Underwater communication is usually affected by ambient noise, which may be generated by different sources, such as the wind origin sea-surface sources, ships and under water life. The properties of background noise, which are non-stationary in nature, depend on location, sea depth, wind speed and sound propagation conditions in the area. Overall performance of underwater acoustic instruments can be improved by denoising the underwater signals. This paper proposes a novel denoising method using empirical mode decomposition (EMD) technique. Frequency domain based thresholding has been used to denoise the signal, which involves three steps: (i) EMD is applied to the noisy signal to decompose the signal into intrinsic mode functions (IMFs). (ii) Thresholding is applied to each IMF in the frequency domain to remove the noise. (iii) Thresholded IMFs are added to obtain the denoised signal. Real-time experiments were performed to validate the proposed technique on the basis of the records of the ambient noise data recorded at sea for various wind speed. It was observed that the experimental results are in good agreement with the proposed algorithm under different wind-noise levels.