A simple measure of confidence for fuzzy land-cover classification from remote-sensing data

Abstract

Representing the quality of thematic maps derived from remote-sensing image classification is important in assessing its fitness for use. Conventional approaches to represent the quality in terms of accuracy need information from the reference data at the same scale. Error-prone or dubious reference data may have an impact on the assessment of quality. Therefore, measures that complement the conventional accuracy measures are required to represent the quality. Uncertainty and confidence are such measures that do not require reference data. Few studies have been attempted to derive pixel-level confidence. However, these measures are not widely adopted by the remote-sensing community due to their limitations. In this article, a simple measure of confidence is derived to represent the quality of fuzzy classification. To derive the confidence value for a pixel, two values, viz. first highest class membership value as evidence and an associated degree of certainty, are required. When the difference between first and second highest membership values is used as degree of certainty in the proposed approach, the confidence measure derived is equal to the complement of existing measure of uncertainty, viz. confusion index in difference form.