Sensing the occurrence and dynamics of Cercospora leaf spot disease using UAV-supported image data and deep learning

The most damaging foliar disease in sugar beet is Cercospora leaf spot (CLS), caused by Cercospora beticola Sacc. The pathogen is expanding its territory due to climate conditions, generating the need for early and accurate detection to avoid yield losses. In Germany, monitoring and control strategies are based on visual field assessments, with the parameter disease incidence (DI). This parameter triggers warning systems when a threshold is achieved, and decision-making takes place for fungicide application. However, visual scoring is a time-consuming activity that requires well-trained personnel and is the principal bottleneck for CLS control. Digital technologies can support this process. Thus, the present work is based on two trial fields conducted and monitored in 2020 using an unmanned aerial vehicle (UAV) equipped with a multispectral camera. Image data were collected in time series during the vegetation period. Trials were sown with different sugar beet varieties; for field management, there was employed diverse fungicide strategies, and artificial inoculation took place in a spot manner. Parallel to the flight mission and additional assessment of DI, disease severity (DS) via KWS scale was collected by experts as so-called ground truth (GT). Combined with image-processing, it was possible to catalogize plants in field trials, identify them over time, and use them for training and testing models. A convolutional neural network (CNN) supported by cataloged data was trained to perform classification of the disease presence in time-series, and performance was evaluated. As the last image processing step, maps were generated showing site-specific distribution of the diseased plants in the field. Generated maps can serve as a basis for application maps in practical cultivation or the evaluation of variety performance in variety trials. The presented methodological approach provides high precision and sensitivity in CLS detection and offers the potential to automate processes of CLS monitoring for different application areas.