中文摘要：

针对经典多端元光谱混合模型（MESMA）存在着计算量大,端元预选繁琐以及过拟合等缺点,提出了一种改进的多端元解混算法。该算法根据正交子空间投影具有分离感兴趣信号与不感兴趣信号的特点,将像元投影到全部地物端元（每类地物选择一条类内光谱）构成的正交子空间上,按照投影值确定构成混合像元每类地物的类内光谱,在下一步迭代求解的过程中,分离出已确定地物类内光谱的像元,降低计算量,然后根据重构误差变化量确定最优端元个数,避免过拟合。实验结果表明,改进的算法反演丰度误差和解混时间都比原有算法降低很多。

英文摘要：

The classical multi-endmember spectral mixture analysis model has shortcomings in computation intensity,cumbersome endmember preselection and over-fitting. To overcome these shortcomings,an improved multi-endmember unmixing algorithm is proposed here. Using the characteristics of orthogonal subspace projection that can distinguish signals of interest,it projects pixels onto the orthogonal subspace composed of all of endmembers of the entire surface feature class. Each class selects only one intra-class spectrum. Then it determines the intra-class spectrum of every feature class to which pixels belong according to their projection values. These pixels are isolated in the next iteration in order to reduce computation. Then the optimal number of endmember combinations can be determined according to the reconstruction error variation,which avoids over-fitting. Experiment results show that the inversion abundance error and unmixing time of the improved algorithm are reduced compared to the original algorithm.