Similarity Assessment of NLDAS Multi-Model Ensemble Outputs

Multi-model ensembles have been used in numerous weather, climate, and hydrologic projects to produce ensemble mean estimates with increased simulation skill of model outputs. The North American Land Data Assimilation System, NLDAS, uses a multi-model ensemble of land-surface models to produce outputs of states (e.g., soil moisture and temperature, snowpack), energy (e.g., sensible, latent, ground heat) and water (e.g., evapotranspiration, total runoff/streamflow) fluxes Because of different development philosophies, formulations, and parameterizations applied in the models, the current NLDAS models (Mosaic, Noah, SAC, and VIC) yield different simulations of land surface states. It is expected that the ensemble members will have different levels of similarity and dissimilarity depending on factors such as climatic regimes, seasonality, model parameters, topography, and geography. If multi-model outputs are very similar, it can be argued that they add little additional information to the multi-model ensemble. On the other hand, if one model is very dissimilar, it may indicate errors in the simulation, leading to increase in the mean bias and the variance of the ensemble. Therefore, a similarity assessment of NLDAS multi-model ensemble outputs is necessary for assessing the fidelity and usefulness of the ensemble, as these ensembles are often used for applications such as probabilistic prediction of drought. In this presentation, we will provide an analysis of the models of similarity of the NLDAS ensemble members and quantify their spatial and temporal dependence. Tradeoffs between the similarity metrics and accuracy measures (developed through comparisons against independent observations) will also be developed, for important water cycle components such as soil moisture, evapotranspiration, snow, and runoff. We expect this work to contribute towards the formulation of procedures for evaluating the utility of future additions (such as the Noah-MP model) and explore the impact of member size on the NLDAS multi-model ensemble.