The most common method for labeling multispectral image data classifies each pixel entirely on the basis of its own spectral signature. Such a method neither utilizes contextual information in the image nor does it incorporate secondary information related to the scene. This exclusion is generally due to the poor cost/performance efficiency of most contextual algorithms and a lack of knowledge concerning how to relate variables from different sources. In this research, several efficient spatial context measures are developed from different structural models for four-nearest-neighbor neighborhoods. Most of these measures rely on simple manipulations of label probabilities generated by a noncontextual classifier. They are efficient computationally and are effective in improving classification accuracy over the noncontextual result. Among other schemata, the measures include: average label probabilities in a neighborhood; label probabilities; combined as a function of a metric in the label probability space; and context through semantic constraints within a Bayesian framework. In addition, an efficient implementation of a contextual classifier based on compound decision theory is developed through a simplification of the structure of the contextual prior probability^ No accuracy is lost through the simplification, but computational speed is increased 15-fold. Finally, a procedure to combine label probabilities from independent data sources is proposed. A mechanism for combining the label probabilities from each of the sources as a function of their independent classification accuracies is created and evaluated.
Date of this Version