A New Tree-based Classifier for Satellite Images

Remote sensing methods have extensively been applied in many areas, for instance, hydrological modeling, wildlife habitat modeling, forest degradation monitoring, wetland biodiversity conservation and disaster management. Different types of sensors mounted on a satellite generate images with different spatial, spectral, radiometric and temporal resolutions. Due to synoptic view and repetitive coverage, remote sensing is a fast emerging source of large datasets on natural resources in spatial format. One of the most important products of a raw satellite image is the classified map which labels the image pixels in meaningful classes. Though several parametric and non-parametric classifiers have been developed so far, reliable prediction of pixel-wise class labels still remains a challenge. The inaccuracy and uncertainty in prediction can often be attributed to the complexity of study area terrain, sensor characteristics, spectral mixing and size of the training data. In this chapter, we are promoting a new classifier called mBACT – a multiclass generalization of Bayesian Additive Classification Tree (BACT) for classifying satellite images. BACT is based on the ensemble of trees model called Bayesian Additive Regression Tree (BART) proposed by Chipman, George & McCulloch (2010), as a flexible regression model. Chipman et al. (2010) also developed a binary classifier called BART-probit for a drug discovery application. Independently, Zhang & Hardle (2010) used BART to develop BACT for classifying binary data in credit risk modeling setup. We compare the performance of mBACT with several state-of-the-art classifiers in remote sensing literature (i.e., support vector machine (SVM) and classification and regression tree (CART)) for predicting pixel-wise class labels of a satellite image. The data considered in this study is a portion of a LANDSAT 5 TM image (with six reflectance bands) covering the town of Kentville, Nova Scotia, Canada (see Agarwal, Ranjan & Chipman, 2013 for a detailed case study).