Incorporating Knowledge into Unsupervised Model-Based Clustering for Satellite Images

The identification and classification of landcover types from remotely sensed data is traditionally based on the assumption that pixels with similar spatial distribution patterns belong to the same spectral class. However, spectral data on its own has proven to be insufficient for classification. In addition, it is difficult to obtain enough accurate labelled samples from such data. Contextual data can be incorporated or fused´ with spectral data to improve the estimation of class labels and therefore enhance the accuracy of the classification process as a whole when labelled data is not available. In this paper we use Dempster-Shafer theory of evidence to fuse the output of an unsupervised model-based clustering (MBC) technique and contextual data in the form of a digital elevation model. The final classification accuracy is shown to improve when using this approach.