The complex terrain and variable climatic conditions in mountainous regions often cause cloud and fog occlusions, which hinder the generation of high-resolution Digital Elevation Models (DEMs) (below 10 m) using optical satellite photogrammetry. To address the issues of insufficient accuracy and data gaps in cloud-affected DEM areas, this study proposes a Time-Conditioned Deformable Convolution Attention Super-Resolution Network (TIDA-SR). The network performs super-resolution reconstruction using open-source DEMs and integrates multiple DEM sources through a feathered blending strategy to achieve high-accuracy results. Its architecture incorporates a diffusion process, deformable convolutions, and a Convolutional Block Attention Module (CBAM), combined with a composite loss function, to enhance the recovery of complex terrain details. TIDA-SR network, combined with the cloud-affected 5 m DEMs generated from GF-7 stereo imagery and open-source 30 m DEMs, is employed to reconstruct and fuse cloud-affected regions with 5 m resolution. The experimental results in the Loess Plateau and the Rocky Mountains demonstrate that, compared with traditional interpolation methods and existing deep learning approaches, TIDA-SR reduces RMSE and MAE by approximately 5%–78% on the validation dataset and by 3%–25% on the open-source DEM dataset. Slope accuracy improvements of approximately 3%–42% on the validation dataset and 3%–8% on the open-source DEM dataset are observed. The feathered blending strategy effectively mitigates stitching artifacts between cloud and noncloud areas, enhancing overall spatial continuity. TIDA-SR exhibits superior performance in high-resolution DEM reconstruction for cloud-affected mountainous regions and shows strong potential for practical applications, including surface process simulations, mountain hydrological modeling, geomorphological analysis, and other terrain-driven geoscience tasks.
扫码关注我们
求助内容:
应助结果提醒方式:
