Anomaly Kriging Helps to Remove Bias in Spatial Model Runoff Estimates

The low spatial density of streamflow gauging stations limits the accuracy of spatial streamflow estimates in many parts of the world. Strategies to improve runoff estimates in the absence of dense measurements have tended to focus on estimating parameters of runoff models in ungauged regions, through so‐called parameter regionalization methods. However, parameter regionalization can be affected by overdependence on calibration at gauged sites, model parameter equifinality, and ensuing estimation errors. As a result, spatial model runoff estimates typically exhibit spatially correlated biases. This analysis attempts to enhance the use of observations in spatial runoff estimation. Specifically, we assessed the potential to reduce systematic errors by spatially interpolating residuals (i.e., errors) between prior grid‐based streamflow estimates for Australia at 0.05° × 0.05° grid from the Australian Bureau of Meteorology's calibrated, operational Australian Water Resources Assessment Landscape model (AWRA‐L) and streamflow gauging records from 780 unimpeded, relatively small catchments. We analyzed spatial autocorrelation in residuals and tested an efficient two‐step correction approach involving a uniform correction and subsequent kriging of residuals. The approach removed an average of 41% of systematic bias in the model estimates and also improved other model performance measures. Further reduction in errors at shorter timescales may be achievable through a temporally hierarchical correction scheme. We developed a method to reduce spatially variable bias in modeled runoff at the continental scaleThe two‐step regression and kriging scheme removed an average 41% of bias in cross validationCorrections were applied to continuous daily fields of runoff at a 0.05 × 0.05° grid across Australia