RSPCN: Super-Resolution of Digital Elevation Model Based on Recursive Sub-Pixel Convolutional Neural Networks

The digital elevation model (DEM) is known as one kind of the most significant fundamental geographical data models. theory, method and application DEM are hot research issues in geography, especially geomorphology, hydrology, soil other related fields. In this paper, we improve efficient sub-pixel convolutional neural networks (ESPCN) propose recursive (RSPCN) to generate higher-resolution DEMs (HRDEMs) from low-resolution (LRDEMs). Firstly, structure RSPCN described detail based on recursion theory. This paper explores effects different training datasets, with self-adaptive learning rate Adam algorithm optimizing model. Furthermore, adding-“zero” boundary introduced into a preprocessing method, which improves method’s accuracy convergence. Extensive experiments conducted train till optimality. Finally, comparisons made traditional interpolation methods, such bicubic, nearest-neighbor bilinear methods. results show that our has obvious improvements both robustness further illustrate feasibility deep methods processing area.