3. On the Physical and Operational Rationality of Data-Driven Models for Suspended Sediment Prediction in Rivers

Suspended sediment remains an important variable for prediction in river studies. Knowledge of suspended sediment concentration or load at different downstream locations within a channel allows the temporal and spatial patterns of catchment sediment yield to be determined, as well as within-channel sediment budgets that provide important insight into the patterns and processes governing channel siltation and scour. However, in many rivers, the comprehensive and long-term downstream monitoring of discharge contrasts with relatively sparse and temporally discontinuous monitoring of suspended sediment. As a result, the generation of suspended sediment data through physical and empirical modelling approaches is commonplace. The emergence of a data-driven modelling paradigm in the last two decades has resulted in the adoption of new methods for suspended sediment modelling. To a large extent, these methods mirror traditional empirical approaches, except that the a priori determination of the form of the response function by the modeller is replaced by machine learning and artificial intelligence algorithms that ‘learn’ the response function (both its form and associated parameters) directly from data. Data-driven models have been shown to result in improved goodness-of-fit metric scores, but many hydrologists remain critical about the lack of attempts by data-driven modellers to demonstrate the physical rationality and operational validity of their models. In this chapter, we examine this criticism; highlight specific research challenges facing data-driven, suspended sediment modellers; and detail the research directions through which advances may be made.