Multitemporal UAV Data for Spring Wheat Breeding: Seasonal Transfer, Line Ranking and Comparison with Regional Standards

Spring wheat breeding compares candidate lines across seasons with different growing conditions. We evaluated seasonal transfer of models based on multitemporal UAV observations from a northern Kazakhstan breeding nursery, considering absolute yield, line ranking and yield margins over standards of comparable maturity. Models developed based on 2024 and 2025 data were tested at six retrospectively defined cutoffs in 2026. Pooled development R2 was 0.814 for a random forest and 0.779 for a gated recurrent unit (GRU), whereas within-season R2 for line means ranged from 0.140 to 0.375. In 2026, line-mean R2 was negative at every cutoff, and no positive rank correlation was supported by its bootstrap interval. Referencing candidates to maturity-matched standards removed 67–98% of shared prediction bias but mean zero-margin skill remained negative. For the 20 candidates previously grown in the nursery, historical mean yield gave higher point-estimate rank correlations than either model. Among 16 candidates without historical yields, neither model showed a demonstrated ranking advantage over NDVI from the latest acquisition. Exploratory calibration using harvested yields from 20 lines partly improved prediction and ranking for 16 unseen candidates but did not represent an early-season procedure. In this nursery, reducing shared seasonal prediction bias did not recover reliable estimates of the yield advantages of individual lines.