Multimodal signatures of stress: Revealing variations in blackleg expression across seed potato cultivars with UAV remote sensing

• We tracked blackleg development at individual plant-level with UAV multimodal data (84) • Hyperspectral, thermal and LiDAR metrics captured distinct responses to infection (84) • Substantial differences in captured symptoms were found amongst tested cultivars (82) • Their magnitude varied over time affecting disease detection capabilities (75) • Combining modalities in detection models offered small improvement (up to 4%) (80) Remote sensing technologies offer a promising approach for detection of plant diseases, enabling timely interventions to prevent the spread and minimise financial losses. However, a major gap exists in understanding how cultivar and growth stage variations might influence the stress responses captured by different modalities, and subsequently affect disease detection capabilities. This study aimed to bridge this gap by exploring the variation in plant-level responses to blackleg onset (caused by Dickeya and Pectobacterium species) in six cultivars with varying susceptibility levels through investigation of metrics extracted from UAV hyperspectral, LiDAR, thermal data. Whilst we found clear responses to infection in all modalities, substantial cultivar-based variations were present due to different levels of symptom expression. Hyperspectral data emerged as the most crucial for blackleg detection, with specific feature importance varying over the season. Early-season responses were most strongly reflected in PRI, while later in the season, PSRI became more informative. Thermal and structural metrics, while showing promise, exhibited varying sensitivity to infection across cultivars and growth stages, with significant variability observed even among healthy plants of different cultivars. Combining modalities in SVM detection models offered small improvement in disease detection capabilities, though the use of hyperspectral and LiDAR data together yielded the most consistent performance across the investigated dates (balanced accuracy of 90-94%). Still, models’ performance was substantially affected by cultivar variations. These findings highlight the critical need to account for cultivar-specific responses and the dynamic nature of disease symptom expression when developing remote sensing-based disease detection models for arable crops.