[Objective] High-precision, large-scale identification of crop growth stages in irrigation areas has long been one of the core demands for the development of smart agriculture. Based on Sentinel-2 satellite imagery data, this study systematically analyzes the effectiveness of spectral reflectance and vegetation index (VI) curves for identifying different growth stages of winter wheat, and proposes a remote sensing identification method for winter wheat growth stages that integrates multi-spectral features with ensemble learning. [Methods] Using a feature selection + heterogeneous ensemble learning approach, 83 Sentinel-2 Level-2A images were collected, covering all 9 growth stages of winter wheat (emergence, tillering, overwintering, regreening, jointing, booting, heading, grain filling, and maturity). The reflectance patterns of 12 original Sentinel-2 bands and 8 vegetation indices (NDVI, EVI, RECI, NDRE, GCI, LSWI, MASVI, GNDVI) across all growth stages of winter wheat were systematically analyzed. Recursive feature elimination combined with the XGBClassifier was applied to select key feature parameters. A Stacking ensemble architecture was used to heterogeneously integrate five different types of machine learning models—support vector machine (SVM), random forest (RF), extremely randomized trees (ERT), backpropagation neural network (BPNN), and k-nearest neighbors (KNN)—for the identification of winter wheat growth stages. [Results] The 13 feature parameters retained by recursive feature elimination were NDVI, LSWI, GCI, NDRE, EVI, B5, B9, B12, B1, B7, B8a, B11, and B4. Analysis of the selected parameters showed that NDRE, B5, and B7 were all related to the red-edge bands, confirming the unique advantage of red-edge bands in capturing key physiological changes of winter wheat (such as chlorophyll content and leaf structure). Furthermore, the overall importance of vegetation indices was significantly higher than that of original spectral bands, highlighting that vegetation indices derived from multi-band combinations could more effectively characterize changes in crop physiological status and reduce background interference. After accuracy validation, it was found that all six models achieved relatively high remote sensing classification accuracy for the growth stages of winter wheat, with overall accuracy exceeding 0.907 5, and Kappa coefficient and F1-Score above 0.891 6. Additionally, significant changes in spectral reflectance and vegetation indices were observed in the winter wheat canopy during certain growth stages, providing a crucial basis for distinguishing key growth stages. [Conclusions] (1) Significant differences are observed in the spectral reflectance and vegetation index values across different growth stages of winter wheat. Furthermore, the spectral reflectance curve of the winter wheat canopy shows completely opposite trends in the visible and near-infrared bands. The vegetation index curves generally exhibit consistent trends throughout all growth stages of winter wheat, but numerical differences between the curves are significant during specific stages. (2) Red-edge bands can effectively improve the accuracy of identification models for winter wheat growth stages. Compared to using spectral reflectance alone, vegetation indices are more effective in identifying different growth stages. (3) The identification model constructed using the Stacking ensemble achieves significantly higher accuracy than the other models and is suitable for research on growth stage identification. The overall model accuracy ranks as follows: stacking ensemble model>random forest model>extremely randomized trees model>k-nearest neighbors model>backpropagation neural network model>support vector machine model.