Initialization of the N-FINDR Algorithm Based on the Max-min Distance Method

Because of the complexity of hyper spectral imagery, the analysis of mixed spectral become one of the difficulties and key points in the hyper spectral remote sensing image processing, and end member extraction occupys an important position. Based on the linear model of the automatic, N-FINDR algorithm is easy to be realized by computer. But the result of extraction is very extremely unstable, and it is due to the initial end member matrix. According to the relation between the pure pixel and mixture, this paper uses the max-min distance method in clustering to initialize the end member matrix, this method enlarges the difference of the initial pixel to obtain a stable and large volume. Through the construction of data and examples of data for testing, this paper indicates that this methods is better than the random initialization, it gets a stable and extracted end members that conform to the actual situation.