Multistage unsupervised classification of spatially continuous imagery

A new system of unsupervised analysis has been developed that integrates low-level segmentation and high-level classification of the image processing system. An hierarchical spatial clustering algorithm using the closest neighbor chain has been employed for image segmentation. It results in a unique, exact hierarchy and is very computationally efficient for remotely sensed images acquired over large geographic areas. An appropriate measure of homogeneity for establishing similarity criteria for region growing and selecting the number of the classes for region classification is defined by the approach based on information criteria for model selection. Given an image partition from low-level image segmentation, the regions related to the partition have been classified using the expectation maximization method which is mathematically formulated as an optimization problem through the maximum entropy principle. The class parameters are generated iteratively by maximizing the expected-likelihood based on order statistics to increase robustness of the classification in the presence of outliers