A Spanning Tree Hierarchical Model for Land Cover Classification

Abstract

Image segmentation persists as a major statistical problem, with the volume and complexity of data expanding alongside new technologies. Land cover classification, one of the largest problems in Remote Sensing, provides an important example of image segmentation whose needs transcend the choice of a particular classification method. That is, the challenges associated with land cover classification pervade the analysis process from data pre–processing to estimation of a final land cover map. Multispectral, multitemporal data with inherent spatial relationships have hardly received adequate treatment due to the large size of the data and the presence of missing values. In this chapter we propose a novel, concerted application of methods which provide a unified way to estimate model parameters, impute missing data, reduce dimensionality, and classify land cover. This comprehensive analysis adopts a Bayesian approach which incorporates prior subject matter knowledge to improve the interpretability, efficiency, and versatility of land cover classification. We explore a parsimonious parametric model whose structure allows for a natural application of principal component analysis to the isolate important spectral characteristics while preserving temporal information. Moreover, it allows us to impute missing data and estimate parameters via expectation-maximization. We employ a spanning tree approximation to a lattice Potts model prior to incorporating spatial relationships in a judiciousway and more efficiently access the posterior distribution of the pixel labels. We achieve exact inference of the labels via the centroid estimator. We demonstrate this series of analysis on a set of MODIS data centered on Montreal, Canada.