The advantage of neural network method for estimating water quality parameters of complex water body has been approved. Using in-situ measurement data of chlorophyll-a concentration, imageries of MODIS 250m and retrieval model of water temperature, we develop two single-hidden-layer BP neural network models for estimating chlorophyll-a in Lake Taihu: Model NN1 without temperature input and Model NN2 with temperature input. The training method is used by Levenberg-Marquardt algorithm, and the early-stage determinationin the modeling is used to improve generalization. The results show that: the estimation precision of the two models is high, in which the estimation precision of neural network input with temperature has been improved although the test is not significant.