Thematic validation of global land cover data sets-procedures and interpretation methods

The International Geosphere-Biosphere Programme DISCover Global Land Cover Data Set (v 1.0) was assembled to provide a high spatial resolution data set for global studies of climate, biogeochemical cycles, atmospheric chemistry, water, energy, vegetation, and ecosystems. DISCover was compiled at a resolution of 1 kilometer using data from the National Oceanographic and Atmospheric Administration (NOAA) Advanced Very High Resolution Radiometer (AVHRR). Following production of DISCover, a study was performed to verify its thematic accuracy and provide statistically valid accuracy estimates on an area-weighted and by-class basis. Landsat Thematic Mapper and SPOT satellite image data were used to verify 379 primary core samples selected from DISCover 1.0 using a stratified random sampling procedure. The goal was to verify a minimum of 25 samples per DISCover class; this was accomplished for 13 of the 15 verified classes. Three regional expert image interpreters independently verified each sample. A majority decision rule was used to determine sample accuracy. For the 15 DISCover classes validated, the average class accuracy was 59.4% with accuracies for the 15 verified DISCover classes ranging between 40.0% and 100%. The area-weighted accuracy of the data set was determined to be 66.9%. Considering only samples for which a majority of interpreters agreed were classified in error (with majority agreement on the correct class) as well as those samples interpreted by consensus as correct, the average class accuracy of the data set was calculated to be 73.5%. This paper describes procedures utilized to validate the thematic accuracy of DISCover. Data set sampling design and core sample interpretation procedures are summarized. Interpretation procedures and interpreter confidence (recorded for each sample by each participant during the validation) is analyzed by category and interpretation region. This DISCover 1.0 validation demonstrates that Landsat TM and SPOT imagery can be efficiently used to validate high-resolution global land cover products. This study demonstrates that the 15 validated IGBP land cover classes were not equally interpretable on the TM and SPOT imagery. The utility of and confidence that may be placed in this technique depends upon the land cover classification scheme used or a subset of categories.