Context and objectives
The aim of this study was to demonstrate the benefit of different tools for the visualisation of classification uncertainty and to stress the need to incorporate error modelling strategies in quantitative remote sensing studies. Especially when remote sensing is used for the monitoring of large regions, as it is the case for the MARS project, accuracy assessment issues become particularly important. Since estimates of the total area of different crops at European scale are based on crop area estimation at the level of 53 sites, it is obvious that, apart from the error introduced by the extrapolation procedure, the accuracy of regional estimates will strongly depend on the reliability of site-specific data. Hence attention should be focused on a) improving the quality of crop discrimination at the level of the site; b) deriving unbiased area estimates for different crop groups; c) providing an indication of the uncertainty of crop area estimates, taking classification error into account.
Expected scientific results
In this study a strategy for uncertainty assessment has been proposed which takes both classification error and the spatial structure of the classified scene into account. It was shown that the uncertainty of area estimation for each class depends on the spatial pattern of the class, the area that is covered by the class and the spectral confusion with other classes. For the majority of classes area uncertainty proved to be relatively low (less than 3%). However, in connection with the estimation of crop area evolution, this level of uncertainty may be considered important. Characterising the uncertainty of crop area evolution for one site can be seen as a first step towards the assessment of the reliability of estimates at European scale.