Feasibility and Optimization of DEM Super-Resolution Prediction Using Multi-Factor U-Net Model

Deep learning methods have been making significant progress in Digital Elevation Model (DEM) super-resolution (SR), with superior performance achieved compared to traditional spatial interpolation methods. By utilizing deep learning techniques to identify the relationship between remote sensing factors and terrain, DEM super-resolution predictions can be made under different conditions, which holds great significance for landslide susceptibility assessments and geological disaster predictions. This study was conducted at the Qingyuan Jiuan landslide test site in Guangdong, where the UNet model was used for DEM super-resolution prediction. The 0.1-meter resolution DEM obtained by LiDAR on February 28, 2023, was used as the output data, and 11 input factors, including multi-source remote sensing data such as Sentinel-1, Sentinel-2, and differential DEM, were incorporated. A comparison was made between two methods of assigning weights to the 11 influencing factors: equal weighting and empirical preset weighting. The results showed that the preset weighting method was able to predict the DEM more accurately, with the predicted results aligning with the LiDAR-derived DEM data. This research demonstrates the feasibility of using the U-Net model for DEM super-resolution prediction under multi-factor conditions and provides an effective reference for landslide prediction. Additionally, future optimization suggestions, including studies on multi-factor weighting and the integration of low-resolution and high-resolution models, have been proposed to enhance model accuracy and generalization capabilities