针对N-FINDR算法计算速度慢、搜索范围较大的特点,提出改进的快速N-FINDR算法,通过提供一个像元个数较少的候选端元集合,为N-FINDR算法提供一个较小的搜索范围。在N-FINDR算法中,所有的端元被认为是处于所有像元构成的单形体的顶点位置,表示这些像元远离像元聚类中心。因此,利用卡方分布的分位点可以分离出这些像元,形成数量较少的候选端元集合。利用合成的和真实的高光谱数据对该算法的性能进行了验证。实验表明,在与N-FINDR算法有相同的端元提取精度的前提下,该算法计算速度更快。

N-FINDR algorithm were employed for endmember extraction for decomposing the mixed pixels, which searches for each pixel from the dimension reduced feature space induced using principal component transformation or maximum noise factor transformation method. Due to the large search range for the endmembers, the efficiency of the N-FINDR algorithm is low. In this paper, we proposed the improved fast N-FINDR algorithm aiming to decrease the computation cost by providing a relative smaller search range, i.e. the candidate endmember set which was only a subset of the entire feature space. N-FINDR algorithm assumed that all the endmembers located at the vertexes of the simplex, which means that these pixels should be far away from the central part of all the pixels. Therefore, the percentile of chi-square distribution can be used to segment out these possible endmembers into a candidate set, which has much smaller size. The performance of the proposed algorithm has been verified using both synthetic and real hyperspectral data. Under the same endmember extraction precision, the modified N-FINDR algorithm has faster computation velocity and a higher overall efficiency.