System-Wide Calibration of River System Models: Opportunities and Challenges

Abstract:

Semi-distributed river system models are traditionally calibrated using a reach-by-reach calibration approach from that starts from headwater gauges and moves downstream toward the end of the system. Such a calibration method poses a unique problem since errors related to over-fitting, poor gauging data and uncertain physical connection are passed downstream. Reach-by-reach calibration, while efficient, cannot compensate for limited/poor calibration data of some gauges. To overcome the limitations of reach-by-reach calibration, a system calibration approach is proposed in which all the river reaches within a river basin are calibrated together using a global objective function for all stream flow gauges. In this approach, relative weights can be assigned in the global objective function for different gauges based on the magnitude and quality of available data. The system calibration approach was implemented in a river network covering 11 stream flow gauges within Murrumbidgee catchment (Australia). This study optimises flow at the selected gauges within the river network simultaneously (36 calibrated parameters) utilising a process-based semi-distributed river system model. The model includes processes such as routing, localised runoff, irrigation diversion, overbank flow and losses to groundwater. Goodness of fit is evaluated at the 11 gauges and a flow based weighting scheme is employed to find posterior distributions of parameters using an Approximate Bayesian Computation. The method is evaluated against a reach-by-reach calibration scheme. The comparison shows that the system calibration approach provides an overall improved goodness-of-fit by systematically de-valuing poor quality gauges providing an overall improved basin-wide performance. Clusters of viable parameter sets are determined from the posterior distributions and each examined to assess the effects of parameter uncertainty on internal model states. Such a method of calibration provides a lot more flexibility in the calibration process, however because of the equifinality problem; it has increased potential to allow the model to produce unrealistic model states. The use of supplementary data (flood extent) is tested as a means of reducing parameter uncertainty, and producing more realistic model states.