The collected field seismic data is disturbed by random noise, which is not conducive to seismic data interpretation, the noise must be removed from seismic data. Soft and hard threshold methods of wavelet transform can eliminate this noise of seismic data, but they have some shortcomings in signal denoising. So this paper proposes an improved threshold method to seismic data denoising. The improved threshold method can eliminate random noise effectively, which is evaluated by means of root mean square error (RMSE) and signal-to-noise ratio (SNR). Furthermore, according to the decomposition layers of wavelet transform, the selection of reasonable threshold value is obtained, and the simulation results proved that the improved method can well remove noise and the wavelet basis function of wavelet transform is sym3 and the best decomposition level of the new method is 3 layers. Using the improved threshold method to remove noise in actual seismic data, the results show that lots of noise is removed, and the resolution of seismic data profile after noise reduction is high and the texture of it is clear, especially, the effect of the improved threshold method in seismic data denoising is better than that of soft and hard threshold methods.