Neural Network–Based Sensitivity Analysis of Summertime Convection over the Continental United States

Abstract

Abstract Although land–atmosphere coupling is thought to play a role in shaping the mean climate and its variability, it remains difficult to quantify precisely. The present study aims to isolate relationships between early morning surface turbulent fluxes partitioning [i.e., evaporative fraction (EF)] and subsequent afternoon convective precipitation frequency and intensity. A general approach involving statistical relationships among input and output variables, known as sensitivity analysis (SA), is used to develop a reduced complexity metamodel of the linkage between EF and convective precipitation. Two additional quantities characterizing the early morning convective environment, convective triggering potential (CTP) and low-level humidity (HIlow) deficit, are included. The SA approach is applied to the North American Regional Reanalysis (NARR) for June–August (JJA) conditions over the entire continental United States, Mexico, and Central America domain. Five land–atmosphere coupling regimes are objectively characterized based on CTP, HIlow, and EF. Two western regimes are largely atmospherically controlled, with a positive link to CTP and a negative link to HIlow. The other three regimes occupy Mexico and the eastern half of the domain and show positive links to EF and negative links to HIlow, suggesting that both surface fluxes and atmospheric humidity play a role in the triggering of rainfall in these regions. The regimes associated with high mean EF also tend to have high sensitivity of rainfall frequency to variations in EF. While these results may be sensitive to the choice of dataset, the approach can be applied across observational, reanalysis, and model datasets and thus represents a potentially powerful tool for intercomparison and validation as well as for characterizing land–atmosphere interaction regimes.