A Spatial-Temporal Framework for Integrating UAV, GNSS, and Lidar Data

Thesis Abstract

This study addresses the challenge of achieving consistent, high-accuracy spatial-temporal integration of heterogeneous data from unmanned aerial vehicles (UAV), Global Navigation Satellite Systems (GNSS), and Light Detection and Ranging (Lidar) sensors to support precise 3D mapping and dynamic natural-and-built environment modeling. The aim is to develop a spatial-temporal framework that fuses UAV-imagery, GNSS trajectories, and lidar point clouds within a unified probabilistic and machine-learning-based pipeline to produce temporally coherent geospatial products at sub-decimeter accuracy across varying terrains and acquisition conditions. Specific objectives are (i) to characterize sensor biases, temporal synchronization errors, and georeferencing uncertainties through a controlled calibration ensemble; (ii) to formulate a modular fusion architecture that integrates time-stamped UAV imagery, GNSS trajectories, and multi-echo lidar returns using a Bayesian state-space model augmented by a learned correspondence module; (iii) to implement a time-aligned, geodetic-aware co-registration routine that leverages feature-based matching and probabilistic data association; (iv) to assess framework performance across diverse study areas (urban, forested, and coastal) and phenological stages using a multi-sensor dataset; and (v) to demonstrate practical applicability in engineering, environmental monitoring, and disaster risk assessment by validating generated products against ground truth and independent reference datasets. The methodology adopts a mixed-methods research design combining quantitative sensor fusion experiments with qualitative expert validation. The population comprises calibrated UAV campaigns, GNSS trajectories, and terrestrial lidar scans collected over three pilot sites spanning 2.5 square kilometers, including one urban, one forested, and one coastal environment. A stratified sampling approach yields 60 UAV flight lines, 180 GNSS reference points, and 450 lidar scans, collected over four seasonal windows to capture temporal variability. Data collection instruments include high-resolution RGB and multispectral UAV sensors (2 cm nominal ground sampling distance), dual-frequency GNSS receivers with real-time kinematic (RTK) corrections, and terrestrial/mobile lidar systems with 5 mm vertical accuracy. The analysis employs a hierarchical Bayesian fusion framework to estimate latent state variables representing terrain surface, vegetation structure, and built-form geometry, integrated with a neural-network-based feature correspondence module for lidar-uav image alignment. Spatial statistics are conducted via variogram analysis, while time-series coherence is evaluated with dynamic time warping and Kalman-filter–based smoothing. Model specification embeds a probabilistic sensor model, temporal alignment priors, and a scene-specific priors derived from Open Geospatial Consortium standards. Validation uses independent reference datasets, including terrestrial total station measurements (n=240) and airborne lidar-derived digital elevation models (DEMs) with 0.05 m root-mean-square error across sites. Reliability is assessed through test-retest procedures and cross-validation with kappa statistics for feature correspondences. Ethical considerations address data privacy and security in urban environments, with institutional review board clearance obtained for fieldwork. Expected findings indicate that the proposed spatial-temporal fusion framework reduces georeferencing and alignment errors by an average of 38% relative to baseline single-sensor processing, achieving sub-decimeter to decimeter-level accuracy for DEMs and 3D feature models across all study sites. The approach is anticipated to demonstrate robustness to GNSS outages and variable lidar sampling densities, with the Bayesian component providing probabilistic uncertainty quantification that informs end-users about product reliability. A key contribution lies in the integration of a temporal coherence constraint within the co-registration process, improving consistency of 3D reconstructions over successive epochs, which is critical for change detection and monitoring applications. The theoretical contribution includes operationalizing a hybrid data-association scheme between point clouds and image-derived features within a probabilistic framework, and extending existing theories of multi-sensor fusion by incorporating temporal priors and spatially adaptive weighting informed by scene semantics. The study advances knowledge by delivering a replicable, scalable framework suitable for deployment in municipal planning, environmental management, and post-disaster assessment. It provides publicly shareable datasets and a modular software prototype implementing the fusion pipeline, along with guidelines for sensor configuration, calibration, and quality assurance. Recommendations focus on standardizing temporal metadata schemas, enhancing real-time fusion capabilities for near-live mapping, and extending the framework to incorporate additional modalities such as hyperspectral imagery and synthetic aperture radar. The conclusion emphasizes that spatial-temporal integration of UAV, GNSS, and lidar data, when structured within a probabilistic fusion architecture with semantic-informed priors, yields consistent, high-fidelity geospatial products essential for accurate decision-making in dynamic environments.