DEM super-resolution framework based on deep learning: decomposing terrain trends and residuals

Hongen WangLiyang XiongGuanghui HuHaoyu CaoSijin LiGuoan TangLei Zhoua School of Geography,Nanjing Normal University,Nanjing,People's Republic of Chinab Key Laboratory of Virtual Geographic Environment,Ministry of Education,Nanjing Normal University,Nanjing,People's Republic of Chinac Jiangsu Center for Collaborative Innovation in Geographical Information Resource Development and Application,Nanjing,People's Republic of Chinad School of Geographic and Biologic Information,Nanjing University of Posts and Telecommunications,Nanjing,People's Republic of China

DOI: https://doi.org/10.1080/17538947.2024.2356121

IF: 4.606

2024-05-23

International Journal of Digital Earth

Abstract:Deep learning-based super-resolution is an essential technique for acquiring high-resolution digital elevation models (DEMs) by enhancing the spatial resolution of low-resolution DEMs. However, current deep learning-based approaches for DEM super-resolution lack comprehensiveness in terrain information reconstruction, resulting in the need to strengthen the rationality of terrain representation. Furthermore, the limited adaptability and extension potential of these approaches restrict their practical applicability and scope, hindering further advancement. As a solution, we introduce a broadly scalable detrending-based deep learning (DTDL) spatially explicit framework for DEM super-resolution. The framework aims to improve DEM reconstruction through data processing and augmentation. It employs detrending to distinguish between large-scale terrain trends and small-scale residuals in DEMs, thereby enhancing the neural network's capacity to learn terrain information. We integrate DTDL with classical super-resolution methods (SRCNN, EDSR, and SRGAN) and conduct experiments in the Alps, Himalayas, and Rockies. The experimental results indicate that the fusion of DTDL with deep learning-based methods enhances the accuracy of terrain reconstruction and the rationality of terrain feature representation, demonstrating strong compatibility and robustness.

geography, physical,remote sensing

What problem does this paper attempt to address?

The problem that this paper attempts to address is the inadequacy of existing deep learning-based Digital Elevation Model (DEM) super-resolution methods in terrain information reconstruction. Specifically:

1. **Incomplete Terrain Information Reconstruction**: Current methods lack a reasonable representation of terrain features, resulting in inaccurate reconstruction outcomes.
2. **Limited Adaptability and Scalability**: Existing methods are complex and difficult to transfer, limiting their practical application scope.
3. **Loss of Detail Information**: Current methods mainly focus on prominent terrain features (such as valleys and ridges) while neglecting finer terrain details (such as surface textures, slope gullies, small protrusions, and terrain edges), which are crucial for precise terrain analysis.

To address these issues, the authors propose a detrended deep learning (DTDL) spatial explicit framework for DEM super-resolution reconstruction. This framework enhances the neural network's ability to learn terrain information by decomposing the DEM into large-scale terrain trends and small-scale terrain residuals through data processing and enhancement techniques. Experimental results show that when combined with classical super-resolution methods (such as SRCNN, EDSR, and SRGAN), DTDL can significantly improve the accuracy and reasonableness of terrain reconstruction, demonstrating strong compatibility and robustness.