Fuzzy Classification of Remote Sensing Image.

Abstract

The conventional method for classification of remote sensing image is based on Bayes' theorem. The applied condition is to be “each pixel must belong to either of the classes”. In other words, neither the pixel belonging to more than one class ('mixed' pixel) nor the pixel belonging to none of the classes ('unknown' pixel) can be allowed. However, the 'mixed' pixel is necessarily existent in the case of satellite imagery. The existence of 'unknown' pixel is also inevitable as the number of class settings is restricted. This paper proposes the fuzzy classification of the remote sensing image. The classes are defined as fuzzy sets. With this the 'mixed' and 'unknown' pixels can be theoretically considered by the fuzzy set operations. An important problem is how to give a membership function for each class in multi-spectral space. In this paper the membership function is calibrated on the least squares criteria from the training data by the back propagation algorithm of neural network model. The performance of the proposed fuzzy classification method is evaluated in comparison with the conventional supervised classification method. This paper also discusses a method to effectively visualize the fuzzy classification result using RGB color composite.