The turbulent latent and sensible heat fluxes play an important role in the global energy cycle. Modeling of these fluxes over the ocean using flux-gradient or bulk aerodynamic methods relies crucially upon the near surface vertical temperature and humidity gradients. Often, near surface temperature and specific humidity are predicted from satellite measurements using standard multiple linear regression algorithms trained with collocated observations. However, the relationships are not necessarily linear and can be better modeled using a non-linear approach. Neural networks have been used successfully previously to predict monthly averages of temperature and humidity. This study focuses on the use of the neural networks to examine the improvements capable in comparison to the standard multiple linear regression approach. It is shown that using this non-linear approach can greatly improve the prediction of both near surface temperature and humidity.