Assessing the Accuracy of National Land Cover Dataset Area Estimates at Multiple Spatial Extents

Abstract

Site-specific accuracy assessments evaluate fine-scale accuracy of land-use/land-cover (LULC) datasets but provide little insight into accuracy of area estimates of LULC classes derived from sampling units of varying size. Additionally, accuracy of landscape structure metrics calculated from area estimates cannot be determined solely from site-specific assessments. We used LULC data from Rhode Island and Massachusetts as reference to determine the accuracy of area measurements from the National Land Cover Dataset (NLCD) within spatial units ranging from 0.1 to 200 km2. When regressed on reference area, NLCD area of developed land, agriculture, forest, and water had positive linear relationships with high r2, suggesting acceptable accuracy. However, many of these classes also displayed mean differences (NLCD REFERENCE), and linear relationships between the NLCD and reference were not one-to-one (i.e., low r2, β0 ≠ 0, β1 ≠ 1), suggesting mapped area is different from true area. Rangeland, wetland, and barren were consistently, poorly classified.