DEM Super-Resolution Assisted by Remote Sensing Images Content Feature

Deep learning has achieved promising progress in the super-resolution (SR) of digital elevation models (DEMs) by effectively exploiting the correspondences of terrain features at different spatial scales. However, the existing deep learning based DEM SR methods rarely consider the integration of the available multi-modal data, which can provide auxiliary high-frequency information for resolution enhancement. In this paper, we explore the integration of optical remote sensing images for assisting DEM SR, where the images have a higher resolution than the DEM dataset to be reconstructed. This is achieved by constructing a terrain-guided content feature learning network, where the core idea is to transfer the feature representation of high-resolution image textures with the guidance of informative terrain features, for adapting DEM SR learning. Specifically, a content feature learning module is established to calibrate the content details to the learned terrain feature space, by exploiting the multi-dimensional attention mechanism. Extensive experiments were conducted on simulated datasets. The results indicated that the proposed method can reconstruct a DEM with high-fidelity elevation and sharper terrain details by incorporating the complementary information from satellite images, and outperforms the competitive comparison methods.