The traditional convolutional neural network model has substantial spatial feature information redundancy exists in the spatial dimension of the feature maps in hyperspectral image classification， and the spectral band data on a single pixel of the hyperspectral image are regarded as a disordered high dimensional vector for data processing， which does not conform to the characteristics of spectral data， which greatly affects the operational efficiency of the model and the performance of classification. In order to address this problem， a hyperspectral images classification method combined three-dimensional Octave convolution with bi-directional recurrent neural network attention network is proposed. Firstly， the 3D Octave convolution is exploited to capture spatial features of the hyperspectral image，and reduce spatial feature redundant information. Secondly， Bi-RNN spectral attention network is applied to regard spectral bands data as an ordred sequence to obtain spectral information of the hyperspectral image. Then， the spatial and spectral feature maps are connected by means of the fully connected layer to achieve features merge. Finally， the results of classification are outputed through softmax. Experimental results demonstrate that the classification accuracy of the method proposed reaches 99.97% and 99.79% in Pavia University and Botswana datasets. Compared with other mainstream methods， the proposed method can fully exploit spectral and spatial feature information， and own more competitive classification performance.