Forest characterization using multi-resolution satellite data and in situ measurements of Chequamegon National Forest in Wisconsin

Many land cover maps have been created, especially involving forested landscapes and because of this a better understanding of the accuracy of land cover maps is required. Multi-resolution crown cover data were classified for part of the Chequamegon National Forest in northern Wisconsin. A very high spatial resolution likelihood map for crown/no crown cover was created using a QuickBird image of the area, training sites, and a classification decision tree. A threshold of 50% was applied to the likelihood map and then reprojected in to a sinusoidal projection so that it could be averaged and then scaled up to MODIS scale, for comparison with four different iterations of the global MODIS derived VCF product. The four MODIS VCF iterations were taken from 2000 and 2005 data, and were compared to the in-situ field measurements and the QuickBird data at the MODIS scale. The data were compared firstly, for the whole QuickBird image area (all 293 MODIS pixels), and then compared for only the 18 field site pixels. The 2005 MODIS VCF bottom 5 product demonstrated the highest correlation with the QuickBird data and in-situ field measurements. The QuickBird data were rescaled using the field data and then compared again to the field measurements and the MODIS VCF products. These comparisons were used to analyze and evaluate if QuickBird data can be directly used as validation for the MODIS VCF products, or if it needs to be refined first. This research has found that the unscaled QuickBird data overestimates the accuracy of the crown cover. When the QuickBird data are rescaled, however, they can then be used as direct validation for MODIS derived VCF products. The in-situ field measurements were also compared to the QuickBird data separately, which resulted in a high correlation between the two data sets at MODIS scale. There was disagreement, however, in the classified no crown cover areas at the field measurement pixel level between these two data sets.