Total suspended matter concentration is one of the important parameters of water component. Traditional retrieval model ignored the difference of case 2 water which has complex optical properties. In the present study, we developed a method of water classification based on optical classification using HJ-1 multispectral data. We divided the water into two types and developed a retrieval model suitable for different water type. The results indicate: (1) the accuracy of retrieved results based on spectral classification has been improved; (2) exponential model reflects the relationship between suspended sediment concentration and retrieved factor and is better for type Ⅰ water, while linear model is better for type Ⅱ water.