Direct statistical downscaling of monthly streamflow from atmospheric variables in catchments with differing contributions from snowmelt

Abstract

The climate-hydrological modelling chain is the most widely used approach for assessing likely future changes in streamflow but is also associated with large uncertainties. We present an alternative approach for modelling monthly streamflow directly from large-scale atmospheric variables using statistical downscaling to streamflow for 42 catchments with flow generation regimes ranging from largely snowmelt-driven to entirely rainfall-driven. It aims to find the most appropriate combination of predictors and methods, which can best include information pertaining to snow storage and provide robust streamflow simulations under changing climate conditions. The effects of selection of probable predictors (five groups of potential variables from reanalysis data, three moving windows, calibration periods of two different lengths) and methods (eight combinations of pre-processing methods and statistical models) were considered. Results indicate that the following selection of probable predictors gave a good and robust model performance: a combination of 13 independent climate variables and their lagged information, a 3-month moving window and a calibration period of 31 years. Three combinations of methods performed best: (a) principal component analysis (PCA) and the relevance vector machine (RVM) statistical model using gridded data, (b) PCA and RVM using interpolated data and (c) the Pearson correlation analysis and RVM using interpolated data. Monthly streamflow was well reproduced for most catchments, with a median NSE (Nash and Sutcliffe efficiency) of 0.82 and PBIAS (percentage bias) of 0 in the calibration period and a median NSE over 0.60 and PBIAS less than −6 in the validation period. The best performance of the methods was associated with snow-dominated catchments in inland regions. We also established that lagged information on atmospheric variables gave better results than snow variables from reanalysis data. The poorest results were associated with rainfall-dominated catchments, and this was

An integrated multi‐GCMs Bayesian‐neural‐network hydrological analysis method for quantifying climate change impact on runoff of the Amu Darya River basin

“…Alternatively, statistical models were widely used for runoff simulation due to their ability in capturing arbitrary input–output relationships, without explicit descriptions of the underlying physical processes (Li et al ., 2015; Zhuang et al ., 2017; Huang et al ., 2020). For instance, Li et al .…”