Advertisement

Pattern Classification from Raster Data Using Vector Lenses, Neural Networks and Expert Systems

  • B. Archie Bowen
  • Jianli Liu
Part of the NATO ASI Series book series (volume 65)

Abstract

Satellite data is often used for a variety of applications in assessing the condition of the state of the earth’s surface or the immediate atmosphere. This problem is usually attacked by processing the raw data (infrared, photographs, radar data, etc.) into a pixel stream and then composing a picture, which is used by experts to form the appropriate decisions. Such applications in oceanography include the assessment of ocean and ice states, in geography in terrestrial resource evaluation, in meteorology in cloud classification, etc.

The assessment of such pictures usually requires an expert whose decisions are based on experience, which is difficult to quantify into algorithmic form or even to reduce to a set of rules. Replicating this experience demands a system which can be taught the non quantifiable aspects of the classification procedure.

This paper deals with the problem of automatically classifying large pictures by utilizing several applications of artificial intelligence. The logical architecture of the classification scheme consists of a system of linear neurons acting as vector lenses, which provide input to a neural network. This combination acts both to reduce the dimensionality of the pixel stream representing the picture and to learn the various classifications. The expert system is used to provide expert opinion in case of difficult or inconclusive decisions and to adjust the parameters of the lens system to account for pattern inhomogeneity.

An analysis of the computational complexity of neural computing [6] shows that large pixel streams demand very large data storage and performance in achieving classical back-propagation training. This large computational complexity renders impossible the direct utilization of neural networks. The immediate need is for a reduction of the dimensionality of the pixel space.

The representation of a picture by a non orthogonal set of vectors [4] establishes the theoretical background for the lens concept. Many types of lenses are possible depending on the application. For classification, a particularly simple choice is the set of prototypical patterns representative of each class to be identified.

Result are presented for a simulation of a set of cloud patterns. The robustness and generalization of the system is explored. These results indicate that a robust system can be generated that can consistently perform classifications.

A design methodology is presented which acts as a guide line for adapting this approach to other similar problems.

Keywords

Neural Network Expert System Data Stream Input Vector Input Neuron 
These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

Preview

Unable to display preview. Download preview PDF.

Unable to display preview. Download preview PDF.

References

  1. 1.
    David E. Rumelhart and James L. McClelland, “Parallel Distributed Processing: Explorations in the Microstructure of Cognition, Volume 1: Foundations,” 1986, The MIT Press, Cambridge, Mass., and London.Google Scholar
  2. 2.
    Louis Garand, “Automatic Recognition of Oceanic Cloud Patterns and its Application to Remote Sensing of Meteorological Parameters,” PH. D. Thesis, September 1986, University of Wisconsin, Madison, Wis.Google Scholar
  3. 3.
    Teuvo Kohonen, “Self-Organization and Associative Memory,” Springer-Verlag, 1987.Google Scholar
  4. 4.
    John Daugman, “Complete Discrete 2-d Gabor Transforms by Neural Networks for Image Analysis and Compression,” IEEE Transactions on Acoustics and Signal Processing, vol. 36, No. 7, July 1988, pp. 1169–1179.CrossRefMATHGoogle Scholar
  5. 5.
    Eric Suand, “Dimensionality-Reduction Using Connectionist Networks,” IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol II, No. 3, March 1989, pp. 304–314.CrossRefGoogle Scholar
  6. 6.
    Jianli Liu, “Analysis of The Computational Complexity of Back Propagation Training.” Technical Report, CompEngSery Ltd., Carling Avenue., Ottawa, Ontario Canada K2B 2E1Google Scholar

Copyright information

© Springer-Verlag Berlin Heidelberg 1990

Authors and Affiliations

  • B. Archie Bowen
    • 1
  • Jianli Liu
    • 2
  1. 1.CompEngServ Ltd. and Carleton UniversityCanada
  2. 2.Carleton UniversityOttawaCanada

Personalised recommendations