Atmospheric correction of remote sensing image is the precondition of inversing water color parameters. An atmospheric correction method for lake Taihu based on 6S the radiative transfer model and the NIR water-leaving reflectance model is proposed. Two neural networks are built to model radiative transfer, and thereafter, estimate parameters such as aerosol optical thickness in 550 nm with the optimization of four near infrared bands (i.e, 754, 779, 865 and 885 nm) of medium-resolution imaging spectrometer (MERIS) data. Then, the parameters are extrapolated to visible wavelength to fetch the parameters in those visible bands. This model by MERIS Level 1p data on November 11th, 20th, 21th, 2007 and November 20th, 2008 and the in situ date-sets is validated. The result indicates that the radiative transfer optimized model has a high precision in inversing the water remote sensing reflectance with an average relative error from 20% to 40% in the 13 bands. Comparing with the atmospheric correction methods of 6S and Beam 4.9, it has the highest correction precision and strongest stability, and is applicable in lake Taihu.