Recently, deep learning has achieved considerable results in hyperspectral image (HSI) classification. However, when training image classification models, existing deep networks require sufficient samples, which is expensive and inefficient in practical tasks. In this article, a novel Combining Spatial-spectral Features for Hyperspectral Image Few-shot Classification (CSFF) framework is proposed, attempting to accomplish the fine-grained classification with only a few labeled samples and train it with meta-learning ideas. Specifically, firstly, the spatial attention (SPA) and spectral query (SPQ) modules are introduced to overcome the constraint of the convolution kernel and consider the information between long-distance location (non-local) samples to reduce the uncertainty of classes. Secondly, the framework is trained by episodes to learn a metric space, and the task-based few-shot learning (FSL) strategy allows the model to continuously enhance the learning capability. In addition, the designed network not only discovers transferable knowledge in the source domain (SD) but also extracts the discriminative embedding features of the target domain (TD) classes. The proposed method can obtain satisfactory results with a small number of labeled samples. Extensive experimental results on public datasets demonstrate the versatility of CSFF over other state-of-the-art methods.