Calibration of hydrologic models: The role of input errors

Hydrologic models are widely used in water resources and catchment studies. However, significant sources of error exist in input and response data, and are exacerbated by uncertainty in the model structure itself. Established uncertainty estimation techniques typically assume independent residuals in the response variable whilst neglecting potential errors in forcing data. Such techniques therefore simplistically transfer the effects of multiple sources of uncertainty onto the uncertainty of the parameter estimates alone. In this study, a rainfall-runoff model (TOPMODEL) is forced with corrupted rainfall data and calibrated to synthetic data to exclude model error. It is shown that corrupt rainfall induces persistent auto-correlation in the runoff error series. The slow dynamics of storage processes exacerbate this auto-correlation and invalidate the principal assumptions of least-squares methods. As a result, parameter tractability and model's predictive capacity are degraded. A new approach is then developed, which explicitly accounts for temporally distributed errors in the forcing data.