Land Cover Classification of High Resolution Images Using Superpixel-based Conditional Random Fields

Abstract

It is a difficulty of classifying remote sensing imagery with high spatial resolution(HSR) ensuring high precision. Conditional random fields(CRFs) have advantages over long-range spatial dependency and directly modeling posterior probability of object class in image processing. Superpixel-based CRFs have more ability to express spatial and even semantic information in an image. Due to the advantages, the paper presents a superpixel-based CRFs model with an association potential defined as an indirect probability output of decision function from Support Vector Machine (SVM) and an interaction potential weighted by common boundary of neighboring superpixels for urban land-cover classification of HSR remote sensing images. Evidential experiments on typical urban scene from Quickbird satellite image and comprehensive analysis have shown that our proposed superpixel-based CRFs model from over-segmented image at finer scales has shown better performance than pixel-based CRFs in classification, and less time is consumed in classification when ?-expansion algorithm based on graph cut is used to infer the model than those methods based on message passing.