Modeling grain-size dependent bias in estimating forest area: A regional application.

Abstract A better understanding of scaling-up effects on estimating important landscape characteristics (e.g. forest percentage) is critical for improving ecological applications over large areas. This study illustrated effects of changing grain sizes on regional forest estimates in Minnesota, Wisconsin, and Michigan of the USA using 30-m land-cover maps (1992 and 2001) produced by the National Land Cover Datasets. The maps were aggregated to two broad cover types (forest vs. non-forest) and scaled up to 1-km and 10-km resolutions. Empirical models were established from county-level observations using regression analysis to estimate scaling effects on area estimation. Forest percentages observed at 30-m and 1-km land-cover maps were highly correlated. This intrinsic relationship was tested spatially, temporally, and was shown to be invariant. Our models provide a practical way to calibrate forest percentages observed from coarse-resolution land-cover data. The models predicted mean scaling effects of 7.0 and 12.0% (in absolute value with standard deviations of 2.2 and 5.3%) on regional forest cover estimation (ranging from 2.3 and 2.5% to 11.1 and 23.7% at the county level) with standard errors of model estimation 3.1 and 7.1% between 30 m and 1 km, and 30 m and 10 km, respectively, within a 95% confidence interval. Our models improved accuracy of forest cover estimates (in terms of percent) by 63% (at 1-km resolution) and 57% (at 10-km resolution) at the county level relative to those without model adjustment and by 87 and 84% at the regional level in 2001. The model improved 1992 and 2001 regional forest estimation in terms of area for 1-km maps by 15,141 and 7,412 km2 (after area weighting of all counties) respectively, compared to the corresponding estimates without calibration using 30 m-based regional forest areas as reference