluster means

New cluster
boundaries

Adapted from Jensen, 2005

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 57 / 72
 Unsupervised Classification
ISODATA nth pass

Final multi-band
cluster means

Final cluster
boundaries

Adapted from Jensen, 2005

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 58 / 72
Unsupervised Classification

Adapted from Jensen, 2005

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 59 / 72
 Unsupervised Classification
 ISODATA in ERDAS Imagine
 Raster tab Classification Group Unsupervised

Use only for hybrid classification:

1) ISODATA of small, representative area
2) Cluster interpretation by analyst
3) Cluster Signature Set used in Maximum
Likelihood classifier (with multiple
same-class sigs colored the same)

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 60 / 72
 Unsupervised Classification
 ISODATA in ERDAS Imagine
Step 1 As a generall rule,
A l thi
this should
h ld bbe att
Step 5 least 3x the number of legend
classes you desire + 3 – 5 more
Step 2

Step 3

Step 4 Minimum Size = 1 – 5%

Step 6
Maximum SD = 3.0
3 0 – 4.0
40

Minimum Distance = 4 – 5
Step 7 (for 8-bit images)
Maximum Merges = 10 - 15

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 61 / 72
 Unsupervised Classification
 ISODATA in ERDAS Imagine

Step 8

Step 9

Typically requires 15 – 20 iterations

Step 10
to stabilize (35 – 40 for 16-bit images)

Step 11 Set to 0.95 (0.96– 0.98 will take many

more iterations)

Step 12
Set to X:1,,Y:1
(i.e., do not skip any pixels)

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 62 / 72
 Unsupervised Classification
 After each iteration, the normalized percentage of
pixels whose assignments
p g are unchanged
g since
the last iteration is determined.
 When this number reaches T (the convergence
threshold), the program terminates.
 As a QA
Q check, ensure that the program
terminated because it reached the convergence
threshold and not because it reached the
maximum number of iterations.

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 63 / 72
 Unsupervised Classification
 File Session View Session Log
Performing
g iteration: 1 Convergence:
g 0.000

Performing iteration: 2 Convergence: 0.260

Performing
g iteration: 3 Convergence:
g 0.525

Performing iteration: 4 Convergence: 0.862

Performing
g iteration: 5 Convergence:
g 0.920

Performing iteration: 6 Convergence: 0.938

Performing
g iteration: 7 Convergence:
g 0.949

Performing iteration: 8 Convergence: 0.957

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 64 / 72
 Unsupervised Classification

20 Clusters
6 Iterations

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 65 / 72
 Unsupervised Classification
 Cluster Labeling Attribute Table
Right-click for
pop-up menu.

Select “Display
Attribute Table
Table”

Attribute
Table

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 66 / 72
 Unsupervised Classification
 Cluster Labeling Attribute Table
1. Left-click on Opacity column heading to select all

2. Right-click on Opacity column heading to

open “Column Options” pop-up menu

3. Select “Formula” from menu

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 67 / 72
 Unsupervised Classification
 Cluster Labeling Attribute Table

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 68 / 72
 Unsupervised Classification
 Cluster Labeling Attribute Table
Inspect the
included LULC
classes in each
cluster
l t using
i ththe
SWIPE tool
(HOME tab)

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 69 / 72
 Unsupervised Classification
 Cluster Busting
 It is common when performing unsupervised classification that
the algorithm will generate n clusters, but you have no confidence
in labeling q of them to an appropriate information class (e.g., 15
out of 30 clusters contain multiple LULC categories).
 Step 1 - all the pixels associated with the 15 clusters that are mixed (e.g.,
clusters 3, 5, 6, 9, 11, 13, 15, 17, 18, 20, 22, 23, 25, 28 and 30)) are all RECODED
to a value of 1 and a binary mask file is created.

 Step 2 - A mask program is run multiplying the binary mask file times the
original imagery file.
file The output of the mask program is a new multiband
image file consisting of only the pixels that could not be adequately
labeled during the initial unsupervised classification.

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 70 / 72
 Unsupervised Classification
 Cluster Busting
 Step 3 - Perform a new unsupervised classification on the output file
f
from Step
St 22, perhaps
h requesting
ti ththe original
i i l number
b off clusters.
l t

 Step 4 – Repeat Steps 1 and 2, keeping all of new clusters that are
dominated byy a single
g LCLU (e.g.,
( g , 7).
) There are likelyy still some mixed-class
clusters, but the proportion of these is definitely smaller.

 Step 5 - Repeat Step 3 performing a third iterative unsupervised

classification.
l ifi i P Perhaps
h an additional
ddi i l 3 good
d clusters
l are extracted.
d

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 71 / 72
 Unsupervised Classification
 Cluster Busting
 In this hypothetical example, the final cluster map would be
composed of :
 15 good clusters from the initial classification,
 7 good clusters from the first cluster
cluster-busting
busting pass (recoded as values 31
to 38), and
 3 good clusters from the second cluster-busting pass (recoded as values
39 to 41).

 The final cluster map file is re-assembled using a simple GIS

maximum-dominate
maximum dominate overlay function.
function The final cluster map is then
recoded to create the final classification map.

David P. Lusch, Ph.D. GEO 827 – Digital Image Processing and Analysis October 2015 72 / 72