Use of data assimilation to improve rainfall-runoff model structure for climate change projections

Data assimilation is a powerful tool that has been used to correct states and parameters of rainfall-runoff models based on recent streamflow, remotely sensed soil moisture or groundwater data. Data assimilation is now routinely applied by forecasting centres around the world to improve simulations and increase forecast skill. In this work, we are less concerned with the direct benefits of data assimilation on model outputs, but more on the nature of the corrections introduced and how they can be analysed to diagnose structural deficiencies in rainfall-runoff models.

Rainfall-runoff models have been shown to lack extrapolation capacity in simulating dry and wet periods that are more extreme than calibration conditions. This is particularly concerning in the context of climate change studies where more climate extremes are generally predicted for expected. This is the case in South-Eastern Australia where annual rainfall is expected to decrease significantly under most climate scenarios. Consequently, the improvement of rainfall-runoff model structures to better simulate dry flow regimes is critical to obtain robust estimates of water resources availability.

In this work, we assimilated streamflow data in the GR2M monthly rainfall-runoff models for 100 catchments in South-East Australia. The assimilation was conducted during a wet period between 1970 to 1995 and used to identify model structure deficiencies, particularly in the function computing water exchanges with nearby catchments. An attempt of correcting these deficiencies was undertaken using a simple regression approach. Finally, the correction was applied during a dry period (1995-2010) and performance was compared with the original (uncorrected) model. The results suggest that the corrected simulations better capture streamflow extremes, especially low flows. Further work is also discussed related to the use of additional data such as LAI and groundwater data to better constrain the correction regression.