Distributed Acoustic Sensing (DAS) offers a cost-effective solution for long-term acoustic monitoring in shallow water environments. However, field data from multiple DAS experiments reveal significant noise, particularly at frequencies above 1 Hz, which challenges its use for acoustic studies. Notably, these high-frequency noises (>1 Hz) are strongly correlated with 0.1–0.5 Hz ocean gravity waves, with noise levels increasing at the peaks and troughs of the gravity waves' amplitudes. To address this issue, we developed a curvelet-based machine learning method to remove noise and enhance the signal-to-noise ratio (SNR). Using 2 years of DAS data collected at the Martha's Vineyard Coastal Observatory, we trained and validated the denoising model. The method was applied to signals from wind farm pile-driving and whale calls, resulting in significant SNR improvement for both cases. The denoised results were benchmarked against data from a co-located hydrophone and a hydrophone array deployed at the wind farm site. Comparisons demonstrated strong agreement between the hydrophone and denoised DAS data. This highlights the potential of the denoising approach to uncover signals masked by noise in DAS data and enable the detection of signals previously hidden in noisy data, significantly enhancing the utility of DAS for shallow-water acoustic studies.