Local Endmember Extraction to Improve Unmixing Results Accuracy
In classic linear mixture modeling, each pixel is classified or decomposed individually without considering its context, using class or endmember distributions derived from training data. This article tries to derive local distributions directly from the image data, using spatial information based on knowledge of the application domain. Contrary to traditional decomposition methods which decompose all image pixels, assuming the same endmember distributions for the entire image, and using all available endmembers at the same time; we try to make use of known spatial relationships between the pixels to locate the mixed pixels that need to be split, to determine local endmember distributions from the image itself, and to select only those endmembers that are probable components of a mixed pixel.
In this method, the smallest possible symmetrical neighborhood - a 3×3 window centered at the mixed pixel - is examined to determine the pure pixel clusters. Furthermore, only a small fraction of all possible combinations are tried in order to save time as well as to prevent unlikely class mixtures from being chosen.
The local endmember extraction method has overcome three important weaknesses of the classic decomposition approach. The first improvement is that, instead of decomposing all pixels, a discrimination between pure and mixed pixels is made in order to handle them by classification and decomposition, respectively. The second enhancement is achieved by taking local endmembers, instead of endmembers that are global to the entire image. The third weakness that is resolved is the necessity of decomposing a mixed pixel into all endmembers simultaneously; our method is able to predict the probable components of a mixed pixel prior to its decomposition.
As it verified, the proposed method is also more accurate than the classic decomposition approach because of the way in which it defines the endmember distributions.