Abstract: High-resolution remote sensing image scene classification is an indispensable part of remote sensing image processing, and existing convolutional neural networks cannot overcome the problems of large intra-class diversity and high inter-class similarity of remote sensing images. This paper proposes a two-stream remote sensing image scene classification model named ConvNeXt-SPCECA-GCN based on improved efficient channel attention (ECA) combined with graph convolutional network. The spatial attention mechanism was introduced on the basis of ECA, so that the network could extract feature information from the spatial and channel dimensions. The long-distance spatial information was extracted through the graph convolutional network, and the additive fusion strategy was used to fuse the local key features and the long-distance space features implement classification. In the experiment, a variety of data augmentation methods were used to train the model, which effectively alleviated the impact of insufficient data on the model and enhanced the generalization ability of the model. Experiments were carried out on the datasets UCMerced Land-Use and AID DataSet, and the average accuracy rates were 99.03% and 96.87%, respectively.