An MSOM framework for multi-source fusion and spatio-temporal classification

Presents a unified neural network framework, known as MSOM, for multi-source data fusion and spatio-temporal classification. MSOM was originally developed as a classifier-design framework and is now extended for joint scene-modeling (i.e., jMSOM) attempting to “fully” exploit the potential of multi-source data of spectral and categorical features as well as their spatio-temporal attributes in a compound fashion. Difficulties of high dimensionality, disparate statistical and geometrical characteristics, and joint spatio-temporal modeling are addressed. Experiments with a bitemporal set show significant improvement by jMSOM over its SOM or GMLC counterparts and any of its sub-models if only part of data sources is used