A Terrain Feature Guided-Diffusion Model for Void Filling of Digital Elevation Models

The scientific research and engineering application of digital elevation models (DEM) is affected by data voids. Spatial interpolation methods and the generative adversarial network (GAN)-based methods are widely used in void filling. However, they suffer from accuracy degradation, artifacts, and the failure to reconstruct complex terrain features. To address these deficiencies, a terrain feature-guided diffusion model (TFDM) is proposed to fill the DEM data voids. To the best of our knowledge, this is the first work where a diffusion model has been applied to DEM void filling. The TFDM is characterized by the generation of seamless DEM surfaces and stable terrain contours in response to terrain conditions. The TFDM outperforms common methods in filling voids according to visual inspection and quantitative comparison and improves elevation accuracy with low bias in regions of varying gradients.