Abstract: Accurate monitoring of crop nitrogen status can be used to predict grain quality before harvesting. It is often required for precision nitrogen management and high-quality wheat in sustainable agriculture. However, information redundancy can reduce modeling efficiency and weaken effective spectral responses to crop nitrogen status, because canopy hyperspectral data contains numerous highly correlated bands. In addition, the formation of grain protein is a continuous physiological process, including nitrogen accumulation, remobilization, and translocation over multiple growth stages. One growth stage cannot fully characterize the dynamics of grain protein accumulation. In this study, leaf nitrogen content (LNC) inversion was developed using spectral transformation and characteristic-band selection. LNC information was further used to predict grain protein content (GPC) at multiple growth stages. Canopy hyperspectral reflectance data were collected at the jointing, flowering, and grain-filling stages of winter wheat. The original spectra were transformed using first- and second-derivative processing. Characteristic bands were selected using Pearson correlation analysis and the Boruta algorithm. Gaussian process regression (GPR), random forest regression (RFR), and support vector regression (SVR) were then used to establish stage-specific LNC inversion models. In parallel, two GPC prediction pathways were constructed: a direct pathway with canopy spectral features and an indirect pathway with LNC derived from hyperspectral data as an intermediate physiological variable. In the indirect pathway, LNC information was evaluated from individual growth stages and combinations of multiple growth stages. A systematic investigation was implemented to determine the contribution rates of integrated nitrogen status to GPC prediction. The results showed that the first-derivative reflectance was more strongly correlated with LNC than the original spectra, with the maximum absolute Pearson correlation coefficients of 0.74, 0.79, and 0.88 at the jointing, flowering, and grain-filling stages, respectively. Characteristic-band selection further improved LNC estimation, compared with full-band input. The coefficients of determination (R²) were 0.67, 0.80, and 0.77, respectively, with the root mean square errors (RMSEs) of 0.38%, 0.38%, and 0.36% for LNC estimation at the three growth stages after Boruta screening. The first-derivative transformation with Boruta feature selection also reduced spectral redundancy to retain bands closely correlated with the variations in LNC. In direct GPC prediction, the performance varied among growth stages. The flowering-stage GPR model achieved the highest accuracy, with the validation R² of 0.87 and an RMSE of 0.56%, indicating that canopy spectral information at flowering was closely correlated with subsequent grain protein formation. In indirect GPC prediction, the model with LNC information from the jointing, flowering, and grain-filling stages performed better than those from a single growth stage, with a validation R² of 0.75 and an RMSE of 0.62%. The nitrogen status at different developmental stages also provided complementary information related to grain protein formation. LNC at the jointing stage represented nitrogen accumulation during vegetative growth. Flowering-stage LNC represented nitrogen status around the transition from vegetative to reproductive development, and grain-filling-stage LNC was correlated with nitrogen remobilization and translocation to developing grains. These growth-stage-specific nitrogen signals therefore provided a physiological linkage between canopy hyperspectral responses and temporal processes underlying GPC formation. The “canopy hyperspectral reflectance–multi-growth-stage LNC–GPC” pathway was finally established to serve as the hyperspectral inversion framework for winter wheat grain protein driven by leaf nitrogen content. The finding can provide the theoretical basis to monitor crop nitrogen nutrition for grain quality before harvest.