The traditional convolutional neural network (CNN) model has a large amount of spatial feature information redundancy in remote sensing scene classification, which greatly affects the classification accuracy and operational efficiency of the model. In view of this problem, the paper proposes a DCNN model based on octave convolutional(OctConv). Firstly, the feature map outputted by the convolutional layer is decomposed into two parts according to the frequency, and using global average pooling to compress the lowfrequency part with less feature mapping information into a quarter of the current size, then using OctConv to replace the traditional convolution operation, to achieve highlow frequency feature selfrenewal and information interaction, finally, introducing transfer learning to improve the robustness of the model and making up for the lack of data. The experimental results show that the proposed method can achieve 9925% classification accuracy under the UC_merced_Land_Use public data set, which is 2% higher than the same type method, which shows the superiority and effectiveness of the method.