Adaptive segmentation of remotely sensed images based on stochastic models

This paper discusses the application of stochastic labeling of remotely sensed images. A cooperative, iterative approach to segmentation and model parameter estimation is defined which is a stochastic variant of the expectation maximization (EM) algorithm, adapted to our model. Classical statistical modeling forces each pixel to be associated with exactly one class. This assumption may not be realistic, particularly in the case of satellite data. Our approach allows the possibility of mixed pixels. The labeling used in this technique involves two parts: a hard component, which describes pure pixels, and a soft component, which describes mixed pixels. The technique is illustrated by the classification of a SPOT HRV image. Because of the high resolution of these images, the pixel size is significantly smaller than the size of most of the different regions of interest, so adjacent pixels are likely to have similar labels. In our stochastic expectation maximization (SEM) method the idea that neighboring pixels are similar to one another is expressed by using Gibbs distribution for the priori distribution of regions (labels). This paper also presents a statistical model for the distribution of pixel values within each region. The initial parameters of the model can be estimated by using a K-means clustering or ISODATA, in the case of unsupervised segmentation. These parameters are then modified in each iteration of SEM. In the case of supervised segmentation, the initial parameters can be obtained from a classifier training data set and then re-estimated in SEM method. The reason for this re-estimation is that a set of classification parameters obtained from a classifier training data set may not produce satisfactory results on images which were not used to train the classifier. Our study shows that this SEM method provides reliable model parameter estimators as well as segmentation of the image