The turbulent heat fluxes play a pivotal role in the exchange of energy between the atmosphere and ocean. The calculation of these fluxes over the global oceans requires the use of bulk aerodynamic or flux-gradient methods that rely on estimates of the sea surface temperature (SST), near-surface wind speed, air temperature, and specific humidity. Errors in current methodologies of satellite retrievals of near-surface properties have been shown to be the main sources of error for calculation of the fluxes. A new neural network technique is presented here that significantly improves the error characteristics of the air temperature and specific humidity compared to previous methods. Improvements in predicting near-surface wind speed and SST are also seen. Additional improvements are also made by accounting for the effects of high cloud liquid water contents, the effects of which can be mitigated through the use of regime-specific linear and nonlinear retrieval methods. The use of a first-guess SST is shown to result in significant improvement in retrieval accuracy.