Todays Schedule

1
Todays Schedule

• 830-900 am Continental breakfast
• 900-1030 am Open lab time
• 1030-1100 am Break
• 1100-1230 pm Workshop 3 Advanced
  Classification and Texture in Definiens
  Professional 5
• 1230-130 pm Lunch
• 130-300 pm Presentation 3 Wetlands, Coastal,
  and Riparian Habitats
• 300-330 pm Break
• 330-500 pm Working Group Discussion Special
  Issues, Proposal Development, Future
  Directions
• 500 pm Adjourn

2
Workshop 3Advanced Classification and Texture
Analysisfor Definiens Professional 5

• Instructor Tim De Chant
• dechant_at_nature.berkeley.edu
• OBIA Symposium 2007
• June 7-8, 2007
• Geospatial Imaging Informatics Facility
• http//giif.cnr.berkeley.edu

3
Outline

• Goal To improve feature exploration and
  selection with a focus on texture
• Advanced classification
• Feature View tool
• Feature Space Optimization
• Texture analysis
• Inherent vs. Contextual features
• Haralick texture
• Contextual texture

Exercise Explore features using Feature View and
Feature Space Optimization, analyze and
implement texture
4
Advanced Classification

• Feature View
• visualizing the differences between objects
  statistics and attributes
• Feature Space Optimization
• getting a handle on the myriad of features to
  better separate your classes

5
Feature View

• Useful tool in determining classification rules
• aids in isolating objects in the landscape by
  their spectral, textural, or contextual
  characteristics

6
Pros and Cons of Feature View
Advantages
Disadvantages

• Quicker than clicking from object to object to
  compare object values and characteristics
• Helps in visualizing potential trade-offs in
  the fuzzy classification
• which features may help classification, which may
  hurt it

• May still need to click from object to object to
  determine precise values for fuzzy rulesets
• Can be slow
• especially so when using Texture after Haralick
  features

7
Feature Space Optimization

• Intended to provide an ordination-like way of
  creating standard nearest neighbor rules for
  classification
• Can also be used to provide guidance in your own
  classification through its ordination-like
  results

8
Pros and Cons of FSO
Advantages
Disadvantages

• Helps in distinguishing which features may best
  separate classes of interest
• Saves time when exploring hundreds of possible
  features

• Does not perform well when results applied to
  classes (personal experience)
• Seems to require all classes with samples be
  analyzed

9
Other Notes on FSO

• FSO results are best used as guides
• other features may still be important in the
  classification
• Like most ordinations, only the first few
  dimensions may be useful

10
Texture Analysis

• Inherent vs. Contextual Features
• Background
• Haralick texture
• Contextual texture

11
Inherent vs. Contextual Features

• Inherent features
• features essential to an images pixels
• Contextual features
• features that express the relationship of objects
  to one another at the same or different scales

12
Inherent vs. Contextual Features

• Whats available in Definiens Professional
• Inherent
• spectral means, ratios, and standard deviations,
  Texture after Haralick, and more
• Contextual
• context-based texture (mean SD of sub-objects,
  area of sub-objects, etc.), relation to
  neighbors, position in image, and more
• Whats not available
• PCA, NDVI, tasseled cap, etc.

13
Basics of Texture

• Formalized by Robert M. Haralick in his often
  cited 1973 and 1979 papers
• Image texture is essentially the spatial
  variation of DN within an area
• for practical purposes, the area is usually user
  defined

Robert M. Haralick
14
Haralicks definition

• The number and types of an images primitives
  and the spatial organization or layout of its
  primitives1
• organization may be random, pairwise dependent,
  or n-wise dependent
• dependence can be structural, probabilistic, or
  functional

Robert M. Haralick GLCM Dissimilarity Green
(10, 20) Level 10
1Haralick, Robert M. 1979. Statistical and
Structural Approaches to Texture. Proceedings of
the IEEE 67(5) 786-804.
15
Differing texture concepts

• Haralicks texture is an inherent feature
• segmentation is not required to get at Haralicks
  texture measures (think ENVI)
• Definiens then calculates an objects value from
  these inherent features
• Contextual texture is a newer concept used in
  Definiens Professional
• how a parent objects sub-objects vary
• example - mean area of the sub-objects

16
Observations on Texture

• Texture is highly dependent on your image
• How was the image taken or processed?
• Is it mosaicked?
• Has it been properly registered and
  radiometrically corrected?
• What time of day was the image taken?
• What is in the image?
• Is it a patterned landscape or more random?
• Noise can also influence the usefulness of
  texture.

17
To precalculate or not to precalculate?

• Can speed analysis and classification
• Texture after Haralick in Definiens Professional
• Features that may not be available
• PCA, tasseled cap, other enhancements
• while custom features are available, implementing
  something like PCA in a custom feature, while
  good for posterity, may be a waste of your time
• Texture at variable moving window sizes

18
Summary

• Definiens Professional provides two useful tools
  for visualizing and analyzing your data
• Feature View
• Feature Space Optimization
• Texture is cool
• Cool things may or may not be useful.

may be useful
may not
19
Exercises

• Open Definiens Professional 5 and start with the
  Grassland project (C\OBIA\Workshop3)
• Begin exercises in workshop handbook

20
Key Points

• Feature View
• Feature Space Optimization
• Texture
• Texture after Haralick in Definiens Professional
• Contextual texture
• Using precalculated texture layers (also
  Haralick-based)