Analyzing streamflow variation in the data-sparse mountainous regions: An integrated CCA-RF-FA framework

Abstract

Abstract In this study, an integrated CCA-RF-FA framework (abbreviated as CRFF) is developed for analyzing the streamflow variation in the mountainous watershed. CRFF incorporates cross-correlation analysis (CCA), random forest (RF), and factorial analysis (FA) within a general framework. CRFF can identify the time lag effect in the runoff mechanism (both rainfall and glacier/snow meltwater), tackle the problem of the simulation performance degradation caused by the time lag effect, as well as investigate the individual and interactive effects of meteorological and physical factors on runoff simulation. CRFF is applied to the Amu Darya River Basin (ADRB), a typical mountain watershed in Central Asia. Bayesian neural network (BNN) and stepwise cluster analysis (SCA) are used for comparison to illustrate the advantage of RF in streamflow simulation. The main findings reveal that (i) compared with BNN and SCA, RF has the better simulation capacity; (ii) the time lag effect of heat conditions such as the temperature (T) and shortwave radiation (SWR) is weak, with the lag time being less than 30 days; the time lag effect of the precipitation conditions such as snow precipitation rate (SPR) is strong, with a lag time ranging from 30 days to 90 days; (iii) in snow-melting, non-melting and the entire periods, T has the dominant impact on the variation of the runoff in ADRB; (iv) the interaction of T and SWR has important effect on the streamflow; SPR can still considerably affect the runoff generation in non-melting period.