Variable selection plays a significant role in multivariate calibration for spectral analysis.However,majority of variable selection methods are linear algorithms,which often focus on one aspect of the wavelengths and ignore the impact of the sample sets as well as non-linearity,thus resulting in a poor result.In this regard,a new novel non-linear variable selection strategy,ensemble mean impact value(EMIV),is proposed by taking into account the impact of both the wavelength and the sample set in a non-linear way,which is expected to provide a more robust calibration performance.The EMIV selected variables were then used as the input of support vector machines(SVM)to process the complex spectral data sets at hand,no matter what kind of data structures will present.We test and refine the EMIV with several data sets,and the results reveal that the proposed method handles all complex spectra very well and produce a good calibration performance,thus providing a promising tool for complex spectral analysis.This would definitely promote the application of spectroscopy in process analysis and control.