Taking the cloud top temperatures and their differences from previous an hour temperature, which cause corresponding rainfall over the area of the automatic observatory, as the input layer and the hourly rainfalls as output layer, four kinds of artificial neural network prediction models, which include dot model, line model, little-area model and large-area model are built for June—September tropical cyclone rainfall forecasting in Zhejiang province. After a lot of tests the results show that little-area model could grasp main rainfall factors in infrared satellite image, which include moving direction, moving speed, temperature, developing rate and TBB gradient. It features better ability of forecast and TBB preventing over-fitting.