N. A. Best, J. W. Elliott, I. T. Foster, "Synthesis of a Complete Land Use/Land Cover Data Set for Conterminous United States," Preprint ANL/MCS-P1998-0112, January 2012. [pdf]
The PEEL0 land cover data set for the conterminous USA characterizes each of its five arcminute 50 cells in terms of sub pixel area fractions for fifteen land use and natural cover classes, with the fractions for each cell summing to unity. This dataset has three advantages relative to other products. First, its cover classes address distinctions important in studies of the economic dimensions of land use and land cover change, by for example distinguishing between land uses associated with human and natural processes and among different crops while simultaneously providing a complete representation of cover in each cell. Second, aggregates for the various cover classes in PEELo compare more favorably with national level statistics from the USDA Major Land Uses MLU census data for 2002 than do other sources. Aggregate cultivated land is within 0.8 percent of the MLU value, as compared to 16.2 percent for the Modis Land Cover Type MLCT primary cover product, 1.8 percent for the National Land-Cover Database (NLCD), and 1.2 percent for the Agricultural Lands in the Year 2000 dataset. Aggregate water, natural, and urban cover classes are within 2.2, 1.0, and 6.1 percent resp. as compared to deviations of 24.0, 0.01, and 69.2 percent for MLCT primary and 2.2, 1.35, and 6.1 percent for NLCD. Third, the spatial distribution of cultivated land is also substantially improved; PEELo per cell sub pixel fraction root mean square error for cropland relative to the NLCD is 0.149 versus 0.175 for the 2001 MLCT primary classification, a 16 percent improvement in RMSE. PEELo, improved performance is due to the multi source guided aggregation decomposition method used to construct this dataset. This method combines information from multiple sources, including spatial data sets and agricultural production statistics, to guide both aggregation of multiple fine-scale land use land cover classifications, and decomposition of hybrid classes to separate their constitutive land uses and natural covers. We describe how we use this method to construct PEELo by incorporating information from the 2001 MODIS Land Cover Type data set, the 2001 NLCD, and the Agricultural Lands in the Year 2000 data set. In the case of the MODIS data set we demonstrate that considering all of its information components (primary classification, secondary classification, and confidence level) leads to more accurate aggregate measures of cultivated acreage. We also describe our evaluation process. This procedure is adaptable to other regions, data sets, and requirements.