MODELING ERRORS IN REMOTE-SENSING IMAGE CLASSIFICATION

Standard error assessment techniques in image classification have been primarily concerned with identifying errors in individual pixel assig nments. However, these techniques overlook a fundamental fact that ima ge classification is basically a process of generalization. The output s of this process are often intended to be cartographic objects (e.g., polygons) which are abstract models of reality and may not be verifia ble at each pixel. Linking errors with cartographic objects in image c lassification is a challenging problem in remote sensing. This article proposes a new error assessment methodology for image classification (an error model) in which uncertainties involved in the classification process are estimated through simulations of various steps in image c lassification. Two error models have been developed to estimate the un certainties involved in class modeling (training) and boundary generat ion (boundary pixel allocation). Results derived from two case studies show the validity of the proposed error concept for image classificat ion and its potential for improving image classification.