ISBN: 1-84626-171-6, ISBN13: 978-1-84626-171-8 A General Method for Assessing the Uncertainty in Classified Remotely Sensed Data at Pixel Scale Yanchen Bo 1, 2 + and Jinfeng Wang 3 1. State Key Laboratory of Remote Sensing Science, Jointly Sponsored by Beijing Normal University and Institute of Remote Sensing Applications, CAS, 2. Research Center for Remote Sensing and GIS, School of Geography, Beijing Key Laboratory for Remote Sensing of Environment and Digital Cities, Beijing Normal University, Beijing, 100875, China 3. LREIS, Institute of Geographical Science and Natural Resources Research Chinese Academy of Sciences, Beijing, 100101, China Abstract. The uncertainty assessment on the classification of remotely sensed data is a critical problem in both academic arena and applications. The conventional solution to this problem is based on the error matrix (i.e. confusion matrix) and the kappa statistics derived from the error matrix. However, no spatial distribution information of classification uncertainty can be presented by this method. A probability vector-based method has been developed for assessing classification at pixel-scale. However, the use of this method is severely limited because the probability vector can be derived only through the Bayesian classification. In practice, other classifiers such as Artificial Neural Network Classifier, Minimum Distance Classifier, Mahalanobis Distance Classifier and Fuzzy Classifier are wildly used for remote sensing data classification. To assess the uncertainty of thematic maps by these classifiers at pixel-scale, a general method is presented in this paper which extends the probability vector-based method to the assessment on the uncertainty classified by classifiers beyond Bayesian classifier. This extension is realized through a transformation method, which transforms the “Membership Vector” in various classifiers to the “transformed probability vector” so that it is comparable to the probability vector in Bayesian Classifier. The uncertainty measurements could be derived from the probability vector were evaluated and, the probability residual and entropy that derived from the extended probability vector are used as indicators to assess the absolute and relative uncertainty perceptively. The uncertainties by different classifiers are compared at pixel scale. Some examples of the uncertainty of maps from distance classifiers were presented and compared with that of maps from MLC classifier. Keywords: scale, a posteriori probability vector, uncertainty measure. 1. Introduction Categorical spatial data set such as land use/land cover and vegetation type derived from remotely sensed data is the critical inputs in ecological and environmental modeling. As the classified maps from remotely sensed data are uncertain in nature, assessment on the accuracy and uncertainty of remotely sensed data classification is essential to the propagation and assessment on the uncertainty of models’ output and, further more, the risks of decisions based on the models outputs. An essential aspect of the increasing sophistication of ecological and environmental models is the use of spatially explicit inputs and outputs. Thus, the spatial distribution of the classified maps from remotely sensed data has become a new challenge. The widely accepted method for assessing the accuracy of thematic maps from remotely sensed data has been the error matrix, or confusion matrix (Congalton and Green, 1999). A limitation of the error matrix is that it allows only one reference class for each reference site. This is problematic in sites where selection of + Corresponding author. Tel.: +86-10-58802062; fax: +86-10-58805274 E-mail address: boyc@bnu.edu.cn Proceedings of the 8th International Symposium on Spatial Accuracy Assessment in Natural Resources and Environmental Sciences Shanghai, P. R. China, June 25-27, 2008, pp. 186-194