Estimating per‐pixel thematic uncertainty in remote sensing classifications

Abstract

Standard methodologies for estimating the thematic accuracy of hard classifications, such as those using the confusion matrix, do not provide indications of where thematic errors occur. However, spatial variation in thematic error can be a key variable affecting output errors when operations such as change detection are applied. One method of assessing thematic error on a per‐pixel basis is to use the outputs of a classifier to estimate thematic uncertainty. Previous studies that have used this approach have generally used a single classifier and so comparisons of the relative accuracy of classifiers for predicting per‐pixel thematic uncertainty have not been made. This paper compared three classification methods for predicting thematic uncertainty: the maximum likelihood, the multi‐layer perceptron and the probabilistic neural network. The results of the study are discussed in terms of selecting the most suitable classifier for mapping land cover or predicting thematic uncertainty.

Acknowledgements

A version of this paper was previously published in The Proceedings of 7th International Symposium on Spatial Accuracy Assessment in Natural Resources and Environmental Sciences, M. Caetano and M. Painho (Eds), 5–7 July 2006, Lisbon, Portugal (Lisbon: Instituto Geográfico Português) (Brown et al. 2006). The results presented in that paper are given here and additional results are presented that clarify issues related to per‐pixel thematic uncertainty estimation. This paper is published with the permission of the Instituto Geográfico Português. The Environment Agency and English Nature partially funded this study. The authors thank Rob Wolstenholme and Elly Hill of English Nature for their help and advice during the study. We are grateful to the referees for their constructive comments on the article.