Automatic Refinement of 3D Building Models from LOD1.3 to LOD2.2 : Parametric Roof Reconstruction with Machine Learning, without Point Clouds

Sammanfattning

Background. Three-dimensional city models support applications such as solar analysis, flood simulation, and energy modelling, but their usefulness depends on geometric detail. While most national datasets provide buildings at LOD1.x, higher-detail LOD2.2 models remain scarce and are typically generated using LiDAR data. Objectives. This thesis investigates whether LOD1.3 building models can be upgraded to LOD2.2 using only information available in the lower-detail representation, without LiDAR or other external data. Methods. A hybrid pipeline was developed combining Random Forest roof-type classification, regression-based roof-parameter prediction, and deterministic geometric reconstruction. The approach was evaluated on 41 tiles from the Dutch 3DBAG dataset using cross-validation and held-out tile testing. The study was restricted to rectangular buildings with flat, gabled, or hipped roofs. Results. Overall classification accuracy reached 75.31% under cross-validation and 68.10% on held-out tiles. This figure is driven almost entirely by the flat and gabled classes (roughly 68–80% accuracy); the hipped class performed poorly, reaching only 9.17% under cross-validation and 5.52% on held-out tiles, because gabled and hipped roofs are frequently indistinguishable from LOD1.3 geometry alone. Ridge-height prediction achieved mean absolute errors below 0.31 m, with over 92% of buildings reconstructed within 1 m ridge-height error and over 93% within 20% relative volume error. Conclusions. The results demonstrate that LiDAR-free upgrading from LOD1.3 to LOD2.2 is feasible for rectangular buildings with flat, gabled, and hipped roofs. The main limitation is distinguishing between gabled and hipped roofs from LOD1.3 geometry alone. When roof type is correctly identified, the proposed method produces practical LOD2.2 reconstructions suitable for regions lacking LiDAR coverage.