Integrating Optical Sensing and Environmental Data for Predicting and Monitoring Plant Diseases in Sugar Beet

Abstract

Cercospora leaf spot (CLS), caused by the fungus <em>Cercospora beticola</em>, is a major foliar disease in sugar beet that shows increased incidence under favorable climatic conditions. This dissertation develops a modular and scalable framework for the detection, quantification, and prediction of disease dynamics under real field conditions. The framework is primarily based on imagery obtained from uncrewed aerial vehicles (UAVs) and integrates remote sensing, machine learning, and epidemiological modeling. <br/> A georeferenced, plant-level monitoring system was established using UAV-acquired red-green-blue and multispectral images. This system enables consistent identification and tracking of individual plants throughout the growing season and supports the generation of standardized datasets for phenotyping. Using this foundation, convolutional neural networks were trained to detect and classify CLS severity directly from UAV imagery with high accuracy. The methodology was successfully applied to Rhizoctonia crown and root rot, demonstrating its robustness and adaptability to diseases with distinct symptomatology. <br/> To evaluate scalability and interoperability, CLS detection was further investigated using satellite and ground-based datasets. A U-Net model was trained to convert UAV images into synthetic satellite observations with characteristics similar to those of PlanetScope data for disease classification. Complementarily, a multisensor ground platform equipped with hyperspectral, thermal, RGB, and LiDAR sensors allowed for presymptomatic detection and detailed canopy characterization. These extensions demonstrate the ability to detect CLS across multiple spatial scales and sensor modalities. <br/> Finally, a hybrid epidemiological model was developed by integrating image-derived features with environmental data and mechanistic, phase-specific predictors. The epidemic process was divided into biologically meaningful phases, including incubation, sporulation, dispersal, and yield impact. This approach enhances both predictive accuracy and biological interpretability. <br/> This dissertation makes three primary contributions: the development of a temporally structured, plant-based monitoring system; the demonstration of a cross-platform sensing strategy spanning UAVs, ground-based sensors, and satellites; and the implementation of a hybrid modeling pipeline that combines mechanistic epidemiological understanding with remote sensing and artificial intelligence. Together, these advances provide a strong foundation for predictive, explainable, and scalable disease monitoring aimed at sustainable crop protection