Graph convolutional networks (GCNs) have recently received increasing attention in hyperspectral image (HSI) classification, benefiting from their superiority in conducting shape adaptive convolutions on arbitrary non-Euclidean structure data. However, the performance of GCN heavily depends on the quality of the initial graph. Conventional GCN-based methods only adopt spectral–spatial similarity to build the initial graph without extracting other contextual information from neighboring nodes. In addition, most GCN-based methods use shallow layers, which cannot extract deep discriminative features from HSIs under the limited number of training samples. To solve these issues, we propose a superpixel feature learning via offset graph U-Net for HSI classification, which can learn deep discriminative features from HSIs. Multiple strategies of measuring similarity among superpixels are utilized to build the initial graph, including spectral information, spatial information, and context-aware information among nodes, making the initial graph more accurate. Furthermore, the graph U-Net structure, containing the graph pooling layer and the graph unpooling layer, is helpful in constructing deep GCN layers and learning multiscale features, which can alleviate the oversmoothing problem. Moreover, an offset module is introduced to emphasize the local spectral–spatial information. Finally, we comprehensively evaluate the proposed method on three public datasets. The experimental results demonstrate the superiority of the proposed approach compared with other state-of-the-art methods.