ABSTRACT

Accompanying the dramatic increase in availability of high spatial resolution imagery, linear spectral unmixing is rapidly gaining popularity in the remote sensing community for its ability to extract fine-scale urban features. The vegetation-high albedo-low albedo spectral unmixing model has recently been proposed and used to characterize urban surface features with higher accuracy. In a two-dimensional spectral mixing space, the performance of these linear spectral unmixing models is largely dependent on the triangle structure (i.e., straight or convex edges) of the selected scatterplot. Yet the choice of an appropriate triangle structure is still largely left up to visual interpretation. To address this gap, we propose an indicator to quantify the triangle structures of the scatterplots in all two-dimensional spectral mixing spaces, which would enable the selection of the best feature space for subsequent linear spectral unmixing. The effectiveness of the proposed indicator is evaluated using 8-band WorldView-2 imagery. The unmixing accuracy is highly related to the proposed indicator and this simple yet robust indicator is able to objectively select suitable feature space when performing spectral unmixing and holds great potential for automated high-resolution remote sensing processing.