The Hetao Plain and Xing’an League are the major cultivated areas and main grain-producing areas in Inner Mongolia, and their crop planting structure significantly affects the grain output and economic development in Northern China. Timely and accurate identification, extraction, and analysis of typical crops in Xing’an League and Hetao Plain can provide scientific guidance and decision support for crop planting structure research and food security in ecological barrier areas in Northern China. The pixel samples and the neighborhood information were fused to generate a spectral spatial dataset based on single-phase Sentinel-2 images. Skcnn_Tabnet, a typical crop remote sensing classification model, was built at the pixel scale by adding the channel attention mechanism, and the corn, sunflower, and rice in the Hetao Plain were quickly identified and studied. The results of this study suggest that the model exhibits high crop recognition ability, and the overall accuracy of the three crops is 0.9270, which is 0.1121, 0.1004, and 0.0874 higher than the Deeplabv3+, UNet, and RF methods, respectively. This study confirms the feasibility of the deep learning model in the application research of large-scale crop classification and mapping and provides a technical reference for achieving the automatic national crop census.