Digital Image Classification Techniques

Digital Image Classification
Uses the spectral information in one or more

spectral bands.
Classifies each individual pixel based on this
spectral information. Also known as spectral
pattern recognition or cluster analysis.
 The objective: To assign all pixels in the image to
particular classes or themes (e.g., water,
coniferous forest, deciduous forest, crops, bare
soil, etc.)
Geo 424 – Advanced Remote Sensing


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

 The two classes of approach:
 Per-pixel classification
• The algorithm categorizes each input pixel into a
spectral feature class based solely on its
individual multispectral vector (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. The objects are classified using both
spectral and spatial cues.

Within Per-Pixel Classification there are 3 types:
 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
solely on the statistical structure of the data. Then
the analyst must label each statistical cluster,
placing them in the appropriate categories.
 Rule-based classification
 Spectrally categorized pixels are classified using
ancillary data in a model.

Aside about cluster analysis in general…
- Back to our example dataset!

- TMAX versus clay content

[when a company you’ve shopped with before

decides to send you an ad or a coupon, that’s
partially thanks to cluster analysis!]
Think about your last online order…
In Remote Sensing, the data we often use is

spectral, but could be any raster data.

 You decide the ‘centroids’ of the clusters.
 In the imagery, you identify 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 cover types
present in the image. Visual interpretation skills are
mandatory for success.

Spectral
Training Sites Signatures
Conifer
The computer software compiles

Conifers
Water
Used to classify
Water all unknown
Deciduous pixels
Deciduous
Digital Image

 Supervised classifiers (algorithms)
 Parallelepiped – 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 Mapping – class membership based on

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

 Supervised classification process
 Step 1: Define training data
 Each training site should appear homogenous and
representative of the legend class
 Delineate several training sites within 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)

 Training 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

 Example of a homogeneous training site
Check all bands

 Example of a heterogeneous training site

 Training data
 Should be normally distributed if the maximum
likelihood classifier is going to be used
Darker DNs Brighter DNs

 Supervised Classification Process
 Training Sites – use several for each class 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 - 1σ +1σ - 1σ +1σ Brighter DNs

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

 Supervised Classification Process
 Training Sites – for variable classes do NOT
merge signatures
Darker DNs - 1σ +1σ Brighter DNs
µ
The disadvantage of merged signatures:
shifted mean DN value (atypical of the class) and a large variance
 Training site pixels
Red
NIR

 Parallelepiped decision rule
Decision spaces
based on range Unknown pixels
(min – max)
in this example
1&2
are assigned to
class R
Unknown
pixels:
•1
Unknown pixel
3 is unclassified
Red
•2
•3 because of
decision-space
overlap
NIR
 Advantages
 Sensitive to different degrees of class variance in the data
 Extremely fast, since it’s a basic table look-up function
 Disadvantages
 Problems occur in regions of spectral overlap
 Creates “oversized” decision spaces in the presence of large
inter-band covariance, leading to between-class overlaps

 Minimum Distance-to-Means rule
Unknown pixel 1
Euclidean distance from
unknown pixel 1 to the is assigned to
three class mean values
class H because
D1-H < D1-R < D1-J
Unknown
pixels:
•1 Unknown pixel 2
•2 is assigned to
•3 class J because
D2-J < D2-R < D2-H
Red
Unknown pixel 3
n-dimensional is assigned to
class means class H because
D3-H < D3-R < D3-J
NIR
 Euclidian distance uses…
 Pythagorean Theorem!?

 Euclidian distance
Red
NIR
 Euclidean distance measures
Distance – two dimensions
d = ( x2 − x1 ) + ( y2 − y1 )
2 2
Distance – n dimensions
n =1
∑ (DN )
n 2
d a −b = ai − DN bi d = x2 – x1
2
i =1

 Advantages
 Fast, because it’s computationally efficient
 Disadvantages
 Insensitive to different degrees of class variance and

covariance in the data (because you only care about the
mean), which can cause misclassification

 Supervised classifiers (algorithms)
 Parallelepiped – 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 Mapping – class membership based on

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

 Advantages
 Sensitive to different degrees of class variance in the data
 Extremely fast, since it’s a basic table look-up function
 Disadvantages
 Problems occur in regions of spectral overlap
 Creates “oversized” decision spaces in the presence of large
inter-band covariance, leading to between-class overlaps

 Advantages
 Fast, because it’s computationally efficient
 Disadvantages
 Insensitive to different degrees of class variance and

covariance in the data (because you only care about the
mean), which can cause misclassification

