Abstract: In view of the difficulty in accurately distinguishing between “same spectrum foreign objects” and “same object foreign spectrum” in hyperspectral images caused by traditional machine learning methods, and the time-consuming processing of high-dimensional remote sensing data by deep learning models, 19 land cover types（mainly agroforestry vegetation） in Matiwan village, Xiong’an New Area, Hebei Province, were taken as the research object, and a hyperspectral classification method for agroforestry vegetation based on improved U-Net was proposed. This method was based on U-Net framework. Firstly, Principal Component Analysis（PCA） was used to extract the main spectral bands to reduce spectral redundancy. Then, the feature extraction module was proposed, which used the depthwise separable convolution to replace the traditional convolution in U-Net to extract multi-scale features of hyperspectral images to reduce the complexity of the network, and uses the better nonlinear h-swish（hard-swish） activation function to improve the generalization performance of the network. Finally, residual connection was introduced into each feature extraction module to extract deep semantic information and improve the classification accuracy. The results showed that the overall accuracy of the improved U-Net for 19 cover classification was 96.68%, which was 4.47%, 2.92% and 0.45% higher than that of Mobile-UNet, U-Net and Res-UNet, respectively, and the training time was 23.5% less than that of Res-UNet with similar classification accuracy. Therefore, residual connection improved the accuracy of network classification and made the model perform well in describing vegetation edges and details. Depthwise separable convolution made the model lightweight and reduced the training time. The improved U-Net model can accurately and rapidly distinguish agroforestry vegetation in the study area.