Using genetic algorithms to improve interpretation of satellite data

ABSTRACT

We investigated a problem involving the automatic classification of satellite pixels. Each image is a 512 by 512 matrix of pixels, each of which consists of 4 channel values. The classification of these pixels into one of five classes is ordinarily an arduous process. A variety of search algorithms have been created to solve optimization problems. One of these search algorithms, genetic algorithms, has been developed from the concepts of Darwinian evolution and natural selection. They have several advantages over other search methodologies which are of use to this problem. They do not need expert knowledge, they evaluate a large number of potential solutions quickly and nearly simultaneously, and they are able to identify near optimal solutions while searching for better answers. The method employed uses genetic algorithms to identify a good representative are chosen, classification category. Once the representatives are chosen, classification of pixels in similar images is easily automated. The results indicate that genetic algorithms can be used to classify pixels from satellite images quickly and with a good degree of reliability.

REFERENCES

• 1: David E. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning, Addison-Wesley Longman Publishing Co., Inc., Boston, MA, 1989
• 2: Grefenstette, John J., Lawrence Davis, Daniel Cerys, <u>Genesis & OOGA: Two Genetic Algorithm Systems<</u>, The Software Partnerships, Melrose, Massachusetts, 1991.
• 3: Tran, Phuong, "Application of the Fuzzy K-Nearest Neighbor Algorithm in Remote Sensing", Poster Session at the ACM/CSC Conference, Phoenix Arizona, March 1994.

INDEX TERMS

Keywords:
automatic classification, genetic algorithms, satellite data, satellite images, search algorithms