基于SURF和顾及维数的高光谱影像端元提取新算法

摘要

端元提取,丰度反演是高光谱遥感技术的重要内容,其中端元提取是关键的步骤。首次将特征提取算法speed-up ro-bust features( SURF)引入到高光谱影像端元提取中。兼顾高光谱影像丰富的光谱信息改进了SURF算法,提出了在多维尺度空间内寻找极值点作为端元的高光谱影像端元提取新算法,即多维 SURF ( multi-dimensional speed-up robust features, MD-SURF)算法；将其应用于美国EO—1卫星获取的云南中甸普朗地区的Hyperion高光谱影像,并成功提取了影像端元。为了进一步验证结果的可靠性,设计两组对比实验,分别利用N-FINDR和连续最大角凸锥( sequential maximum angle convex cone, SMACC)算法在同等条件下提取实验影像的端元,然后对三种方法的结果进行综合评价和分析,得出MD-SURF算法提取端元的观感效果较好、精度最高、质量最好。提出了一种新的高光谱影像端元提取算法,实验结果表明新方法具有精度高、鲁棒性好等特点,证明了基于新物理机理的MD-SURF算法是一种可行的高光谱端元提取算法。%Endmember extraction and abundance inversion are important study contents of hyperspectral remote sensing technology therein the endmember extraction technology is a sky step on its applications. The feature extrac-tion algorithm speed-up Robust features ( SURF) is introduced into the endmember extraction of hyperspectral im-age for the first time. A new algorithm named the multi-dimensional speed-up Robust features ( MD-SURF) was put forward for hyperspectral image endmember extraction based on the feature detector of SURF and considering the hyperspectral dimension. MD-SURF is to find the extreme points as endmembers in multi-dimensional scale space. MD-SURF was used to extract successfully the endmembers of the Hyperion hyperspectral image obtained by Ameri-ca EO-1 satellite in Pulang Region of Yunnan province in China. In order to verify furtherly the reliability of the re-sults, the most commonly used algorithms of N-FINDR and sequential maximum angle convex cone (SMACC) were used simultaneously to extract endmembers of the experimental image in the same conditions, and the MD-SURF was proved to have the highest precision and the best quality according to the comprehensive evaluation and compar-ative analysis results of the MD-SURF, N-FINDR and SMACC applications. A novel endmember extraction algo-rithm of hyperspectral image was presented, and the result indicates that the new algorithm has the characters with high precision, good robustness and so on. Therefore, the MD-SURF is a feasible algorithm for hyperspectral end-member extraction.