This master thesis addresses the unsupervised statistical classification to remotely sensed images based on mixture estimation. The application of two well-known techniques, Expectation Maximization (EM) and Stochastic EM (SEM) algorithm to multidimensional image data is to be investigated, where Gaussian mixture is assumed. The initialization parameters are estimated by two different procedures namely the K-means algorithm and the rough sets algorithm. Relative entropy is adopted as the criterion to measure the discrepancy between the original image histogram and the estimated probability distribution function from mixture estimation. The set of parameters yielding the smallest discrepancy is treated as the result. Since most remotely sensed images are nonstationary, adaptive algorithms, AEM and ASEM, will also be explored by localizing the estimation process. Both color and multispectral data are in the experiments for performance analysis. In particular, comparative study is conducted to quantify the differences among versions. |