A DEM Super-Resolution Reconstruction Network Combining Internal and External Learning

The study of digital elevation model (DEM) super-resolution reconstruction algorithms has solved the problem need for high-resolution DEMs. However, DEM algorithm itself is an inverse problem, and making full use a priori information effective way to solve this problem. In our work, new method proposed based on complementary relationship between internally learned methods externally methods. presence large amount repetitive within DEM. Using internal learning approach learn prior DEM, low-resolution dataset rich in detailed features generated, this, training constrained external network constructed discrepancy data pair. Finally, it introduces residual accelerate operation rate degradation brought about by deepening network. This enables better transfer deeper mappings, which turn ensures accurate information. utilizes specific as well achieves results experimental results. Bicubic method, Super-Resolution Convolutional Neural Networks (SRCNN), very deep convolutional networks (VDSR), ”Zero-Shot” (ZSSR) paper were compared, average RMSE five 8.48 m, 8.30 8.09 7.02 m 6.65 respectively. mean error at same resolution 21.6% than that 19.9% SRCNN, 17.8% VDSR 5.3% ZSSR method.