Abstract Mesoscale convective systems (MCSs) are large, organized convective storms that frequently produce flash floods and other severe hazards such as damaging winds, hail, and tornadoes. Developing an observationally based MCS hazard climatology is important for establishing a baseline to evaluate the representation of these events in numerical models. This study constructs such a climatology using a 13‐year MCS data set, storm reports, and atmospheric reanalysis. MCS‐related and near‐storm environmental variables are extracted and used to train object‐based machine learning (ML) models. Three models are developed to predict flash floods, severe (including all wind, hail and tornado events), and significant‐severe events, with the latter representing higher‐impact hazards. The flash flood and severe models perform well in distinguishing hazard‐producing MCSs from non‐producing ones, while the significant‐severe model shows limited skill, likely due to sample size constraints. The flash flood and severe models are then applied to the full MCS archive to reconstruct a more complete warm season hazard climatology, addressing the potential underreporting in storm reports and gaps in flash flood reports during early years. The study also examines the application of these models to convection‐permitting model simulations. While the spatial distribution of simulated MCS hazards generally aligns with observations, event frequencies differ considerably. These discrepancies are attributed to biases in both the derived input variables and the representation of MCS properties within the model.
Cui et al. (Thu,) studied this question.