Classification of Remote Sensed Data A Soft Computing Approach

1
Classification of Remote Sensed Data-A Soft
Computing Approach

• Aditya Saurabh (2001001)
• Raghu B.V. (2001016)
• An MHRD Project
• Under
• Dr. Anupam Agrawal

2
What we did?

• Collection of Data
• Satellite Images LISS and PAN
• Preparation of Data
• Image to Map Registration
• Fusion
• PCA
• Classification
• ANN-BP
• GA-BP
• Modular ANN
• Accuracy Assessment

3
Data Preparation
LISS99
PAN99
Register
Register
Map
Map
Registered PAN
Registered LISS
Fuse
4
Classification
LISS Image
Fused Image
Classify
ANN-BP
GA-BP
Modular ANN
5
Accuracy Assessment
Reference Data or Classified Map B
Classified Map A
Accuracy Assessment
Kappa Coefficient
Producers Accuracy
Confusion Matrix
Users Accuracy
Overall Accuracy
6
Back-propagation Network

• It is widely used in the classification of multi
  spectral images.
• Limitations
• Can be easily trapped into Local Minima.
• Learning rate is slow.
• Huge number of training samples are required.
• Uses gradient descent (Greedy Algorithm). So,
  transfer function must be differentiable.

7
GA-BP

• Uses Evolutionary Algorithm for initialization of
  weights.
• Back-propagation for fine tuning the Neural
  Network weights.
• Population (wi , n1)
• Selection Process Tournament Selection Scheme.
• Mutation wi Gaussian-Random(0,1)n1
• Fitness Function Ranking on MSE

8

• Advantages
• Evolution Learning Adaptation.
• Probability of reaching global minimum higher.
• Less sensitive to initial weights.
• Disadvantages
• Computationally expensive for small networks.

9
Modular Neural Networks

• Reverse Engineering Biological System.
• Optimal Number of Modules is square root of
  classes.
• Stages
• Task Decomposition.
• Module Training.
• Multi-Module Decision Making.

Task Decomposition (ART)
Module Training (BP)
Multi-Module Decision Making (Voting Layer
Democracy)
10
(No Transcript)
11

• Advantages
• Cooperation between intelligent workers yields
  better capability than summation of Individuals.
• Scalability
• Extendibility.
• Competition Cooperation
• Parallelizable.
• Disadvantages
• Computationally Expensive on 1p machine.

12
Papers

• 1 G. Auda, M. Kamel, and H. Raafat. Modular
  neural network architectures for classification.
  In International Conference on Neural Networks,
  volume 2, pages 1279-1284, Washington D.C., June
  1996.
• 2 G. Bartfai. Hierarchical clustering with ART
  neural networks. In World Congress on
  Computational Intelligence, volume 2, pages
  940-944, Florida, June 1994.
• 3 G. Auda, M. Kamel. Cooperative Modular Neural
  Network. IEEE, 1997.
• 4 XIN YAO. Evolving Artificial Neural Networks.
  PROCEEDINGS OF THE IEEE, VOL. 87, NO. 9,
  SEPTEMBER 1999.

13
Books

• Computer Processing of Remotely-Sensed Images, An
  Introduction, Second Edition, Paul M. Mather.
  1999 http//www.geog.nottingham.ac.uk/mather/
• Introductory Digital Image Processing A Remote
  Sensing Perspective (2nd Edition) by John R.
  Jensen
• Machine Vision and Advanced Image Processing in
  Remote Sensing by Ioannis Kannellopoulos, Graeme
  G. Wilkinson, Theo Moons 1999
• Classification Methods for Remotely Sensed Data
  by Brandt Tso and Paul M. Mather 2001

14
Thank YouQuestions?
http//profile.iiita.ac.in/asaurabh_00/MyProjects/
7thSemester/