Abstract: An accurate and timely determination of the crop growth stages is one of the key steps in precision agriculture. This study aims to realize the rapid, non-destructive, and precise identification of the winter wheat growth stages. A methodological framework of “multi-source parameter fusion—parameter optimization and dimensionality reduction—probabilistic model” was established to validate the identification. Specifically, a Bayesian classification model was constructed to integrate the skew distribution parameters of the canopy red-green-blue (RGB) color with the cumulative light–heat meteorological factors. Field experiments were conducted over four consecutive years at the observation sites in three ecological zones in Shandong Province, China, namely HeZe, Rai'an, and HuanTai. The high-resolution canopy red-green-blue (RGB) images and concurrent meteorological data were systematically collected throughout the growth cycles. Twenty parameters of the canopy color skew-distribution and four indices of the cumulative light-heat were extracted using manually observed phenological stages as the prior knowledge. A correlation analysis was performed to identify the key variables. Bayesian classification models were then constructed using canopy color parameters, meteorological factors, and their fusion. Model performance was also compared under different input conditions. The cross-year and cross-ecological-zone samples were used to verify the generalization and robustness. The results showed that both canopy color skew distribution parameters and cumulative light-heat indices were significantly associated with the growth stages of the winter wheat, thereby providing a theoretical foundation for intelligent identification. The Bayesian model with only meteorological parameters outperformed the model with only canopy color information. The performance of the model was often required for further enhancement over the different years and ecological regions. While the initial model was integrated with the canopy and meteorological parameters. The potential model also exhibited a risk of overfitting. Subsequently, the parameter optimization and correlation analysis showed that a refined and most effective parameter set was identified for the model input: red channel skewness, green channel kurtosis, accumulated temperature, and accumulated photosynthetically active radiation (ar_e). The Bayesian model was constructed with the optimal parameter set. The superior performance was achieved in a classification accuracy of 91.70%. More importantly, the optimal model significantly enhanced the robustness and generalization. The high and stable identification accuracies ranged from 84.00% to 100% for the independent samples over different years and ecological zones. The improved model substantially outperformed the single-source parameter models. The accuracy and robustness of the winter wheat growth stage were effectively improved by the canopy digital image features derived from RGB imagery, with the key meteorological variables, after digital image processing and machine learning techniques. The "multi-source parameter fusion-optimization-probabilistic determination" framework was specifically developed using a Bayesian model. The finding can offer valuable technical support for precision crop management, agricultural meteorological services, and field-level disaster early warning, thereby contributing to the advancement of agricultural intelligence.