Abstract
Reliable assessment of yield models in breeding trials requires evaluation on lines that have not contributed observations to model development. We examined this problem in an advanced spring wheat yield nursery in northern Kazakhstan during the 2024 and 2025 seasons. Of 300 plots, 176 from 22 entries formed the model development set, while 124 plots representing 31 breeding lines were completely excluded from training. Eleven variables were retained from 82 candidates using the training data alone; the candidates were derived from UAV multispectral imagery, ERA5-Land reanalysis, LiDAR-derived static plot microtopography, phenology and maturity group. LightGBM yielded a pooled R² of 0.587 for the excluded lines; resampling breeding lines within season gave a mean R² of 0.564 with a 95 % confidence interval from 0.252 to 0.776. The pooled value partly reflected the strong contrast between seasons rather than discrimination among lines within a season. Random division of plots produced an R² of 0.890, indicating that apparent predictive accuracy depended strongly on validation design. Agreement between observed and predicted line rankings was moderate, with Spearman correlations of 0.612 in 2024 and 0.650 in 2025, although uncertainty was wide because only 14 and 17 independent lines were available. Six modelling approaches produced overlapping uncertainty intervals, with no clear evidence of superiority, and the complete combination of data sources was not separable from simpler configurations retaining the multispectral block. A separate analysis in which screening, selection and fitting used 2024 data alone and evaluation used 2025 data showed a severe loss of performance, with R² = −8.81 and a Spearman correlation of −0.082 that was indistinguishable from zero. Poor transfer persisted among entries grown in both seasons, so genotype novelty alone was insufficient to explain the failure. Breeding lines absent from training could therefore be ranked with moderate consistency when both observed seasonal regimes were represented during model development, whereas prediction in an unobserved season remained unreliable. The study covers one nursery complex and two seasons, so transfer across locations and broader conditions remains to be established.