An unmixing algorithm based on vicinal information

Pixel unmixing is a highly undetermined procedure. Due to sensor measurement noise, environmental changes, and intraclass variance, it is very difficult to make a high quality classification. It is even more unlikely to decompose the mixed pixel with a high degree of confidence. Many previous algorithms for linear pixel unmixing were only concerned with a single pixel or a small set of pixels with the same mixing proportions. In this paper we present a novel method for pixel unmixing based on the classification information of vicinal pixels. Since the end members' signatures are non-orthogonal, there are multiple possible combinations of these signatures to produce a particular mixed pixel. Thus, the selection of the end member for a mixed pixel becomes a vital problem. In addition to the widely used residual root mean square error (RMSE), four new performance metrics are proposed for comparing quantitatively the classification accuracy of the proposed method and conventional single pixel unmixing algorithm. While conventional methods aim to minimize residual errors, our method tries to achieve the best possible correct end member combination. The case study shows that in terms of both RMSE and new performance metrics, the proposed method achieves significant improvement over conventional algorithms.