基于线性变换的高光谱图像端元提取算法  

Hyperspectral Image Endmember Extractio Algorithm Based on Linear Transformation

作　　者：王瀛[1] 郭雷[1] 梁楠[1] 

机构地区：[1]西北工业大学自动化学院,陕西西安710129

出　　处：《计算机仿真》2012年第12期329-333,348,共6页Computer Simulation

基　　金：国家自然科学基金项目(60802084)

摘　　要：研究了高光谱遥感图像的端元优化提取问题,针对现有特征空间中最大体积转换思想的端元提取算法中所存在的运算量巨大、对原始数据需要预处理、提取精度较差等问题,分析了图像数据在高维特征空间的相关性,提出了采用线性变换的端元提取算法。使特征空间的基变换寻找正交于某个低一维超平面的投影向量,通过数据在向量上的投影运算将低维相关数据压缩成一个点,与点距离最大的孤立像元作为一个端元输出,每步获得的端元反馈作为下一次提取的输入以保证提取的正确性。由于采用在高维特征空间中距离的计算代替体积计算。仿真结果表明,提出的算法在较短的时间内可有效地提取端元,大大减少了计算量;而且每次提取所依据的信息是反映整幅图像数据在特征空间线性相关性的子空间,所以不需要对原始数据进行预处理,避免了丢失小目标的隐患,进而可以提高提取精度。为高光谱遥感图像优化提取提供了参考。In order to overcome the limitations including high computational complexity, the requirement of preprocessing of original data and poor extraction accuracy in the hyperspectral remote sensing image endmember extraction algorithms which produce endmembers based on maximum volume transform in the feature space, an endmember extraction algorithm was proposed after the correlation of image data in the high dimensional feature space was ana- lyzed. A projection vector orthogonalized a less - one - dimension hyperplane, onto which projection operators compress the low - dimension correlation data into a point, was calculated by using base transformation in feature space, hence, an isolate pixel was determined as an endmember which has the maximum distance from the point, and endmembers extracted in each step were fed back as a part of input of the next step to ensure the extraction accuracy. As a result of the calculation of the distance in the high dimensional feature space instead of volume, the proposed algorithm can extract endmembers in a short time effectively and reduce computation significantly. Each extraction step depending on a subspace reflected linear correlation of the whole image in the feature space leads to that the preprocess of the original image become unnecessary, therefore, the extraction accuracy is enhanced due to avoiding the hazard of loss of small targets. Experiments using simulation image and real image show the effectiveness of the algorithm.

关 键 词：高光谱图像 端元 单形体 线性空间 基变换 子空间 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]