Noise reduction method of lidar atmospheric backscattering signal based on empirical mode decomposition (EMD) is developed by Cross-Validation. Considering characteristics of lidar return signal noise and defects of traditional de-noising algorithm, Cross-Validation is applied to identify signal layers and noise layers automatically, and then separate signal and noise by EMD reconstruction. With experiments, the method can select the signal in the instrinsic mode function adaptively, not only removes the random error, but also maintains the effective characteristics of the signal, reduces the loss of signal, and then improve the accuracy in the next phase of data processing.