Local characterization of thematic classification accuracy through spatially constrained confusion matrices

Classification accuracy statements derived from remote sensing are typically global measures. These provide a summary measure of the quality of the entire classification and are typically assumed to apply uniformly over the region represented. Classification accuracy may, however, vary across the region. A simple means of measuring and characterizing accuracy locally, which also facilitates the representation of the spatial variation in classification accuracy, is to constrain geographically the data used for accuracy assessment. The use of this approach is illustrated with a crop classification from Satellite pour l'Observation de la Terre ( SPOT) High Resolution Visible (HRV) data. The global accuracy of the classification was estimated to be 84.0% but accuracy was found to vary locally from 53.33% to 100%. Moreover, accuracy varied from 0-100% over the region on a per-class basis. These variations in accuracy arose mainly as functions of the geographical distribution of the classes and highlight dangers in using a global measure of accuracy that masks spatial variation as a tool in classification evaluation. Local accuracy assessment can, therefore, be a useful analysis and, as the locational information is known, may be achieved at no substantial extra cost to the analysis.