Learning a Spatial Ensemble of Classifiers for Raster Classification: A Summary of Results

Abstract

Given a spatial raster framework F, a set of explanatory feature maps, training and test samples with class labels on F, as well as a base classifier type, the problem of ensemble learning in raster classification aims to learn a collection of base classifiers to minimize classification errors. The problem has important societal applications such as land cover classification but is challenging due to existence of class ambiguity from spatial heterogeneity, i.e., samples with the same feature values may have distinct class labels in different areas. Many existing approaches are non-spatial ensembles (e.g., bagging, boosting, random forest), which assume that learning samples follow an identical distribution. Some spatial ensemble approaches also exist, which simply partition the raster framework into several regular sub-blocks and combine classification results on each sub-block. However, these existing approaches can not address the class ambiguity issue among pixels. In contrast, this paper proposes a new spatial ensemble approach, which partitions the spatial framework into several spatial footprints to minimize class ambiguity of training samples and then learns a base classifier for each footprint. Experimental evaluations on a real world remote sensing dataset show that the proposed spatial ensemble approach outperforms existing approaches when strong class ambiguity exists.