Procedure for classification of forests for CORINE land cover in Sweden

The EU-CORINE Land Cover (CLC 2000) project for Sweden was carried out by the Swedish National Land Survey. The classification of forested land cover types, covering about 65% percent of the land area in Sweden, was contracted out to the Swedish University of Agricultural Sciences (SLU). The method of forest classification used detailed plot information from the National Forest Inventory (NFI), map masks and Landsat data, and was developed to be automated and robust. Within each Landsat scene an average of 2,000 NFI plots were available to use as field data. One use of the NFI plots was to use them as spectral reference targets for correction of within-scene haze differences and correction of slope effects. The forest classification was made with a self-calibrating maximum likelihood algorithm, using the NFI plots as ground truth. In the classification, prior probabilities were iterated until the frequency of each class in the classification corresponded to the frequency of that class according to the field plots. All pre-processing of the image data and the NFI data, in addition to the classification, was done using the automated production line written in-house called MUNIN. In total three products were produced: the CLC 2000 product with 4 forest classes and 25 ha minimum mapping unit (MMU); and two national products with 7 forest classes and a MMU of 1 ha and 25 m (pixel level), respectively. The seven forest classes in the national product were: clear felled areas; young forests; deciduous forests; mixed forests; coniferous forests 5 to 15 m in height; coniferous forests greater than 15 m in height; and, lichen dominated coniferous forests. Accuracy assessment was carried out using both photo interpretation and cross-validation. The accuracies using cross-validation at the pixel level were on the order of 50-70 %. When aggregated to the four CORINE forest classes, the accuracy increased to about 80 % at the pixel level.