| Summary: | Simple linear regression models have been widely employed in the analysis of suspended-sediment concentration (SSC) time series from glacierized catchments, although they have many limitations. This paper builds regression models which address these shortcomings and permit inferences concerning the controls on suspended-sediment transfer from a glacier at 788N in the Svalvard archipelago. A bivariate regression model, deterministically predicting SSC from discharge alone, explained less than 15 per cent of the variance in SSC. A multivariate model, incorporating additional potentially explanatory variables, offered little improvement. Diurnal hysteresis in the data gives rise to quasi-autocorrelation in the residual series from regression models. This was effectively removed by incorporating dummy diurnal variables into the multivariate model. The presence of a first-order autoregressive, stochastic process gives rise to true autocorrelation in the residual series from regression models. This was accommodated by incorporating an ARIMA (1,0,0) term into a multivariate autoregression model. The model-building process yielded a systematic progression in the explanation of variance in SSC, stripping away pattern in the autocorrelation function of the residual series; mean model error was reduced from 54 per cent to 6 per cent. The dependence of SSC on the magnitude of discharge is weak and highly variable, whereas the dependence of current SSC on recent values of SSC, revealed through the stochastic term, is an order of magnitude greater and relatively constant during the melt season. The dominant control on SSC throughout the melt season is therefore short-term sediment availability. The simple and largely unchanging stochastic process generally responsible for generating the observed SSC series implies a simple and unchanging glacier drainage system.
|
|---|