Impact-oriented evaluation of stochastic weather generators for rice blast risk modeling: Implications for agricultural climate services

Process-based plant disease models require daily weather inputs, whereas climate information used in agricultural climate services is often available at coarser temporal scales. We evaluated whether stochastic weather generators (WGs) can reproduce daily weather characteristics relevant to process-based disease modeling using GEM and ClimGen with the EPIRICE rice blast model. Using 30 years of observations from five Korean weather stations, 30 independent 30-year synthetic realizations were evaluated for each station and WG. Performance was assessed using conventional weather statistics, quantitative metrics within the 100-day EPIRICE simulation window, EPIRICE-derived area under the disease progress curve (AUDPC), and process-informed diagnostics of infection-conducive weather conditions. ClimGen reproduced temperature more closely, whereas GEM showed lower precipitation RMSE, and relative-humidity performance varied by station and metric. GEM underestimated mean AUDPC at all stations, whereas ClimGen overestimated it, although ClimGen showed smaller absolute mean bias at four of five stations. Both WGs produced reduced AUDPC variability relative to observed-weather-driven simulations. Process-informed diagnostics showed contrasting strengths: ClimGen more closely reproduced infection-conducive temperature conditions and moderate precipitation events, whereas GEM more closely reproduced high-relative-humidity conditions at most stations. No single WG was consistently superior across weather-statistical and disease-model-oriented metrics. These findings demonstrate that WG suitability for process-based disease modeling cannot be inferred from conventional climatological agreement alone and should be assessed using both weather-statistical and impact-oriented criteria before application in forecast-oriented agricultural climate services.