Indicator-Kriging-Integrated Evidence Theory for Unsupervised Change Detection in Remotely Sensed Imagery

Abstract

This study proposes a novel approach based on indicator kriging and Dempster–Shafer (DS) theory for unsupervised change detection (CD) in remote sensing images (DSK). Indicator kriging is integrated to the standard DS theory. A feature set with four difference images (DIs) providing complementary change information is initially generated. Subsequently, the mass functions for each DI are determined automatically using fuzzy logic, the four pieces of DI evidence are combined by DS theory, and a preliminary CD map is achieved. The preliminary CD map is then divided into three parts adaptively—weakly conflicting part of no change, weakly conflicting part of change, and strongly conflicting part—by calculating the evidence conflict degree for each pixel. Finally, the pixels in the weakly conflicting parts, which have little or no conflict, are labeled as the current class, and the pixels in the strongly conflicting part that contains misclassified pixels are reclassified based on indicator kriging. DSK combines the advantages of different DI features and solves the conflicting situations to a large extent. The main contributions of this study include the following: 1) introducing indicator kriging into CD to manage conflict information during DS fusion and 2) presenting a scheme for producing DI set with complementary change information, developing a novel DSK fusion model for information fusion, and defining the proposed CD framework. Experimental results verify that the proposed DSK is robust and effective for CD.