Monday, 29 January 2024: 4:30 PM
Key 10 (Hilton Baltimore Inner Harbor)
Reanalysis datasets are the products of data assimilation systems and while they exhibit consistency with the observed climate system, their representation of observed climate tends to be biased, especially at a local scale. In this study, we present an innovative approach for assessing the accuracy of four reanalysis datasets - ERA5, MERRA2, North American Reanalysis, and the 20th Century Reanalysis - in capturing daily observed station precipitation anomalies and apparent temperature, in addition to extreme events, in the United States. Traditionally, reanalysis outputs are evaluated directly against observations. However, this approach may lack robustness as a few days with significant biases in the assimilated values can skew the results. To overcome this challenge, we employed predictive modeling to learn the relationship between the assimilated data and observed data during a training period and subsequently use these models to predict observations during the validation period. This process will enable adjusting the biases in the reanalysis datasets and facilitate more rigorous comparisons and potential downscaling. The performance of two predictive models: Feed-Forward Neural Network and traditional linear regression, will be evaluated. The focus will be to examine (i) the best predictive model of daily and extreme values of precipitation anomaly and apparent temperature, (ii) the reanalysis data with the best predictive skill, and (iii) the spatial dependency of the results in (i) and (ii), in the United States.

