The EM-based Maximum Likelihood Classifier for Remotely Sensed Data

Based on parametric density distribution model, the maximum likelihood classification (MLC) might be one of the most popular methods for remote sensing image classification. By comparison with non parametric approaches, MLC has several distinct advantages, such as its clear parametric interpretability, feasible integration with prior knowledge based on Bayesian theory, and relative simple realization, etc. However, remote sensing information has some degree of definite statistical characteristic, but as well as holds the high randomness and complexity, which generally behaves as mixture density distribution in feature space. If the distributions of certain categories in feature space are so discrete that they might not obey to the single assumed distribution, or the training samples are not sufficient enough that they can not represent the overall distributions, it often brings on great bias between obtained results and practical situations. In this article, we firstly introduce into the expectation maximization (EM) algorithm in order to extend the conventional MLC approach to mixture density model. EM assumes that the overall distribution could be decomposed into infinite parametric distributions. The model should be firstly assumed that whole distribution could be separated into infinite parametric density distributions, then by EM iterative computation the maximum likelihood parameters of each proportional distribution can be estimated. Better parameter estimates can be obtained by exploiting a large number of unlabeled samples in addition to training samples using the EM algorithm under the mixture model. Furthermore, we present out the framework and detailed process of EM MLC for remote sensing image classification. By experimental case, the EM MLC classification algorithm is compared with conventional MLC algorithm qualitatively and quantitatively. The results show that the EM MLC could obtain higher accuracy than MLC.