Deep Learning Approaches to DEM Generation: Attention Mechanisms and Gradient-Based Loss Functions

Abstract

[Abstract]: Digital Elevation Models (DEMs) are essential for a wide range of geospatial applications, including urban planning, environmental monitoring, and disaster management. Conventional methods for DEM generation, such as LiDAR and radar, are accurate but costly and constrained by regulatory and logistical challenges. This study explores the use of deep learning techniques for DEM generation from single satellite images, focusing on the integration of advanced attention mechanisms and novel normalization strategies. A custom dataset, featuring Sentinel‑2 and Landsat 9 imagery paired with high-resolution DEMs of the Iberian Peninsula, was developed and made publicly available to support transparency and reproducibility. Two normalization approaches were evaluated: Global normalization, which preserves global elevation relationships, and Global normalization with Shift, which emphasizes local terrain features. Additionally, state-of-the-art architectures, including U‑Net, Pix2Pix, and DRPAN, were adapted with attention mechanisms such as the Global Attention Mechanism (GAM) and gradient-based loss terms. The results indicate that the proposed GAM, when combined with Shift normalization, produces Mean Absolute Error (MAE) values of 65 m when evaluated under the corresponding local terrain reconstruction objective. While the method does not match the accuracy of traditional sensors, it offers a promising, scalable, and cost-effective alternative for DEM generation in contexts with limited data or resources, where relative elevation accuracy is sufficient.