Detecting non-stationary hydrologic model parameters in a paired catchment system using data assimilation

- Time varying model structures can improve prediction in changing catchments.
- Data Assimilation provides an objective framework for developing time varying model structures.
- The Locally Linear Dual EnKF is demonstrated to be an effective tool for time varying parameter estimation.
- Improved streamflow prediction and soil moisture representation for catchments undergoing deforestation are seen within this framework.

Non-stationarity represents one of the major challenges facing hydrologists. There exists a need to develop modelling systems that are capable of accounting for potential catchment changes, in order to provide useful predictions for the future. Such changes may be due to climatic temporal variations or human induced changes to land cover. Extensive research has been undertaken on the impacts of land-use change on hydrologic behaviour, however, few studies have examined this issue in a predictive modelling context. In this paper, we investigate whether a time varying model parameter estimation framework that uses the principles of Data Assimilation can improve prediction for two pairs of experimental catchments in Western Australia. All catchments were initially forested, but after three years one catchment was fully cleared whilst another had only 50% of its area cleared. Their adjacent catchments remained unchanged as a control. Temporal variations in parameters were detected for both treated catchments, with no comparable variations for the control catchments. Improved streamflow prediction and representation of soil moisture dynamics were also seen for the time varying parameter case, compared to when a time invariant parameter set from the calibration period was used. While we use the above mentioned catchments to illustrate the usefulness of the approach, the methods are generic and equally applicable in other settings. This study serves as an important validation step to demonstrate the potential for time varying model structures to improve both predictions and modelling of changing catchments.