Cross-season candidate enrichment in potato breeding using a multi-temporal UAV-based 3D phenotyping framework

Accurate extraction of canopy traits from ridge-cultivated potato plots remains challenging because uneven micro-topography, plastic mulch glare, and saturation of two-dimensional canopy metrics can reduce the reliability of unmanned aerial vehicle (UAV) phenotyping. In this study, we developed a multi-temporal UAV-based workflow to improve the robustness of plot-level phenotyping and its utility for cross-season screening in potato breeding. Using two large field trials conducted in 2024 and 2025 (460 and 706 accessions, respectively), we combined hue-saturation-value (HSV) and normalized difference vegetation index (NDVI) segmentation to suppress mulch-related noise and improve early seedling detection. To reduce terrain-related structural bias, we applied a terrain-constrained two-stage point cloud registration procedure, and we used relative-time alignment to improve comparability of developmental trajectories across seasons. The cascaded HSV-NDVI method achieved 93.0% seedling detection accuracy under complex field conditions, and the registration workflow reduced the vertical residual error from 8.4 cm to 1.4 cm. The resulting τ-aligned descriptors, including canopy height, canopy volume, and dynamic canopy development traits, showed consistent associations with yield across seasons. In independent cross-season tests, top-33% model-based selection increased mean yield by 8.50%–13.18%, depending on classifier and feature set. A comparison of seven classifiers showed that advanced models improved performance in some settings, especially when relative-time descriptors were included, but no single model consistently resolved the difficulty of cross-season high-yield classification. These results suggest that the framework is more suitable for early-stage candidate enrichment than for precise ranking of the entire breeding population.