Potential%20of%20Ant%20Colony%20Optimization%20to%20Satellite%20Image%20Classification

1
Potential of Ant Colony Optimization to Satellite
Image Classification

• Raj P. Divakaran

2
Remote Sensing The Definition
Remote Sensing is defined as the art and science
of obtaining information about any object without
coming in direct contact with it. The term is
associated with the process of obtaining
information about the earth.
3
Schematic Diagram
Reflectance detected by sensor
Incoming Radiation
4
Energy Balance and Spectral Patterns

• E(I) E(R) E(T) E(A)
• Proportion of energy reflected, transmitted or
  absorbed depends on the feature on the earth.
• Same feature has different spectral
  characteristics at different wavelengths
• Different features have distinct spectral
  patterns

5
Spectral Patterns Of Different Features
6
Satellite Image Classification

• Involves grouping pixels into distinct labels
• Necessary for subsequent analysis of the image
• Image Classification Process
• Supervised
• Unsupervised

7
Need For Automated Classification Techniques

• Very large amount of data being generated
• Insufficient skilled manpower
• Result data not utilized to full extent
• Need for quick reference to potentially useful
  imagery

8
ACO

• ACO algorithms take inspiration from the
  coordinated behavior of ant swarms.
• ACO algorithms strive to generate intelligent
  systems by emphasising on emergence,
  distributed-ness and autonomy
• Bottom-up approach

9
How Can ACO Help?

• Does not require the user to create training data
  samples
• Does not assume an underlying statistical
  distribution for the pixel data
• Contextual information can be taken into account.
  i.e., neighbourhood information for a pixel
• Could improve robustness because the solution is
  not hard-coded

10
Kenge GIS Extension to Swarm

• Developed to make GIS layers accessible for
  agent-based simulations
• Multiple layers can be brought into the same
  Kenge object interface

11
Approach

• Create n groups of ants
• Make each group search for a particular spectral
  combination that corresponds to a distinct
  feature - vegetation, clouds etc.
• Allow them to mark a pixel to indicate they have
  'classified' a pixel as belonging to a particular
  category (analogous to trail laying in real ants)
• An ant belonging to group 'a' tends to move onto
  pixels marked as belonging to the same group
  (analogous to trail following in real ants)
• Trails can be strengthened by this method
• Multiple ants can lay trails over the same pixel
  and the pixel is classified as belonging to the
  group with the highest trail strength

12
Work Done

• Created 3 types of ants - vegetation seeking,
  water seeking and bare soil seeking ants
• Assigned variable parallelopipeds (spectral
  bounds) to water seeking and bare soil seeking
  ants
• Assigned the NDVI factor for vegetation seeking
  ants
• NDVI Normalized Difference Vegetation Index
• NDVI (NIR Red) / (NIR Red)
• High values ( values close to 1) mean healthy
  vegetation

13
Ant Movement Rules

• An ant searches for a trail in its immediate
  neighborhood
• If trail exists, it moves to the pixel with the
  trail mark
• Else, moves randomly
• If ant happiness reaches threshold point
  exploratory mode is triggered. The ant moves
  randomly

14
Preliminary Results
Original Image
After Classification
15
Classification (after 1000 time steps)
16
Preliminary Results (contd.)
17
Plus Points

• Ants are able to classify the image!
• Able to identify vegetation
• Able to identify bare soil

18
Minus Points

• Buildings wrongly identified as soil
• Shadows mistaken for water bodies

19
Work That Remains To Be Done

• Formulate mechanisms to identify urban areas,
  clouds
• Incorporate diffusion of trails
• Incorporate recruitment of ants to other groups
• Incorporate contextual searching mechanisms
  (could become vital for delineating urban areas)
• Comparison with standard techniques
• Comparison with ground data