Environment-Specific Guidance Improves Forecast Skill for Isolated Convection Initiation over South China

Abstract Is convection initiation (CI) harder to forecast in some large-scale environments? Using 8,607 radar-detected isolated CI events over South China during 2018–2023 and ERA5 fields, we identify two recurrent pre-CI environments: a thermal environment with broad warm-moist support and a dynamic environment with more organized flow. Across 200 forecast models, dynamic cases generally have lower skill than thermal cases. Motivated by this difference, we develop an environment-conditioned correction that adapts baseline predictions using weights inferred from each sample’s circulation state. Across 20 paired runs, the correction increases the area under the receiver operating characteristic curve (AUC) for dynamic cases by 12.4% and reduces Brier score by 0.040, while retaining thermal-case skill. Exploratory comparisons indicate added skill from environmental weighting when dynamic CI and non-CI controls have contrasting circulation states, with geographical and temporal inputs held constant. These results show how environmental diagnosis can guide adaptive correction for the more difficult CI cases.