Assessing the Potential of Neural Radiance Fields for UAV-Based DSM Generation: A Preliminary Comparison with Photogrammetry Farhan Ardianzaf Putra (a*), Cheng-Hsin Li (b), Chao-Hung Lin (c), Jiann-Yeou Rau (c)
a) PhD Student, Department of Geomatics, National Cheng Kung University, Taiwan
*farhanardianzaf[at]gmail.com
b) Master Student, Department of Geomatics, National Cheng Kung University, Taiwan
c) Professor, Department of Geomatics, National Cheng Kung University, Taiwan
Abstract
Digital Surface Models (DSMs) are fundamental products in geospatial analysis, urban planning, and environmental monitoring. Traditionally, DSMs are derived from aerial imagery through photogrammetry, which performs effectively in well-textured urban areas by reconstructing dense and accurate 3D surfaces. However, photogrammetric methods often encounter challenges in areas with reflective materials, dense vegetation, or low-texture surfaces, where feature detection and matching become unreliable. Recent advances in deep learning-based 3D reconstruction, particularly Neural Radiance Fields (NeRF) and its variants, offer data-driven alternatives capable of learning volumetric scene representations from sparse UAV imagery. This study explores the feasibility of generating DSMs directly from UAV images using several NeRF-based models and compares these preliminary results against those produced by conventional photogrammetric workflows. Initial findings suggest that while photogrammetry currently achieves higher geometric accuracy and completeness, certain NeRF-based models demonstrate promising potential in retaining surface detail and structural coherence, particularly in complex scenes. Notably, one model based on an improved NeRF architecture produced denser point clouds and finer representation of building edges, indicating a pathway for further enhancement. Although these results remain preliminary, they highlight the potential for deep learning-based methods to complement or augment traditional photogrammetric techniques, especially under limited data conditions or where traditional methods face constraints. Continued research and optimisation are expected to narrow the performance gap, offering more robust and efficient solutions for DSM generation from UAV imagery.
Keywords: 3D Reconstruction, Digital Surface Model, Neural Radiance Fields, Photogrammetry, UAV Imagery
Topic: Topic C: Emerging Technologies in Remote Sensing
