Contextual Image Classification Based on Spatial Boosting

Spatial AdaBoost proposed by Nishii and Eguchi (TGRS, 2005) is a supervised image classification method. It is a voting machine based on log posterior probabilities at a test pixel and its neighbors. The method can be obtained by less computation effort with respect to a classifier based on Markov random fields, but still shows a similar excellent performance. Further, the method was modified for applying various settings. This paper considers another extension of Spatial Boost. Consider supervised image classification of geospatial data. Suppose that separated training regions with a single land-cover class are given. In this case, the original Spatial Boost does not work because it incorporates spatial information of the training data. The aim of the paper is to make Spatial Boost applicable to the case. We propose a classifier given by a linear combination of log posteriors whose coefficients are determined by spatial information of test data only. By numerical examples, it shows an excellent performance.