Semantic segmentation technology is an important step in the interpretation of remote sensing images. High spatial resolution remote sensing images have clear features. Traditional image segmentation methods cannot fully represent the information in high spatial resolution images and tend to yield unsatisfactory segmentation accuracy. With the rapid development of deep learning, many researchers have tried to use deep learning algorithms for remote sensing image segmentation. This paper uses U-Net for multiclassification and binary classification of Gaofen-2 high spatial resolution remote sensing image data. Six types of features, which were build-up, farmland, water, meadow, forest and others, were labeled in the image. A “neighborhood voting” method was used to determine the category of uncertain pixels based on spatial heterogeneity and homogeneity. Through U-Net neural network multiclassification, the overall accuracy of the training data is 93.83%; the overall accuracy of the test data is 82.27%; and the test accuracy of the binary classification algorithm is 79.75%. The results show that the two models yield high accuracy and credibility in remote sensing image segmentation.