In the process of rapid scanning or low power excitation, low SNR Raman usually spectra of biological samples can be acquired. In order to remove the noise in the low SNR spectra, we decomposed the spectra by the CEEMD method and separated the noise from spectra according to the Normalization Permutation Entropy in this paper. The method proposed was named as Complementary Ensemble Empirical Mode Decomposition (CEEMD). LCEEMD method can be used to denoise the Raman spectra, which effectively overcame the modal aliasing between high frequency Raman signals and noise components in EMD. Furthermore, CEEMD reduced residual noise, which were presented in EEMD. Simulation experiments showed that LCEEMD method can improve the SNR of data from 10 dB to 39.615 0 db with a standard deviation of 0.001 17 and correlation coefficient 0.999 9. The denoising experiments indicated that the skin Raman spectrum denosied by LCEEMD showed Raman strong characteristic peaks excited by the amide I-belt of cuticle lipid and weak peak of triglycerides (CO), and most peak intensities were consistent with the references. What’s more, the measurement for water-soluble sugar (rice leaf) was modeled with the removal noise data processed by LCEEMD. The prediction coefficient was 0.871 7 and standard error of prediction was 0.912 0, however they were 0.511 4, 1.647 8 and 0.638 2, 1.508 8 in models denosied by EMD and EEMD. In the process of noise removal by LCEEMD, the threshold of the Normalization Permutation Entropy was adjusted according to denoising performance indexes automatically where parameters needn’t to be set and the LCEEMD method is an adaptive noise filtering.