A self-organizing map model for spatial and temporal contextual classification

This paper opens up a new prospect for exploring the Kohonen self-organizing map (SOM) model of multiple SOMs (MSOM, Wan-Fraser, 1994) for spatial and temporal contextual classification of remote sensing (RS) images. Since the model can effectively exploit correlation information within the data in the sense of clustering and autoassociation without parametric models, it can be treated as a unified framework to exploit various correlations as clusters within and among multi-sources and across spectral, spatial, temporal and labeling contexts of images and associated label maps in a recursive manner. It can significantly improve classification accuracy, provided that various spatial and temporal correlations are coded and represented explicitly for compound clustering and autoassociation. The underlying mechanisms for contextual analysis are Markov random field (MRF) and stochastic relaxation by use of recursive neighborhood operators with no parametric models. The capability of the model is demonstrated by a bitemporal TM data experiment