Feature-Enhanced Deep Learning Network for Digital Elevation Model Super-Resolution
Xiaochuan Ma, Houpu Li, Zhanlong Chen
Abstract
High-resolution digital elevation model (HR DEM) plays an important role in hydrological analysis, cartographic generalization, and national security. As the main high-precision DEM data supplementary method, DEM super-resolution (DEM SR) based on deep learning has been widely studied. However, its accuracy has fallen into a bottleneck at present, which is more prominent in complex regions. The reason for this issue is that existing methods are difficult to capture enough local features from the low-resolution (LR) input data, and part of the global information (Contour information of long-distance features such as rivers and ridges) will also be lost in the network transmission process. To resolve this issue, a novel feature-enhanced deep learning network (FEN) is designed in this paper. The proposed FEN includes a global feature SR (GFSR) module and a local feature SR (LFSR) module. The former provides global information by using an interpolation method (Kriging) including geographical laws (spatial autocorrelation); The latter fully captures the features in the input data by integrating powerful feature extraction modules and then provides sufficient local features for DEM SR tasks. Thus, DEM SR tasks for complex regions can be realized by integrating the results of GFSR and LFSR modules. Extensive experiments show that FEN achieves state-of-the-art performance in DEM SR tasks facing complex regions. Specifically, compared with the existing DEM SR method (TfaSR, SRResNet, Bicubic, SRCNN, and Kriging), the result by FEN is closer to HR DEM and can retain more local DEM features. Meanwhile, the FEN is more than 20% ahead of other DEM SR methods based on deep learning in elevation accuracy.