Thematic mapper (TM) based accuracy assessment of a land cover product for Canada derived from SPOT VEGETATION (VGT) data

This paper addresses the accuracy assessment of land cover products derived from coarse-resolution data. The specific product being evaluated covers the landmass of Canada and was derived from the Satellite pour l'observation de la terre 4 SPOT-4 VEGETATION (VGT) data for 1998. A set of representative Landsat frames was identified using a selection algorithm. Recent growing season thematic mapper (TM) (or enhanced thematic mapper plus, ETM+) scenes were digitally classified and precisely registered to the VGT map, and confusion matrixes were produced. The paper addresses methodological issues concerned with geometric and thematic correspondence between the two data sets, the VGT class accuracies, and factors affecting these. It was found that depending on the number of thematic classes (35 to 9) and VGT pixel homogeneity, the agreement between VGT and TM classifications ranged from 20 to 70%. These results are consistent with earlier assessments of similar products using high-resolution land cover maps. Using a TM data set representing similar to8% of the total area, the VGT and TM classifications overestimated the extent of forests by 7.2% (35 classes) and 5.9% (12 classes), respectively. It is shown that the main obstacle to achieving high accuracies of land cover products derived from coarse-resolution satellite data is the heterogeneous land cover at subpixel resolution. The effects of within-pixel land cover heterogeneity, labelling errors, and geographic variations are discussed.