Urban LiDAR-based Building Footprint Extraction and Validation Methodology

Thesis Abstract

Urban LiDAR data have transformed the ability to delineate accurate building footprints in dense urban environments, yet challenges persist in reliably extracting complex roof geometries and validating footprints across heterogeneous urban morphologies. This study addresses the gap between high-resolution LiDAR point clouds and conventional cadastral delineations by developing a robust, scalable methodology for automatic building footprint extraction and multi-level validation. The aim is to produce precise footprints suitable for cadastral, urban planning, and disaster risk assessment applications, while quantifying uncertainty associated with extraction and validation processes. Specific objectives are (i) to design a structured workflow that integrates LiDAR-derived features, machine learning classification, and rule-based post-processing to generate 3D building footprints; (ii) to implement a multi-stage validation framework combining geometric, semantic, and cadastral-consistency checks; (iii) to evaluate performance across a representative urban sample comprising 12 neighborhoods with varied typologies in a metropolitan region; (iv) to quantify positional, dimensional, and semantic deviations relative to authoritative reference datasets; and (v) to assess computational efficiency and scalability for large-area applications. The methodology adopts a mixed-methods research design grounded in object-based image analysis (OBIA) and empirical validation. The population includes high-density urban LiDAR surveys (point spacing ~0.25–0.5 m) and corresponding authoritative footprints from a metropolitan cadastral authority. A stratified random sample of 12 neighborhoods (covering commercial, residential, and mixed-use zones) is selected, yielding an extensive dataset of approximately 1,200 parcels and 6,400 candidate footprints. Data collection involves (a) LiDAR acquisition data curated from municipal repositories, (b) ancillary imagery (0.5 m orthoimagery) for feature corroboration, and (c) cadastral footprint references and building height metadata supplied by the land administration office. Instrumentation includes an in-house processing pipeline implemented in Python and cloud-accelerated GIS tools, incorporating (i) ground and roof segmentation via point cloud normalization and TIN-based roofline extraction, (ii) feature engineering to derive height, planimetric area, volume, edge continuity, and roof type indicators, (iii) supervised classification using Random Forest and Gradient Boosting to label building vs non-building features, (iv) rule-based post-processing to enforce topological coherence and cadastral compliance, and (v) a multi-tier validation framework consisting of geometric accuracy ( RMSE/ MAE of footprint boundaries), semantic consistency (roof type and usage plausibility), and cadastral congruence (overlay accuracy with official records). Analytical methods include spatial accuracy assessment, area and perimeter error metrics, and uncertainty propagation analysis. The study employs quantitative analyses such as RMSE and MAE for boundary and area discrepancies, F1-score for building delineation accuracy, and paired t-tests to compare results against reference footprints. Regression analysis examines the relationship between LiDAR feature quality and extraction accuracy, while ANOVA tests assess performance differences across neighborhood typologies. The validation framework yields confusion matrices for footprint presence, and Bayesian uncertainty analysis provides probabilistic estimates of boundary certainty. The theoretical underpinning integrates the Object-Based Image Analysis (OBIA) paradigm with the theory of spatial accuracy and the cadastral integrity framework, drawing on matched theory from urban geometry and geographic information science. A conceptual model illustrates the data flow from LiDAR-derived features to final validated footprints, highlighting feedback loops for quality assurance. Expected findings include (i) high-precision footprints with mean boundary deviation under 0.8 meters and area error under 2%, (ii) improved performance for parcels with regular roof geometries compared to highly irregular roofs, (iii) a validated multi-tier framework achieving an overall cadastral congruence score exceeding 0.90, and (iv) clear quantification of uncertainty applicable to urban planning and risk assessment workflows. The study contributes to knowledge by delivering a replicable, end-to-end methodology for urban footprint extraction and validation that integrates OBIA, machine learning, and cadastral verification, accompanied by a transparent uncertainty framework. It offers practical implications for municipal authorities seeking accurate, scalable, and auditable building footprints for governance, planning, and disaster management. The main conclusion anticipates that the integrated workflow significantly enhances footprint accuracy and cadastral alignment in diverse urban contexts, with recommendations to adopt standardized LiDAR preprocessing protocols, expand reference datasets for anchor validation, and extend the framework to temporal analyses for monitoring urban growth and subsidence.