Comparative Analysis of UAV-Derived DEMs for Urban Terrain Mapping

Thesis Abstract

Urban terrain mapping relies on high-fidelity digital elevation models (DEMs) derived from unmanned aerial vehicle (UAV) photogrammetry to support planning, infrastructure management, and hazard assessment. Yet urban environments present complex occlusions, varied surface materials, and vertical geographies that challenge DEM accuracy and comparability across sensor systems and processing workflows. This study addresses the problem of inconsistent elevation representations produced by UAV-derived DEMs and evaluates how methodological choices influence vertical accuracy, terrain representation, and subsequent urban modeling applications. The aim is to compare UAV-derived DEMs generated from different flight configurations, sensor modalities, and processing pipelines to determine best-practice practices for urban terrain mapping. Specific objectives are to (i) quantify vertical accuracy and error structures of DEMs produced from RGB and multispectral UAV surveys at varying ground sampling distances (GSDs) and flight heights, (ii) assess the impact of point cloud densification, aerotriangulation, and dense image matching algorithms on DEM quality, (iii) compare georeferencing performance using ground control points (GCPs) and real-time kinematic (RTK)–enabled platforms, and (iv) examine the influence of urban features such as building façades, bridges, and road surfaces on DEM interpolation. The research adopts a comparative cross-sectional design anchored in object-oriented accuracy assessment and model-based validation. The population comprises urban areas with diverse morphologies from a mid-sized metropolitan region. A stratified sample of 24 urban sites is selected to represent high-rise cores, mixed-use districts, and peri-urban fringes. At each site, three UAV campaigns are conducted (i) a high-resolution RGB survey at 60 m AGL with 2 cm GSD, (ii) a multispectral survey at 120 m AGL with 5 cm GSD, and (iii) a dual-sensor survey integrating RGB and LiDAR-derived point clouds at 80 m AGL with 3 cm GSD where feasible. Data collection employs calibrated UAV platforms with dual-frequency GNSS receivers, tightly coupled RTK where available, and standardized flight planning to ensure consistent overlap and odometry. Ground truth data comprise survey-grade GNSS measurements (3–5 cm vertical accuracy) from 120 reference points per site and 80 arbitrary test checkpoints distributed to capture occluded and complex urban features. DEMs are generated through three processing pipelines (a) structure-from-motion (SfM) photogrammetry with dense image matching, (b) multi-sensor fusion using LiDAR-derived point clouds and photogrammetric surfaces, and (c) a rigorous aerotriangulation-based block adjustment incorporating GCPs. Accuracy analyses employ vertical error statistics (RMSE, ME, standard deviation), structure analysis (skewness/kurtosis of residuals), and robust equivalence testing. Statistical methods include regression analysis to relate DEM error to surface type, height, and shadowing effects; ANOVA to compare DEMs across pipelines; and nonparametric tests where assumptions fail. The study also implements a methodology validation step using urban flood modeling and line-of-sight visibility simulations to assess practical implications of DEM differences for decision-making. Expected findings indicate that DEMs generated with RTK-enabled surveys and LiDAR-assisted fusion exhibit the lowest RMSE (targeting ?5 cm in flat surfaces and ?10 cm on complex façades), while SfM-only pipelines show higher vertical errors in steep or occluded regions. It is anticipated that GSD and GCP density significantly influence vertical accuracy, with diminishing returns beyond 2–3 cm GSD for typical urban planning tasks. The research contributes to knowledge by providing a comprehensive, empirically grounded framework for selecting UAV-derived DEM generation workflows in urban contexts, clarifying the trade-offs between data fidelity, cost, and processing complexity. The theoretical underpinning draws on geographic information science theories of error propagation and data quality, along with urban morphology considerations. The main conclusion is that a hybrid workflow combining high-precision RTK-enabled data capture, LiDAR-assisted fusion, and targeted GCP deployment yields the most reliable urban DEMs for terrain mapping and subsequent applications such as 3D city modeling and flood risk assessment. Recommendations include adopting standardized QC protocols, reporting detailed uncertainty metrics alongside DEM products, and developing adaptive processing schemes that adjust to surface complexity and occlusion prevalence. The study also identifies avenues for future work, including automation of fusion routines and exploration of semantic segmentation to improve terrain classification in DEM generation.