Introduction to Classification of Remotely Sensed Imagery

1
Introduction to Classification of Remotely Sensed
Imagery

• John Porter

2
Example Thematic Mapper

• The Thematic Mapper is a satellite sensor used in
  recent LANDSAT satellites
• It captures 7 bands
• 30 m spatial resolution for bands 1-5 7, 120 m
  in band 6
• Extent

3
Thematic Mapper Spectral Bands

• Wavelengths in micrometers
• Band 1 0.45-0.52
• Band 2 0.52-0.60
• Band 3 0.63-0.69
• Band 4 0.76-0.90
• Band 5 1.55-1.75
• Band 6 10.40-12.50
• Band 7 2.08-2.35

4
Remote Sensing to GIS
Raw Image
5
Remote Sensing to GIS
Classification
Raw Image
Sometimes steps are omitted or reordered,
depending on the purpose of the analysis
Raster GIS Data Layer
6
Automated Image Processing

• This lecture will focus on automated image
  processing
• Unsupervised Classification
• Supervised Classification
• Tools for Image Processing

7
Classification

• Classification is the process of taking the
  brightness values associated with each pixel
  and using them into assign a class to the
  corresponding output pixels
• Output pixels are NOT continuous. Instead they
  are discrete values that represent a class or
  category (e.g. land cover classes)
• For example, we might decide that pixels with
  the value (20,30,190) (in band 1 thru 3 ,
  respectively) indicate that the output pixel
  should be assigned a value of 10
  corresponding to class 10forest

8
Classification

• The trick in classification is to come up with
  rules that will allow us to translate image
  values (e.g., 10,20,190) into classes (forest,
  grass, water, marsh, urban)
• Usually the land cover classes are associated
  with numerical codes e.g., 1forest, 2grass,
  3water, 4urban

9
Classification Methods

• There are two fundamental approaches to
  classification
• Unsupervised
• The computer selects classes based on clustering
  of brightness values
• Supervised
• You specify the classes to be used and provide
  signatures for each class

10
Unsupervised Classification

• Unsupervised classification refers to a variety
  of different techniques that share some features
  in common
• They use statistical clustering techniques to
  decide which pixels should be grouped together
• With luck, these clusters of pixels will
  correspond to land cover classes
• But the correspondence may not be 11
• One land cover class may be represented by more
  than one cluster (easily fixed by recoding)
• One cluster may represent more than one land
  cover type (not easily fixed may need to
  specify more clusters)

11
Example

• The Image Analyst extension in ArcView uses the
  ISODATA unsupervised clustering technique where
  all you need to specify is the number of desired
  classes

12
Each color represents a different cluster
pixels that may correspond to the land cover
classes you are interested in
13
Recoding

• Following an unsupervised classification, you
  need to go through and assign meaning to each
  of the classes (e.g., class 1 water)
• You can use editing functions to set multiple
  classes that represent the same land cover type
  to a common value
• E.g., if both class 1 and class 5 are water
  (albeit, deep water vs. shallow water), you may
  want to edit all the 5s and change them to 1s

14
Supervised Classification

• In supervised classification you help the
  computer to select signatures that represent
  each land cover class
• Signatures are statistical descriptions of the
  brightness values of a given land cover type
  (e.g., the mean band 1 value, the mean band 2
  value etc.)
• You select signatures using a tool that provides
  seed values

15
Signature Selection

• The key to a good supervised classification is
  proper selection of signatures
• What makes for a good signature?
• Characterizes a land cover type of interest
• Separability needs to be distinguishable from
  other signatures

16
Collecting Signatures

• There are two main methods for obtaining
  signatures
• Area-based where you use a box on the screen to
  select the area to be statistically characterized
  as a signature
• Growing an area where you select a point and
  other similar adjacent points are added to the
  sample
• Note choice of minimum similarity can have a big
  effect on the results of this method

17
(No Transcript)
18
Used the seed tool and clicked here. The
highlighted marsh area was similar in color and
connected to the point so it was added to the
data used to calculate the signature.
19
The Find Like Areas command highlighted all the
other areas that have a similar color (i.e.
similar spectral values) with skill they will
all be Marsh
20
Classification
Added Water

• The process is repeated to add new classes
• Due to color variations within a class, often
  multiple signatures will be needed to capture a
  single cover class
• The classes can be recoded and lumped (using
  basic editing functions) so that they correspond
  to the desired classes.

• But not all water was captured by the class-
  requiring additional signatures

21
Classification

• Once you have a group of signatures defined, you
  can classify your image. There are several
  methods for doing this
• Paralleliped
• Mahalanobis Distance
• Maximum Likelihood
• And others.

22
Paralleliped Classification

• use maximum and minimum values on each individual
  band (as derived from each signature) to decide
  which pixels fall within a given class
• Advantages
• Fast
• Can be used one signature at a time
• Used by ARCVIEW Image Analyst
• Disadvantages
• Does poorer job separating classes
• Uses only a fraction of the information contained
  in the signature data

23
Maximum Likelihood

• Uses statistical techniques to decide which class
  a pixel falls into
• Advantages
• Uses full signature information (mean, variation
  inter-band covariation) to tease apart similar
  classes
• Disadvantages
• More computer intensive
• Not available in ARCVIEW

24
Graphical Description
Pixel dim on band 1, but bright on band 2

• We can use a graph to display where pixels fall
  on two bands at once

25
Graphical Display

• Here is a sample display where Marsh is dark,
  beach is light on both bands and water is bright
  on one band (presumably blue) and dim on the other

Water
Beach
Marsh
26
Paralleliped
Here, paralleliped would work well

• Paralleliped classification uses boxes based on
  statistical measures of the range (e.g. maximum
  and minimum values)

Max on band 1
Min. on band 1
27
Paralleliped
Areas of overlap lead to uncertain results

• However, sometimes paralleliped may do a poor job
  separating classes

255
Band 2
0
255
0
Band 1
28
Maximum Likelihood

• Maximum likelihood uses seed statistics to define
  ellipses for each class

29
Maximum Likelihood

• Ellipses make it easier to separate similar
  classes

255
Band 2
0
255
0
Band 1
30
Software for Image Classification

• ArcView Image Analyst
• Provides basic image classification capabilities
• Unsupervised ISODATA method
• Supervised Paralleliped only
• Also supports georeferencing and some image
  processing (e.g., sharpen edges)
• Relatively easy to use

31
Software

• ERDAS Imagine
• Full suite of advanced image processing and
  classification features
• Unsupervised many different clustering
  techniques available (including ISODATA)
• Supervised better tools for capturing and
  analyzing signatures, many classification methods
  (including paralleliped and maximum likelihood)
• Harder to use (with power comes complexity)

32
(No Transcript)