Due to the significant differences in the number of bands, spectral range and spatial resolution of different hyperspectral image datasets, the optimal network structures applicable to different hyperspectral image datasets also differ. In addition, manually designed deep learning networks need to tune a large number of hyperparameters, which undoubtedly poses a serious challenge to designing a generalized classification model applicable to various HSI datasets. Therefore, an efficient attention neural architecture search algorithm is proposed to realize the automatic design of deep learning networks. Firstly, in order to construct an efficient search process, a model is constructed based on the search of microable network architecture, which can effectively improve the search speed of hyperparametric networks. Then, in order to achieve high-precision classification results, a novel modular search space is designed. Finally, considering the misclassification problem of small samples in hyperspectral datasets, Poly loss function is used to increase the loss weights of a few categories, so as to improve the model’s ability to recognize these categories. Experimental results on publicly available hyperspectral datasets show that the overall classification accuracy of the proposed method reaches 99.50% and 97.81%, respectively. The proposed method explores the application of neural architecture search in hyperspectral classification tasks, improving classification accuracy and algorithm design efficiency.