Four field investigations into the Lake Taihu were carried out for collecting in situ observed data in Nov., 2006, Nov., 2008, May and Aug. , 2010. On the basis of water optical classification, different retrieval algorithms were developed, specifically for different types of waters. The obtained optimal models were (1) the four-band model for Type 1 water; (2) the first-order differential model for Type 2 and Type 3 waters. Meanwhile, an optimal retrieval model was also established using the same aggregated calibration data. Some comparisons were done between the developed models for the classified and non-classified waters. The compared results showed that models for the classified waters had better performances than that for the non-classified water, in both retrieval accuracy and model stability. The findings of this study are significant for promoting the development of water color remote sensing for optically complex turbid inland waters.