Evaluating the uncertainty of area estimates derived from fuzzy land-cover classification

The use of remotely sensed data as input into geographical information systems has promoted new interest in issues related to the accuracy of multispectral classification. This paper investigates the impact of classification uncertainty on the estimation of area from satellite derived land-cover data. Applying four variants of the maximum-likelihood classifier, it is shown that the estimated area for different land-cover classes is highly influenced by the methods which are used for classifier training. To evaluate the uncertainty of area estimates, a new error modeling strategy is proposed. Assuming that attribute uncertainty in image classification is field-based rather than pixel-based, the image is segmented in fields according to similarities in the probability vectors of adjacent pixels. In simulating uncertainty, this field structure is explicitly taken into account. Using different strategies for image segmentation, it is shown that the spatial correlation of classification uncertainty has a major impact on the assessment of the uncertainty of area estimates.