Objectives  In recent years, hyperspectral images classification based on deep learning has made important progress. In view of the scarcity of training samples for hyperspectral image classification, this paper proposes a lightweight attention depth-wise relation network (LWAD-RN) to solve the problem of small sample hyperspectral image classification.

  Methods  The network consists of an embedding layer and a relation layer. In the embedding layer, a lightweight convolutional neural network combining attention mechanism is used to extract pixel features, and a dense network structure is introduced. The relation value is calculated in the relation layer for classification, and the task-based mode is used to train the network. Three groups of public hyperspectral image datasets are used to implement experiments.

  Results and Conclusions  The results show that LWAD-RN can effectively improve the classification accuracy under the condition of small samples (5 training samples per category), and the efficiency of model training and classification is improved.The proposed LWAD-RN can obtain ideal classification accuracy under the condition of small samples, and the lightweight network structure can improve the model training and classification efficiency. However, under the condition of small samples, the quality of training samples will have an important impact on the performance of the model. Therefore, follow-up studies should be conducted on how to select high-quality training samples more accurately and efficiently to ensure the stability of the model and better meet the needs of practical application.