Hybrid Modeling of Cercospora Leaf Spot Epidemiology: Integrating Mechanistic and Machine Learning Approaches Using Remote-Sensing and Environmental Data

Despite advances in modeling and sensing, no study has previously integrated mechanistic, meteorological and UAV data into a unified predictive framework for Cercospora leaf spot. From 2020 to 2022, field trials with a susceptible variety under contrasting fungicide regimes and artificial inoculation were monitored for disease severity, airborne inoculum, and yield. Significant treatment differences emerged 44 days after sowing, with incubation lasting 7–12 days and spore peaks occurring from day 77, preceding rapid severity increases. Dissemination showed no prevailing direction but was favored by light, variable winds under conducive microclimates. Yield loss reached up to 0.0123 kg root fresh weight per plant per severity point, and both yield and sugar content decreased with earlier onset and higher final severity. Hybrid models were implemented at multiple levels, integrating multisource data. Severity was best predicted by climatic variables with UAV spectral–structural indices; fructification by humidity–temperature thresholds with stress traits; dissemination by wind-variability metrics with sporulation indicators; and yield and sugar content by UAV indices supplemented with mechanistic covariates. High-level hybridization reduced the RMSE to 0.615 (on a 0–10 severity scale), 0.067 ng C. beticola DNA for actual spores, 2.033 ng for cumulative spores, 1.769° for dissemination direction, 0.015 ng day⁻¹ for dissemination magnitude, 0.235% for sugar content, and 0.051 kg plant⁻¹ for root fresh weight, achieving up to a 39% improvement over lower-level configurations. These results enhance disease prediction, improve the understanding of disease epidemiology, and could support more effective plant-disease management.