Edward Cripps, Anthony O'Hagan, Tristan Quaife and Clive Anderson
Australian Institute of Marine Science, Townsville, Australia, Department of Probability and Statistics, University of Sheffield, UK and Department of Geography, University College London, UK.
Publication details: Submitted to Bayesian Analysis, 2008.
The purpose of this paper is to quantify uncertainty associated with land cover maps derived from satellite data. Satellites record energy reflected from the earth's surface and this information may be used to derive maps of land cover. Typically these maps are thematically organised in terms of the land surface vegetation. The final product is reported as recordings of distinct vegetation classes at pixels across a spatial grid. The recordings are possibly misclassified and the accuracy of these land cover maps is often assessed by a confusion matrix derived from a comparison of the map to a field survey at a sample of pixels across the region. The data in the confusion matrix is small relative to the entire map and does not provide spatial information regarding map accuracy. We propose a Bayesian method to address these two issues. The model describes multinomial recordings with misclassification probabilities and incorporates a spatial correlation structure that is suited to the case where little spatial information exists. Our method allows us to estimate the posterior distributions of the land cover proportions for individual sites as well for the entire region, features previously unavailable in accuracy assessment techniques. We present the results of our method applied to a recently developed satellite derived land cover map, the Land Cover Map 2000, for the region of England and Wales.
Keywords: multinomial distribution, misclassification probabilities, spatial correlation, remote sensing, land cover maps, confusion matrix.