No. 1001

Title:

Proposition of New Mixture Model and Mixed Pixel Classification Method for Classification of Multispectral Images

Author(s):

Mitsuhiro, Tomosada (The Institute of Statistical Mathematics);
Hiroe, Tsubaki (University of Tsukuba);

Key words:

mixel; mixture model; mixed pixel classification.

Abstract:

A new mixture model for use in the mixed pixel classification of multispectral images is proposed. The covariance of each possible pixel category is taken into consideration in the mixture model proposed; this kind of analysis has not received much attention in the field of mixed pixel classification to date. This is the first time such a mixture model, which is different from the finite mixture model or independent component analysis, has been proposed. Furthermore, a mixed pixel classification method based on the proposed mixture model is disclosed. The accuracy of the mixed pixel classification method is evaluated by numerical simulation. In an image to be analyzed, there are areas that are regarded as being composed almost entirely of pure pixels and other areas regarded as being composed of pixels each including a mixture of categories, such as at a boundary. The estimated mixing ratio, expected value, and covariance are accurate within ranges accurately estimated from original data. Further, the accuracy of the proposed method is also evaluated when the expected value and covariance for each category, which are calculated from training data, are obtained more accurately. As a result, in the case in which the expected value and covariance are close to the true values, the estimation accuracy of the mixing ratio is high. Finally, the proposed mixed pixel classification method is applied to a real image. Each pixel vector is recovered using the estimated mixing ratio and the expected value of each category. It is found that the difference between the original and recovered values of each pixel vector is the same as the difference derived from numerical simulation.