In the process of signal denoising, because of the obvious “flying away” phenomenon of the endpoint in the calculation of IMF (Intrinsic Mode Function) by two-dimensional EMD (Empirical Mode Decomposition), the fluctuation amplitude of the signal after denoising is still large. Therefore, a reflective terahertz spectral signal denoising method based on improved EMD is proposed. Based on the compound Simpson formula, a multivariate grey model is constructed, and it is brought into the calculation process of two-dimensional EMD to achieve the suppression and improvement of the “flying away” at the endpoint; In the signal denoising stage, after determining the correlation of the original signal noise IMF component, the improved two-dimensional EMD is used to filter the noise IMF component. In the test results, the fluctuation amplitude of the reflected terahertz spectral signal after denoising by the design method can reach a good convergence state.