A Data-Driven Identification Scheme for Winter Weather Events: Integrating Historical Storm Reports and Atmospheric Reanalysis

Abstract Robust and accurate identification of adverse winter weather conditions can substantially contribute to preparedness, thus supporting the effective application of mitigation measures for the protection of human lives and infrastructure. Due to the highly nonlinear nature of the underlying systems, data-driven approaches can offer a versatile tool toward classifying potentially hazardous conditions from incoming storms. In this work, we identify instances of adverse winter weather over the contiguous United States (CONUS), by applying random forest and extreme gradient boosting algorithms, which are trained on ERA5 atmospheric reanalysis data and event reports generated by the National Oceanic and Atmospheric Administration for a period of 27 years (i.e., from 1996 to 2022). To ensure effective and robust model performance, as well as to interpret the obtained estimates, we perform a multifaceted evaluation of the proposed binary classification scheme. As a benchmark for assessing accuracy, we employ ERA5 snowfall, as well as the winter storm severity index for certain major historical storms that occurred over the study domain. The scheme is exhibited to have sufficient skill in extracting elaborate relationships among atmospheric covariates, which can enhance winter storm outlook over CONUS and provide a platform for winter-related impact studies. The uniqueness of this study lies in integrating reports of significant weather phenomena, with atmospheric, climatic, and human–environment data, to develop a robust and versatile model for predicting hazardous winter weather on a continental scale. While the model is trained and tested using reanalysis data in conditions resembling an operational setting, its aim is to be deployed with numerical weather prediction data for real-time forecasting.