基于数据约减和中心化的像元纯净指数端元提取方法

像元纯净指数(pixel purity index, PPI)算法是最为常用的端元提取算法之一,但算法中投影向量的随机性导致多次运行的端元提取结果不一致。为此,提出一种基于数据约减和中心化的像元纯净指数端元提取方法(pixel purity index endmember extraction algorithm based on data reduction and centralization, DRC-PPI)。首先利用自动目标生成算法生成候选端元,并进行无约束最小二乘解混,将解混丰度为负的像元从原始数据中移除得到约减数据。其次,对约减数据进行数据中心化进而获得投影向量,将约减数据投影到这些向量上,然后根据样本点的像元纯净指数选择端元光谱。仿真数据和真实高光谱数据实验结果表明,DRC-PPI算法克服了PPI端元提取结果不一致性,大大减少了投影计算量,其端元提取精度总体上高于PPI算法。

Pixel Purity Index Endmember Extraction Algorithm Based on Data Reduction and Centralization

Pixel purity index is one of widely used endmember extraction algorithm, which identifies endmembers by projecting data sample vectors on a set of projection vectors, however, these vectors must be generated in a random manner which results in the final extracted endmembers are inconsistent and non-reproducible. This study derives a novel endmember extraction algorithm to resolve this issue. It firstly generates endmember candidates and estimates unconstrained abundance using automatic target generate process algorithm and least square estimation method, respectively. And then, it obtains reduction data by removing the pixel with negative abundance from spectral data. The projection vectors can be obtained directly by centralizing and unitizing the reduction data vector. Finally, the reduction data vectors are orthogonally projected onto the projection vectors, and endmember can be identified in terms of pixel purity index of each pixel. Both synthetic and real hyperspectral image experimental results demonstrate that the proposed algorithm could overcome the inconsistency and non-reproducibility caused by random projection vectors, and could reduce projection computation in much degree.