Abstract
Multivariate normal distribution for Maximum Likelihood classification (MLH) is widely used for image classification because it is adequate for multispectral imagery data as well as Synthetic Aperture Radar (SAR) data, and is easy to manupulate and is based on theoretical background mathematically. Spectral variability of the multispectral imagery data and textural feature of the SAR, however, are distributed as Chi-square like Probability Density function (PDF) and are ranged from non-zero value to finite value. So that if the MLH is applied to such that features, then classification performance is not good enough due to a mismatching between the real and assumed PDF or Likelihood Function. In order to overcome such this situation, a Maximum Likelihood classification with a simplified beta distribution is proposed in this paper. A difference between classification performances for the Maximum Likelihood classifications with multivariate normal and the simplified beta distributions is clarified with real satellite remote sensing data.
