A comparative analysis of segmentation algorithms for detecting plant diseases using UAV imagery
TL;DR
The proposed pipeline demonstrates the feasibility of NDVI-based segmentation for agricultural monitoring and provides a foundation for future work incorporating additional spectral features and larger multi-field datasets.
Abstract
UAV-based multispectral imaging combined with deep learning offers significant potential for early crop health monitoring and decision support in precision agriculture. This study, conducted within the DB-KISS research project, addresses the problem of detecting structural field features and vegetation stress patterns in cabbage fields using multispectral UAV data. Five orthomosaics were acquired over a 13 ha field in Dithmarschen, Germany, using a DJI Matrice 350 RTK equipped with a MicaSense RedEdge-P sensor. The Normalised Difference Vegetation Index (NDVI) was computed and used as the sole input representation. The orthomosaics were divided into 640 × 640 pixel tiles, resulting in a dataset of 624 annotated samples with 819 wheel-track instances and 498 disease-related instances. A total of 12 YOLO-based segmentation models from multiple architecture generations were trained and evaluated under identical conditions using five-fold cross-validation. The results show that larger YOLO models consistently outperform smaller variants, achieving an mAP50–95 of up to 59.98, while maintaining stable performance across folds. However, limitations in dataset size and annotation precision affect the detection of small or weakly expressed stress patterns. The proposed pipeline demonstrates the feasibility of NDVI-based segmentation for agricultural monitoring and provides a foundation for future work incorporating additional spectral features and larger multi-field datasets.