 Probability density functions (Max Likelihood)
Unknown
Unknown pixel 1
pixels: is assigned to
•1 class H because
p1H > p1R >> p1J
•2
•3
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
 Probability classification
p(31 RP)
p(31 H)
Unknown
pixel

 Maximum Likelihood
 Advantages
 Sensitive to both variance and covariance in the training data
 More accurate than parallelepiped or minimum- distance-to-
means
 Disadvantages
 Slower, because its computationally intensive
 Training data must be normally distributed

 Spectral Angle Mapper Classifier
255
Unknown (15,155)
155
Known A
α Hyperangle α < hyperangle β,

So the Unknown pixel is
Band B
β classified as a member of
class “A”
Known B
0
165 255
Band A

 Spectral Angle Mapper Classifier
 Advantages
 It’s computationally efficient
 Often more accurate than maximum likelihood
 Less sensitive to illumination differences
 Disadvantages
strongly
Band A
 Numerous training spectra illuminated
required to describe the
moderately
“best” mean vector for illuminated
each class dimly
illuminated
Band B

Advanced Supervised Classification Techniques
(1) “Neural Networks”
- This is what lots of image recognition and facial recognition
software does.
- One example = Support Vector Machine (SVM)
mlgenius.com

(2) “Classification Trees”
- Binary decisions (a ‘tree)
- RandomForest is a common example
classes.cs.uchigago.edu

(2) “Dyadic Partitioning”
- Another binary classifier
- Less frequently used in remote sensing
classes.cs.uchigago.edu

 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
solely on the statistical structure of the data. Then
the analyst must label each statistical cluster,
placing them in the appropriate categories (if
possible).
 Rule-based classification
 Spectrally categorized pixels are classified using
ancillary data in a model.
 Image Classification
 Classifications are never really complete – they end when
time and money run out
 The classification process is iterative – it’s rare to have

high classification accuracy after the first 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

Digital Image Classification Techniques

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Digital Image Classification Techniques

Uploaded by

Copyright:

Available Formats

Digital Image Classification

Uses the spectral information in one or more

Geo 424 – Advanced Remote Sensing

Geo 424 – Advanced Remote Sensing

Geo 424 – Advanced Remote Sensing

Geo 424 – Advanced Remote Sensing

Geo 424 – Advanced Remote Sensing

Geo 424 – Advanced Remote Sensing

Geo 424 – Advanced Remote Sensing

[when a company you’ve shopped with before

Think about your last online order…

In Remote Sensing, the data we often use is

Geo 424 – Advanced Remote Sensing

Geo 424 – Advanced Remote Sensing

The computer software compiles

Geo 424 – Advanced Remote Sensing

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

 Maximum Likelihood – based on probability of class

 Spectral Angle Mapping – class membership based on

Geo 424 – Advanced Remote Sensing

 Delineate several training sites within each legend class

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

Geo 424 – Advanced Remote Sensing

Geo 424 – Advanced Remote Sensing

Check all bands

Geo 424 – Advanced Remote Sensing

Geo 424 – Advanced Remote Sensing

Darker DNs Brighter DNs

Geo 424 – Advanced Remote Sensing

A dark version A bright version

Darker DNs - 1σ +1σ - 1σ +1σ Brighter DNs

Geo 424 – Advanced Remote Sensing

Darker DNs - 1σ +1σ Brighter DNs

Geo 424 – Advanced Remote Sensing

Geo 424 – Advanced Remote Sensing

Geo 424 – Advanced Remote Sensing

Distance – two dimensions

Geo 424 – Advanced Remote Sensing

 Insensitive to different degrees of class variance and

Geo 424 – Advanced Remote Sensing

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

 Maximum Likelihood – based on probability of class

 Spectral Angle Mapping – class membership based on

Geo 424 – Advanced Remote Sensing

Geo 424 – Advanced Remote Sensing

 Insensitive to different degrees of class variance and

Geo 424 – Advanced Remote Sensing

Geo 424 – Advanced Remote Sensing

Geo 424 – Advanced Remote Sensing

α Hyperangle α < hyperangle β,

Geo 424 – Advanced Remote Sensing

Geo 424 – Advanced Remote Sensing

Geo 424 – Advanced Remote Sensing

Geo 424 – Advanced Remote Sensing

Geo 424 – Advanced Remote Sensing

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

 Consider a hybrid classification – part supervised, part

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

Geo 424 – Advanced Remote Sensing

You might also like