Land Cover Mapping Using Combination Classifiers

In recent years, large scale land cover maps constructed from remotely sensed data have become important information sources for resource management. For many applications, poor map accuracy limits their usability. This article investigates methods of combining classification rules for improving accuracy in general, and for exploiting spatial information in particular. We examine the performance of these methods and the stacked regression method of Breiman (1996) and Mojirsheibani (1999) along with several variants. In our applications, a land cover map is a partition of an area into contiguous unclassified polygons which are assigned land cover types via a classification rule. Because polygons tend to differ with respect to land cover type, spatial association patterns are largely absent from polygon maps. However, there is some spatial information carried by the training observations. We propose a spatial classifier that uses spatially-close training observations for classification. While the spatial classifer is not particularly accurate, remarkable improvements in estimated accuracy were obtained when it was combined with linear discriminant and -nearest neighbor classifiers. 5