LDA/NWFE calculates the eigenvectors on the basis of the maximum distance among classes and the minimum distance within classes; it extracts and reduces the dimension of samples by using the calculated eigenvectors.
However, the eigenvectors used for dimension reduction fluctuate considerably locally when training samples are insufficient, which can be assumed as a global convolution kernel. Hyperspectral imagery is a continuous spectrum, and a high correlation exists between such imagery and the local adjacent spectrum.
Therefore, the value of the local adjacent spectrum for classification and the weight of the convolution kernels for local similar spectral features should be similar. A convolution kernel with a considerable local fluctuation is an unreasonable convolution kernel.
To address the above-mentioned problems, LIU Xuebin’s research team from XIOPM proposed a novel method. The proposed method (MEFE, MLDA, or MNWFE) uses a continuous spectrum to guide and correct the eigenvectors calculated with LDA/NWFE.
 It is simple and efficient with limited training samples. The proposed method can effectively improve the classical feature extraction method and obtain better classification accuracy. The experimental results show that MEFE performs better than several conventional feature extraction methods, such as LDA and NWFE.

（Original research article " Signal, Image and Video Processing (2019) https://doi.org/10.1007/s11760-019-01604-3 "）