Abstract: To analyze the impacts of air temperature on annual runoff in the headstream of Yangtze River and explain the physical mechanism behind the glacier melt from Qinghai-Tibetan Plateau, the Generalized Additive Model of Location, Scale and Shape (GAMLSS) is employed to detect the non-stationarity of the annual runoff series (1960-2012) and to quantify the relationships of both the mean and variance of the annual runoff variable to physical factors such as precipitation and air temperature. In the GAMLSS model, the impact of air temperature can be represented in two different ways, one is to directly use the annual average temperature as a predictor and the other way is to use the index of accumulated temperature deficit (ATD) as a predictor, so two kinds of regression models are established in this paper under four different distribution assumptions for annual runoff variable. By comparing the efficiencies of the two kinds of GAMLSS models with different combinations of predictors using a variety of criteria such as the generalized Akaike information criterion, both normal QQ and worm plots of the residuals, Filliben correlation coefficient and kernel density estimation, we investigated the impacts of air temperature on annual runoff in the upper Yangtze River. We found that annual runoff series from the Batang station generally had an unstable trend, with a slight decline during the period 1960-1988 followed by a rise for over twenty years later, while annual precipitation and air temperature values showed continuous, steady and slow increasing trends. The fact that there existed a decline in annual runoff for nearly 30 years can be partly explained by the phenomenon called refreezing of meltwater of glaciers. The fitted model with ATD as a predictor under the lognormal distribution assumption for annual runoff performed better than the direct use of air temperature as the predictor in estimating annual runoff values. The ATD series, as an indirect reflection of heat energy deficit, showed roughly the same change trend as annual runoff series. All the above findings indicate that annual runoff series of the Batang station over the period 1960-2012 might be more strongly influenced by the proposed ATD index rather than directly by the air temperature, suggesting that the ATD index could more effectively explain the physical mechanism of glacial runoff generation that is a significant component of the total runoff in cold regions and hence better describe the temporal characteristics of annual runoff in the upper Yangtze river. In summary, our study may be helpful in predicting annual runoff and understanding the mechanism of runoff generation under the plateau climate.

Key words: the upper Yangtze River, annual runoff, ATD, air temperature, GAMLSS