Spectral mixture analysis (SMA) is one of the most important methods in remote sensing image processing. Traditional SMA assumes a constant spectral signature for each endmember. However, the endmember spectral variability commonly exists, which leads to the lower accuracy of pixel unmixing. In order to solve this problem, a novel unmixing method based on Fisher discriminant analysis (FDA) was developed. FDA aimed to find a linear combination of the spectral bands for getting the largest separation degree among the endmember spectra, i.e. small variability of spectra inside one endmember group but a large difference of spectra among endmember groups. Mixture pixel was unmixed by using transformed spectra, as a result, the adverse impact caused by endmember spectral variability on unmixng accuracy could be diminished to a large extent. A laboratory experiment was designed to obtain a group of mixed spectra with endmember spectral variability. The measured spectra were used to test the performance of the new method and the traditional SMA methods. The comparison results suggest that the new method outperforms the traditional methods with considerable improvement of unmixing accuracy.