Phenological precipitation windows are associated with carotenoid pigmentation in durum wheat: Within-network prediction and genotype stability across contrasting environments

Yellow index, a primary commercial quality trait in durum wheat, shows strong environmental variation, but its climate correlates remain unresolved at the phenological scale. We analysed 871 plots, of which 740 yielded a yellow index, from 35 cultivars × 7 stations × 7 harvest years (2017–2023) in Morocco’s multi-environment trial network. Climate variables from ERA5-Land and CHIRPS were aligned to seven growing degree day (GDD)-defined phenological phases and tested against environment-level yellow index. Of 49 phase × variable combinations, 16 passed false discovery rate correction (q < 0.05). Cumulative maturation precipitation was the dominant environmental correlate (r = −0.706, p < 0.0001; a decline of 0.101 b* per millimetre of maturation rainfall), and remained so after controlling for grain-filling temperature and vapour pressure deficit (partial r = −0.682). On the 23 environments measured natively as b*, without ppm conversion, the association was stronger (r = −0.820; 0.134 b* per millimetre). A moisture-timing asymmetry emerged: vegetative-phase rainfall correlated positively with pigmentation (r = +0.508 to +0.591) while maturation rainfall correlated negatively (r = −0.706). A three-feature precipitation model was cross-validated within the network (leave-one-station-out R² = 0.73, 95% CI 0.41–0.84, falling to 0.66 when feature selection was repeated inside each fold; leave-one-year-out R² = 0.56 with station intercepts estimated from training years only). With only six stations, prediction at genuinely new sites remains to be established. Calendar-based phenology failed spatially (R² = 0.21). Finlay–Wilkinson slopes differed significantly among the 32 registered cultivars (F(31, 641) = 1.80, p = 0.0055) and ranged from 0.51 to 1.40; ‘Nachit’ had the lowest slope among well-sampled cultivars (0.74, 95% CI 0.50–0.98), though no individual cultivar remains significant once the 32 comparisons are corrected for. National and imported releases showed no evidence of a difference (p = 0.69), suggesting that pigment stability may need to be considered explicitly in selection. Network-mean yellow index declined by 0.56 b* yr⁻¹ across the seven harvest years (p = 0.009), a series too short to establish a long-term trend. Together, the rainfall–pigmentation gradient, the cross-validated model and the cultivar rankings provide a basis for anticipating pigmentation within this Moroccan network.