Evaluating Apparent Temperature in the Contiguous United States From Four Reanalysis Products Using Artificial Neural Networks

Abstract Reanalysis data sets are the products of data assimilation systems and while they exhibit consistency with the observed climate, they tend to be biased, especially at a local scale. In this study, we present an innovative approach for the assessment of the accuracy of four reanalysis data sets—European Center for Medium‐Range Weather Forecasts, Reanalysis 5 (ERA5), Modern‐Era Retrospective analysis for Research and Applications Version 2, North American Regional Reanalysis, and the Twentieth Century Reanalysis—in capturing daily and extreme observed apparent temperatures in over 300 stations in the contiguous United States (USA). We modeled the relationship between the reanalysis and observed data during a training period and subsequently used these models to predict daily observations during the test period. This process enabled potential adjustment of the biases in the reanalysis data sets (i.e., the predictors). For predictive modeling, the performance of a Feed‐Forward Neural Network (FFNN) and a linear regression (LR) were assessed. Our results show the data assimilations were closer to observations in the eastern parts compared to the western parts of the USA. Considering the number of stations with the lowest error and highest hit rate (for extreme events), FFNN outperformed LR except for extreme warm events where LR performed better. Among the reanalysis data sets, ERA5 consistently exhibited the highest performance in predicting both daily and extreme values of apparent temperature, with a larger margin of relative accuracy for cold extremes. Thus, ERA5 is more adept at accurately representing apparent temperatures in the USA.