基于鲁棒最大单纯形体积的高光谱图像快速端元提取

Robust maximum simplex volume-based method for fast endmember extraction of hyperspectral data

针对基于最大单纯形体积方法的端元提取算法通常涉及到对全局像素的处理且对噪声敏感的问题,提出了能快速从高光谱图像中提取端元的鲁棒最大单纯形体积端元提取算法。首先,该算法采用主成分分析对高光谱图像降维至p-1维子空间。随后,对于降维的子空间,算法采用凸包检测算法获取子空间下的凸包边界点。接着,对于检测到的凸包边界点,算法迭代选取p个数据点并计算其行列式体积,直至选取出p个能产生最大单纯形体积的数据点。最后将提取的p个数据点逆变换至原始维度空间从而获取去噪后的p个端元。在模拟数据集和真实数据集上的实验结果表明,提出的算法能快速提取弱噪声的端元。该算法能满足高光谱端元提取领域中的高精度,实时性强的要求。

In order to solve the problems that maximum simplex volume-based endmember extraction algorithms(EEAs) involve processing entire pixels and are sensitive to noises, this paper proposes a robust maximum simplex volume-based EEA for quickly selecting endmembers from hyperspectral images. The proposed algorithm first applies principal component analysis to reduce the hyperspectral image into p-1 subspace. It then detects convex hull points from each component pair by employing a convex hull algorithm. Next, it iteratively specifies p points and their simplex volume until they can provide a maximum simplex volume. Finally, it transforms p points into original dimensionality and obtains p denoised endmembers. Experiments conducted on synthetic and real hyperspectral images demonstrate that the proposed algorithm can quickly extract endmembers from the denoised hyperspectral. The proposed method can perfectly meet the requirements of high endmember accuracy and real-time in the field of hyperspectral endmember extraction.