Super-resolution (SR) is an essential method for generating high-resolution (HR) Digital Elevation Models (DEM) in complex mountainous regions. However, existing methods struggle to cope with complex mountainous terrain, thereby limiting the reliability of DEMs in subsequent scientific research. To address this, we propose the Terrain-Guided Fusion Network for Super-Resolution (TGFSR). Its core is a multi-stage architecture that integrates an Attribute-Guided Iterative Evolution Module (AGEM) to provide feature-space structural guidance inspired by terrain organization, while employing a Kernel Prediction Network (KPN)-based adaptive fusion strategy to incorporate complementary Synthetic Aperture Radar (SAR) cues related to surface structure, thereby jointly improving elevation reconstruction, terrain attribute preservation, and hydrology-related terrain usability. In fourfold SR experiments, where 120 m DEMs are reconstructed into 30 m DEMs, on the Hengduan Mountains dataset, TGFSR demonstrates enhanced accuracy, reducing both Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) by over 45% compared to Bicubic interpolation and by more than 8% compared to representative deep learning models, including TfaSR and DuffNet. The model’s superior performance in restoring crucial terrain attributes, including slope, aspect, and extracted river networks, demonstrates its ability to generate DEM with both high numerical accuracy and geomorphological consistency, providing a robust data foundation for hydrological and geomorphological applications in complex mountainous regions.
Wu et al. (Mon,) studied this question.