Development of global diffusion filling model for voids filling of Digital elevation models (DEMs)

A Digital Elevation Model (DEM) is crucial in various applications, but the presence of void regions severely limits its application value. This paper proposes a novel Global Diffusion Filling Model (GDFM) that integrates a multi-attention mechanism into the U-Net architecture and employs a novel hybrid loss function to enhance the model’s ability through learning and recovering terrain features, while the resampling strategy generates realistic DEM samples with accurate geographic features and fine details. The developed GDFM operates independently without requiring external data or predefined masks, making it suitable for both terrestrial and underwater DEM void filling. To validate the proposed model, experiments were conducted on three terrestrial DEM datasets and three sets of underwater DEM datasets using two different types of masks. The results show that GDFM achieves 80%-90% structural similarity in the filled void regions, outperforming traditional methods by 24% and outperforming deep learning models by 14%, demonstrating superior performance on terrestrial DEM datasets. For underwater DEM, GDFM achieved an average improvement of 2%-10% compared to existing methods. The DEM reconstructed by GDFM can retain the characteristics of complex terrain with extremely high accuracy, providing a solid theoretical basis for global high-precision DEM reconstruction.