Spectral unmixing for hyperspectral image classification with an adaptive endmember selection

Hyperspectral classification techniques are widely used for detailed analysis of the earth surface. However, mixed pixels caused by the relatively low spatial resolution of the imaging system are the big burden for traditional pure-pixel-hypothesis based hard classification methods. To address this problem, a novel method, which jointly uses soft classification and spectral unmixing, is proposed in this paper. The confusion matrix is exploited to determine the endmember set for each class. Then the generated endmember is adopted for spectral unmixing. The fractional abundance of training samples, which is generated from spectral unmixing, is utilized to optimize soft multinomial logistic regression classifier. The result of the optimized classifier will result in a more accurate confusion matrix. Thus, this procedure is executed iteratively to achieve required performance. Experimental results on synthetic and real hyperspectral data sets demonstrate the superiority of the proposed method for hyperspectral image classification.