Design-based mapping of errors in remote sensing-based land use/land cover maps

Abstract

For the first time, spatially explicit representation of classification errors of land use/land cover (LULC) maps is approached from a design-based perspective. Since LULC maps are typically derived from non-probabilistic training samples, these maps, like the true LULC map, are fixed in a design-based scenario so that the error maps achieved by comparing the satellite-based and true maps are fixed. Based on a probabilistic sample of locations where the true or “reference” class is obtained (i.e., the “reference” class is considered the best representation of the true class), errors can be assessed at these sample locations by comparing the map classes to the reference classes. Then, the presence or absence of errors is interpolated across the entire survey area using the nearest neighbour technique. Under very common sampling schemes used to collect reference sample data, the interpolated error maps are design consistent. A simulation study confirms the design consistency of the interpolated error maps, which converge to the true error map as the reference sample size increases. The U.S. land cover map from the LCMAP program and the Italian forest/non forest map serve as case studies.

Bibliographic Information