Forecasting Models for Fire Blight Severity in Pear Orchards Under Tunisian Climatic Conditions

Fire blight, caused by Erwinia amylovora, is a devastating bacterial disease affecting pear orchards and has been reported in Tunisia since 2012/2013. Accurate prediction of infection risk is essential for timely disease management. This study aimed to develop forecasting models that correlate epidemiological factors with local microclimatic conditions in a Tunisian pear orchard. The study was conducted in a pear orchard in Ouerdanine (Monastir, Tunisia) from March to June 2023, on the ‘Alexandrine’ variety. Weekly epidemiological monitoring was performed, and disease severity was quantified using a standardized index based on a scale of 0 to 6. An autonomous weather station continuously recorded ambient temperature, flower-level temperature, and relative humidity, which are key variables for pathogen development. Two predictive models were developed after careful data preparation and normalization: (1) a tree-level model using machine learning algorithms, including Decision Tree, Random Forest, and Support Vector Machine, based on individual sensor measurements, and (2) an orchard-level model using Linear Regression with aggregated satellite meteorological data. Local microclimatic conditions strongly influenced disease severity. Machine learning models showed moderate performance in predicting infection at the individual tree level, with the Decision Tree algorithm achieving the highest accuracy (48%), followed by Random Forest (44%) and SVM (29%). The regression model provided reliable estimates of overall orchard severity, allowing for the assessment of disease trends at a larger scale. These forecasting models, adapted to the Tunisian Mediterranean climate, offer practical tools for fire blight management. They can help optimize the timing of chemical and biological preventive measures, improving control of this major phytosanitary threat. However, further validation across multiple orchards, varieties, and seasons is needed to confirm their broader applicability. This approach demonstrates the potential for adapting established models to local conditions in Tunisia.