Similarity Assessment of Land Surface Model Outputs in the North American Land Data Assimilation System

Multimodel ensembles are often used to produce ensemble mean estimates that tend to have increased simulation skill over any individual model output. If multimodel outputs are too similar, an individual LSM would add little additional information to the multimodel ensemble, whereas if the models are too dissimilar, it may be indicative of systematic errors in their formulations or configurations. The article presents a formal similarity assessment of the North American Land Data Assimilation System (NLDAS) multimodel ensemble outputs to assess their utility to the ensemble, using a confirmatory factor analysis. Outputs from four NLDAS Phase 2 models currently running in operations at NOAA/NCEP and four new/upgraded models that are under consideration for the next phase of NLDAS are employed in this study. The results show that the runoff estimates from the LSMs were most dissimilar whereas the models showed greater similarity for root zone soil moisture, snow water equivalent, and terrestrial water storage. Generally, the NLDAS operational models showed weaker association with the common factor of the ensemble and the newer versions of the LSMs showed stronger association with the common factor, with the model similarity increasing at longer time scales. Trade‐offs between the similarity metrics and accuracy measures indicated that the NLDAS operational models demonstrate a larger span in the similarity‐accuracy space compared to the new LSMs. The results of the article indicate that simultaneous consideration of model similarity and accuracy at the relevant time scales is necessary in the development of multimodel ensemble.