Evaluating and Post-processing of the North American Multi Model Ensemble Data Set for Hydrologic Drought Outlook

Reliability and accuracy of hydrologic forecast and drought outlook are directly affected by the uncertainty in the forcing data. Despite recent progresses in climate forecasting, the data sets are still prone to larges biases. Therefore, improving the accuracy of climate forecast has been an ongoing and important research topic, and bias correcting climate forecast is a necessary step in the operational hydrologic forecasting. This study provides an assessment of the North American Multi-model Ensemble (NMME) forecast products over the CONUS and introduces an effective post-processing (bias correction) method, and investigates its application in Ensemble Streamflow Prediction (ESP).

In this process, monthly precipitation of 11 contributing models in the NMME (totaling 128 ensemble members) are evaluated and then bias corrected for the historical period of 1982-2010 and verified for the hindcast period of 2012-2015. For a more detailed analysis, all the models are re-gridded to 0.5-degree spatial resolution. A new Copula-based ensemble post-processing (COP-EPP) method is introduced to improve the performance of NMME forecasts at four different lead-times (lead-0 to lead-3). The proposed technique is rooted in Bayesian networks for conditioning the forecast on the observations. To assess the performance of the proposed method, each of the 128 ensemble members is bias corrected with Quantile Mapping (QM) as a simple and widely used bias correction approach. Results indicate poor performance for the NMME across the western and central US. Both of the bias correction techniques demonstrate significant improvement over the raw NMME. However, COP-EPP is showing superiority over the QM.

The hydrologic forecasting is issued by forcing the Variable Infiltration Capacity (VIC) hydrologic model by the post-processed climate forcing at 1/8^th degree spatial resolution for the Columbia River basin in the Pacific Northwest US and hydrologic drought outlooks are generated at different lead times.  Various probabilistic verification metrics are employed to demonstrate the usefulness and effectiveness of this approach.