Analysis of convergent evidence in an evidential reasoning knowledge-based classification

Abstract

The use of knowledge-based systems (KBSs) that use evidential reasoning for land-cover mapping derived from remotely sensed images is spreading widely. In recent years, KBSs utilizing the Dempster–Shafer Theory of Evidence (D-S ToE) have been found most successful in a wide range of remote sensing applications, partly because of their ability to combine diverse information sources. An important feature of the D-S ToE is that it provides a measure for the evidential support (belief) accumulated for each object class at each pixel. Despite the importance of cumulative belief values (CBVs) in representing the weighting of supportive versus conflicting evidence for each class, their analysis has received little attention in the literature. The objective of the present study was to assess the performance (represented by the kappa coefficient) of a KBS based on D-S ToE and of an unsupervised classification (ISODATA), with relation to the CBV distribution determined for each class. This was done for the task of crop recognition in a wide heterogeneous region in Israel. It was found that while KBS performs very well in cases of conflicts and moderate support, the US classification performed well only in cases of homogeneity and uniqueness. Crop recognition by means of KBS was applied to almost one-third of the country's agricultural areas, and it provided a high level of differentiation among seven crop types, orchards and natural vegetation types.