Quantitative assessment of sugar beet Cercospora infection based on UAV multispectral imaging and U-Net segmentation

Cercospora leaf spot (CLS) of sugar beet, caused by Cercospora beticola Sacc., is a highly destructive plant disease that can reduce yields by up to 40% and significantly impair root crop quality. This study aimed to develop and validate a quantitative disease assessment method utilizing UAV-based multispectral imaging and semantic segmentation. Field trials were conducted in 2023-2024 on commercial sugar beet crops (Agrofirma Start OOO, Buzdyak District, Republic of Bashkortostan). Plots, measuring 20 × 6 rows (≈21.6 m²), included both control and artificially inoculated treatments. UAV imagery was acquired using a Geoscan ChatGPT equipped with a Pollux multispectral camera (Blue, Green, Red, Red-edge, NIR) at an altitude of approximately 30 m. A U-Net model was trained on 420 annotated image tiles (512 × 512 px), using a 6:2:2 split for training, validation, and testing. The model incorporated spectral indices (NDVI, NDRE, MCARI, NSVDI) in addition to geometric features derived from the normal vectors of a Digital Surface Model (DSM). The developed integrative algorithm achieved an overall multiclass classification accuracy of 88.6%. Specifically, an F1-score of 46.0% was obtained for the ‘infected plants’ class, outperforming Partial Least Squares Discriminant Analysis (PLS-DA) by 18.6 percentage points. F1-scores reached 92.5% for ‘Healthy vegetation’ and 68.7% for ‘Soil/Background.’ This methodology confirms the strong applicability of U-Net for the diagnostic segmentation of Cercospora outbreaks, significantly enhancing the objectivity of crop monitoring. The integration of spectral and geometric features proved crucial in improving the detection of weakly expressed disease symptoms.