Deep Learning (DL) has brought revolutionary advancements in Hyperspectral Image (HSI) classification. However, Its accuracy depends on the diversity and labeling of samples. Unfortunately, acquiring labeled samples can be costly, time-consuming, and sometimes impractical. In this paper, we propose few-shot learning for HSI classification based on hybrid attention network (HDFSL). It combines a hybrid attention mechanism with a spectral-spatial fusion network to further enhance feature extraction. Specifically, spectral information is acquired using a pure one-dimensional operator residual network, while spatial information is extracted using a depthwise separable convolutional (DSC) network with a hybrid attention mechanism. The hybrid attention mechanism can effectively balance the utilization of global and local information while improving model performance. It includes both local attention mechanism and global attention mechanism. The local attention mechanism enhances or diminishes features at each spatial position by learning weight allocations, enabling the model to extract information more selectively. The global attention mechanism can help the network more effectively utilize limited sample information, automatically learning the importance of different parts of the data during training, and allocating more attention to key samples. Additionally, by employing deep separable convolutional networks, we are able to achieve efficient computation while under resource constraints. The experimental results on two publicly available HSI datasets indicate that our proposed method achieves state-of-the-art results in HSI classification.