An Improved Land Use Classification Method Based on DeepLab V3+ Under GauGAN Data Enhancement

Land use is a reflection of human activities in surface space, and classifying it helps better understand the relationship between human activities and the spatial environment. However, the imbalance in land use datasets acquired through remote sensing images has become a major obstacle to improving the accuracy of land use classification. To maintain the balance in the samples of the land use dataset and improve the accuracy of land use classification, this article proposes an improved model based on the DeepLab V3+ network under the GauGAN data enhancement strategy. First, regarding the data imbalance problem, this article proposes an attention optimization mechanism to enhance the learning ability of the generator of GauGAN for contextual semantic information, and adds spectral normalization to the discriminator to induce stable model training. Thus, the model can synthesize excellent small-sample feature data. For the land use classification model, this article improves the DeepLab V3+ network by modifying the expansion rate of the ASPP module and adding the proposed feature fusion module to enhance the ability of the model for combining high- and low-level semantic features. Finally, this article implements both of the proposed improvements to achieve high-precision land use classification. The results showed the following. The land use data synthesized by improved GauGAN contained more complex semantic information and detailed features than the synthetic results of other models, and thus better represented the features of land use. The improved DeepLab V3+ model outperformed the U-Net, FPN, DeepLab V3+, MANet, and TransUNet.