An adaptive fuzzy evidential nearest neighbor formulation for classifying remote sensing images

Abstract

The paper presents a novel adaptive fuzzy evidential nearest neighbor formulation for classifying remotely sensed images. The formulation combines the generalized fuzzy version of the Dempster-Shafer evidence theory (DSET) and the K-nearest neighbor (KNN) algorithm. Each of the K nearest neighbors provides evidence on the belongingness of the input pattern to be classified, and it is evaluated based on a measure of disapproval to achieve the adaptive capability during the classification process. The disapproval measure quantifies the lack of support with respect to the belongingness of the input pattern to a given class. Pieces of evidence are ranked based on their degree of disapproval and fused in a sequential manner. The pignistic Shannon entropy is used to estimate the degree of consensus among pieces of evidence provided by nearest neighbors and as a criterion for terminating the evidence fusion process. The paper reports the results of experimental work conducted to evaluate the proposed classification scheme using real multichannel remote sensing images. As will be demonstrated using the experimental results, the proposed classification scheme demonstrated robust performance and outperformed commonly used methods such as the K-nearest neighbor algorithm of Cover and Hart (1967), the fuzzy K-nearest neighbor algorithm of Keller et al. (1985), the evidence-theoretic K-nearest neighbor algorithm of Denoex (1995), and its fuzzy version of Zouhal and Denoex (1997). The performance of these techniques is examined with respect to the K-parameter and classification accuracy.