Spatio-temporal contextual image classification based on spatial AdaBoost

Spatial AdaBoost proposed by Nishii and Eguchi (TGRS 2005) is a contextual supervised classifier of land-cover categories of geostatistical data. It shows an excellent performance similar to that of the MRF-based classifier with much less computational cost. In this paper, we extend the method to the setup with multi spatio-temporal images. We take classification functions by the averages of log posterior probabilities derived by respective training data sets. The functions are sequentially combined by minimizing the empirical exponential risk calculated over samples in all the training data sets. Thus, we obtain a classifier based on a convex combination of the functions. The proposed method is applied to artificial data, and it shows performance similar to that of Spatial AdaBoost based on much larger training data.