Abstract: Underwater ambient noise is one of the main factors affecting sonar performance. The selection of noise model is very important. Gaussian noise model is widely used, but in many cases limitations exist. This paper introduces two kinds of non-Gaussian noise models matching to the actual non-stationary ocean ambient noise: Generalized Autoregressive Conditional Heteroscedasticity (GARCH) model and bimodal model. Through the analysis of the two noise model's probability density functions (PDF) and by comparing them with the Gaussian noise model and the measured underwater ambient noise, the applicability of the two models is clarified. GARCH (1,1) can be fit for the majority of shallow and deep ocean ambient noise by adjusting the parameters. The bimodal model fits for some conditions of shallow water, not for the deep water. The statistical characteristics of the two noise models indicate that they can be used in the non-stationary cases where the Gaussian model has limitations.