Abstract: Stepwise regression prediction models of levels of annual first low temperature disaster to double cropping rice were established based on meteorological industry standards, the daily meteorological data of 708 weather stations located in the planting regions of double cropping rice in the southern China from 1961 to 2010 and 74 atmospheric circulation characteristics from 1960 to 2010. Methods such as factor puffing, correlation analysis, and stepwise regression were used to establish the prediction models that can discriminate different areas according to risk levels and their spatiotemporal change trends. The average basically consistent accuracy rate of the extended prediction of low temperature damage in highly risk area (Ⅰ area) by the stepwise regression prediction models was
100% for early rice, 83.3% for Japonica rice and 83.3% for Indica rice.  Similarly, as to low risk area with a riskincreasing trend (Ⅱ area), the prediction accuracy rate was 100% for early rice, 83.3% for Japonica rice and 83.3% for Indica rice; as to low risk area with a riskdecreasing trend (Ⅲ area), the prediction accuracy rate was 83.3% for early rice, 100% for Japonica rice and 83.3% for Indica rice. The errors of back substitution and prediction of the models to the representative stations of each region were mainly equal to or less than one level. On the whole, the prediction models established in this study had high accuracy.

Key words: CLUE-S, SWAT, Hun-Taizi River watershed, non-point source pollution