Comparison of Minimum Distance and Maximum Likelihood Techniques for Proportion Estimation.

摘要

The estimation of mixing proportions p sub 1, p sub 2,..., p sub m in a mixture density is often encountered in agricultural remote sensing problems in which case the p sub i's usually represent crop proportions. In these remote sensing applications, component densities f sub i(x) have typically been assumed to be normally distributed, and parameter estimation has been accomplished using maximum likelihood (ML) techniques. In this paper the authors examine minimum distance (MD) estimation as an alternative to ML where, in this investigation, both procedures are based upon normal components. Results indicate that ML techniques are superior to MD when component distributions actually are normal, while MD estimation provides better estimates than ML under symmetric departures from normality. When component distributions are not symmetric, however, it is seen that neither of these normal based techniques provides satisfactory results.