A case of coastal sea fog over the Yellow Sea in April 2014 is selected to carry out ensemble forecast experiment using the WRF (Weather Research and Forecasting) model. Based on RMSE (root mean square error) of SLP (sea-level pressure), 2-m temperature, 2-m water vapor mixing ratio, and 2-m relative humidity (RH) in the initial field of each ensemble member, as well as RMSE of ensemble mean, 4 different schemes of ensemble optimization are designed by eliminating/retaining the members whose RMSE values are higher/lower than the RMSE of ensemble mean. A series of numerical forecast experiments are conducted to compare the effects of ensemble forecast with different optimization schemes. The results are as follows. (1) In the ensemble generated by Monte Carlo method, there are many members with poor performance in forecasting sea fog, degrading the skill of ensemble forecast. Therefore, initial field optimization is necessary in sea fog ensemble forecast. (2) The RH-RMSE scheme, which takes RH as the optimization reference variable, performs better than the other three schemes. (3) Compared with the non-optimized ensemble forecast, the optimized ensemble forecast using RH-RMSE scheme not only saves about 50% of the calculation time, but also improves ETS (equitable threat score) by 36%. The RH-RMSE scheme proposed in this study has the prospect of operational application.