An assessment of reference data variability using a "Virtual Field Reference Database"

A "Virtual Field Reference Database (VFRDB)" was developed using field measurement and digital imagery (camera) data collected at 999 sites in the Neuse River Basin, North Carolina. The VFRDB was designed to support detailed assessments of remote-sensor-derived land-cover/land-use (LCLU) products by providing a robust database characterizing representative cover types throughout the study area. The sampling frame incorporated both systematic unaligned and stratified random design elements, to provide both an even distribution of points and sufficient intensification to account for rare classes. Numerous quality assurance procedures were developed and incorporated to ensure both data consistency and repeatability. Two independent interpreters assigned class labels corresponding to a hierarchical classification system based on field measurement and imagery data interpretation. Correspondence between interpreters was analyzed at multiple classification levels. The relatively high 91 percent overall correspondence of interpretations was attributable to the application of the VFRDB, providing a high quality source of measurement and imagery data to guide class assignments. Confusion documented for rangeland and forest classes was consistent with reported results for studies conducted in diverse biological locations. Results demonstrate the requirement for reference data with known variability, to support the quantitative assessments of remote-sensor-derived LCLU products.