부산대학교 도서관

Reference data (“ground truth”) maps are commonly used to quantitatively assess the performance of imaging spectrometer classification algorithms. However, standard reference data scenes typically are not sufficiently detailed to support assessment of spectral unmixing algorithms. Furthermore, commonly used reference data often lack validation reports that estimate error in the reference data itself, and new reference data are prohibitively expensive to generate using traditional methods. This paper presents a summary of our work, which is focused on introducing new methodologies to efficiently generate and validate abundance map reference data (AMRD), which can then be applied to assess the performance of spectral unmixing on real remotely sensed imagery. AMRD, generated using our methodology, had a validated mean and standard deviation error of 3.0% and 6.3%, respectively, which rivaled the accuracy the best traditional methods. A separate experiment designed to replicate our methodology, using different scenes and imagery, confirmed the relative accuracy of our techniques.