以光谱信息熵改进的N-FINDR高光谱端元提取算法

端元提取是高光谱混合像元分解的关键步骤,也是高光谱影像分析的重要前提。N-FINDR算法是一种经典且有效的端元提取算法,但其需遍历所有可能的像元组合,计算量巨大,时间效率不高。本文以光谱信息熵和凸面几何学理论,利用高光谱影像像元,在光谱特征空间形成的单形体顶点附近为相对纯净像元,单形体内部为混合像元的特性,提出了一种结合光谱信息熵的N-FINDR改进算法。该方法根据各波段像元灰度概率计算影像中每个像元的光谱信息熵,将大于光谱信息熵阈值的像元作为混合像元被剔除,在保留的像元组成的单形体上搜索最大体积,并提取最大体积顶点处像元作为端元。最后,使用美国EO-1卫星获取的江西省德兴某铜矿的Hyperion数据,对改进后的算法进行验证。结果表明,改进后的N-FINDR算法在确保较高端元提取精度的同时,大大提高了数据处理的时间效率。

Improved N-FINDR Algorithm on Hyperspectral Endmember Extraction Based on Spectral Shannon Entropy

Endmember extraction is a key step for unmixing hyperspectral mixed pixels and an important prerequisite in the further analysis of hyperspectral imagery. The traditional N-FINDR algorithm is a classical and effective algorithm among various endmember extraction methods. However, the N-FINDR algorithm need to compute all the possible pixel combinations, thus it is time consuming. In order to improve the time efficiency of the N-FINDR algorithm, this paper proposed an improved N-FINDR algorithm on hyperspectral endmember extraction based on spectral Shannon entropy theory and the convex geometry, meanwhile we utilize the characteristic that in spectral feature space, all pixels of the hyperspectral imagery could compose a single shape body, in which the pure pixels are located at the apex and the mixed pixels in the interior or at the surface. The spectral Shannon entropies of all pixels are calculated according to the pixel gray probability, and are used to determine the purity of pixels. The pixel is removed if its spectral Shannon entropy is greater than the threshold value of the spectral Shannon entropy, otherwise it is preserved. Next, the N-FINDR was used to search the largest volume from the single shape body composed by the preserved pixels, and the pixels at the apex of the body with the largest volume would be the endmembers. Finally, we use the Hyperion data of a copper mine in Dexing city from Jiangxi province to testify the improved N-FINDR algorithm. By analyzing the experimental results, the improved algorithm ensured a high accuracy as well as improved the data processing efficiency very greatly in the course of extracting hyperspectral endmembers.