基于流形学习和空间信息的改进N-FINDR端元提取算法

光谱端元提取是对高光谱数据进一步分析的重要前提。由于双向反射分布函数(BRDF)，像元内的多重散射和亚像元成分的异质性等因素，高光谱图像中的混合像元实际上是非线性光谱混合。传统的端元提取算法是以线性光谱混合模型为基础，因此提取的端元精度不高。在光谱非线性混合的基础上，提出一种将流形学习与空间信息结合的改进N-FINDR端元提取算法。首先通过自适应的局部切空间排列算法寻找嵌入在高维非线性数据空间的本质的低维结构，将原始高光谱数据非线性降维到低维空间。接着利用地物分布具有连续性的特点，通过增大空间同质区域的像元的权重进行空间预处理。最后通过寻找最大单形体体积进行端元提取。提出算法很好的解决了高光谱遥感数据非线性结构，并利用了空间信息，提高了端元提取的精度。模拟数据实验和真实高光谱遥感数据实验结果均表明，采用该算法得到的结果优于顶点成分分析(VCA) 算法、基于测地线距离的最大单形体体积(GSVM)算法和空间预处理的N-FINDR(SPPNFINDR)算法。

An Improved N-FINDR Endmember Extraction Algorithm Based on Manifold Learning and Spatial Information

An improved N-FINDR endmember extraction algorithm by combining manifold learning and spatial information is presented under nonlinear mixing assumptions. Firstly, adaptive local tangent space alignment is adapted to seek potential intrinsic low-dimensional structures of hyperspectral high-diemensional data and reduce original data into a low-dimensional space. Secondly, spatial preprocessing is used by enhancing each pixel vector in spatially homogeneous areas, according to the continuity of spatial distribution of the materials. Finally, endmembers are extracted by looking for the largest simplex volume. The proposed method can increase the precision of endmember extraction by solving the nonlinearity of hyperspectral data and taking advantage of spatial information. Experimental results on simulated and real hyperspectral data demonstrate that the proposed approach outperformed the geodesic simplex volume maximization (GSVM), vertex component analysis (VCA) and spatial preprocessing N-FINDR method (SPPNFINDR).