China is a big producer of winter wheat. Obtaining the growth and distribution of winter wheat in a timely and accurate manner can provide a strong basis for China’s agricultural policy and distribution of agricultural products. Complex geometric changes and foreign body phenomena in high-resolution remote sensing images limited the recognition ability of ground objects. The multivariate features of winter wheat were extracted from GF-2. Based on the U-Net model, a deformable full convolutional neural network (DFCNN) model was introduced into the field of automatic interpretation of remote sensing images. In order to improve the ability of the network model to extract geometric features, the idea of deformable convolution was introduced. A trainable two-dimensional offset was added to the front of each convolutional layer in the network to deform the convolution and obtain object-level semantic information. Thus, the expression of winter wheat features with different sizes and spatial distribution was enhanced, and the interference of foreign bodies in high-resolution remote sensing images was eliminated. A deformable convolution module was added to the improved full convolutional neural network model, and the data set was trained and fine-tuned to obtain the optimal network model with an accuracy rate of 98.1% and a time cost of 0.630s. Based on FCNN model, U-Net model and random forest (RF) algorithm, the accuracy of automatic interpretation was 89.3%, 93.9% and 90.0%, respectively. The results showed that the winter wheat based on DFCNN model had the highest accuracy. Moreover, it can express complex geometric change characteristics well and had good generalization ability.