Extracting essential features from massive bands is an important yet challenging issue in hyperspectral image (HSI) classification. Plenty of feature extraction techniques can be found in the literature but most of these methods rely on linear/stationary assumptions. This paper proposes an alternative methodology inspired by the ensemble empirical mode decomposition (EEMD) to gain spectral features of the HSI. To this end, two major aspects are involved: 1) the optimization problems are formulated in each sifting process and solved by the alternating direction method of multipliers (ADMM) algorithm to enhance the benefits of EEMD; 2) the intrinsic mode functions (IMFs) extracted by the optimized EEMD (OEEMD) are summed with appropriate weights automatically gained from the local Fisher discriminant analysis (LFDA). As a consequence, the constructed features (i.e., sum of the IMFs) can then be significantly classified by the state-of-the-art classifiers, i.e., -nearest neighbor (-NN) or support vector machine (SVM). Experiments on two benchmark HSIs validate that the extracted new features achieve higher classification rates as well as greater robustness to the choice of training samples compared with several generally acknowledged methods.