The detailed area and spatial distribution of irrigated and rainfed wheat can help forecast wheat yield and study water use efficiency. However, the similar spectral characteristics of irrigated and rainfed wheat make it difficult to separate them with low-spatial resolution or several high-spatial resolution images on the high heterogeneity of the southern Loess Plateau. To solve this challenge, this study used the Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM) and Enhanced STARFM (ESTARFM) to generate time series of the normalized difference vegetation index (NDVI) and the normalized difference water index (NDWI) at a 30-m resolution by fusing Moderate Resolution Imaging Spectroradiometer and Landsat data. Then, the phenological feature extracted from the predicted NDVI is combined with an auxiliary dataset to classify irrigated and rainfed wheat using the support vector machine classifier. An overall classification accuracy of 93.7% and a Kappa coefficient of 0.91 are achieved. Compared with corresponding high-resolution Google Earth images, the spatial distribution of the classification was consistent with actual land cover. This study demonstrates that the classification approach could classify irrigated and rainfed wheat in high heterogeneity regions and crops with smaller spectral characteristic differences. Moreover, it could be implemented across larger geographic regions.