DEM Super-resolution Reconstruction Method Based on Adaptive Generative Adversarial Network

Objectives: The Digital Elevation Model (DEM) serves as the data foundation for urban modeling, disaster assessment, and 3D visualization. As an input, it can improve the accuracy and reliability of many geographic analysis applications. Methods: To reduce artifacts and noise in the reconstruction of Digital Elevation Models (DEMs) and improve the precision of DEM super-resolution reconstruction, this paper proposes a generative adversarial network (GAN) super-resolution reconstruction model that integrates adaptive convolution and U-Net. (1) The generator's feature extraction module employs a deep residual dense module and an adaptive residual convolution module. The deep residual dense module reduces artifacts and noise generated during SRGAN reconstruction while increasing the receptive field and enhancing the model's feature extraction capability. (2) The adaptive convolution module introduces deformable convolution into the residual network module, enabling the network model to adaptively extract feature information and learn more terrain features. (3) The discriminator incorporates a U-Net structure with an attention mechanism, where U-Net demonstrates good performance and efficient utilization of GPU memory. The attention mechanism captures sufficiently large receptive fields to obtain semantic contextual information, thereby improving the reconstruction accuracy and learning efficiency of the DEM. (4) The proposed model effectively addresses issues such as edge smoothing in DEM reconstruction, the presence of artifacts and noise in reconstructions, and the omission of certain terrain features. Results: The reconstruction accuracy of the DEM is evaluated, and the four datasets of the experiment show a decrease of 0.3%-12.6% in the mean value of RMSE, 3.2%-14.5% in the mean value of MAE, 1.9%-18.3% in the mean value of Slope, and 2.7%- 12.0% in the mean value of Aspect. The reconstruction results are compared with the double-three times and SRGAN methods, and the SRDCGAN method can effectively remove the artifacts and retain more terrain features, and the reconstruction accuracy of the RMSE, MAE, Slope and Aspect values are significantly improved. Conclusions: The results indicate that the SRDCGAN method can establish a mapping model for high-resolution and low-resolution DEM data from a specific area. High-resolution DEMs from other regions can be reconstructed using the mapping model with low-resolution DEMs. SRDCGAN is expected to achieve high practical value in geological applications that utilize high-resolution DEMs.