Super-resolution for terrain modeling using deep learning in high mountain Asia

High Mountain Asia (HMA) is characterized by some of the most complex and rugged terrain conditions in the world. However, high resolution terrain data are not easy to quickly acquire from the area due to difficulties in accessing the region. In this study, we trained a modified super-resolution residual network (MSRResNet) to develop super-resolution (SR) digital elevation models (DEMs) in the HMA areas by using freely available DEM data from the HMA region and limited high resolution (HR) DEMs from other areas to train the model. In this network, a new loss function was constructed that considered the terrain parameters of slope and curvature to constrain the network learning and convergence. The proposed method was applied to and validated by data from the Hengduan Mountains in the southeastern part of HMA, which is a world-famous longitudinal belt of mountains and canyons. A comparative analysis between the current and existing methods (i.e., SRGAN and Bicubic interpolation) was conducted to assess the effectiveness of the proposed approach. The experimental results were also investigated and evaluated by visual inspection and analysis of the terrain parameters. The results demonstrate that the proposed MSRResNet super-resolution process can achieve highly accurate terrain data by downscaling DEMs in HMA. This SR process also outperforms the other comparable methods. Compared to the Bicubic interpolation method, the RMSE and MAE accuracy are improved by 32.17% and 33.97%, and compared to the SRGAN method, the RMSE and MAE accuracy are improved by 39.15% and 32.47%. The HR DEM generated by the new method is more conducive to improving the accuracy of extracted terrain features, such as stream networks. It is promising to apply this model on other areas on Earth or even other planets with terrain similar to that of HMA.