Extracting Cadastral-Scale Building Footprints from UAV-LiDAR Point Clouds in Rural Environments

Extracting building footprints from UAV-LiDAR point clouds is vital for 2D/2.5D urban and rural mapping. We propose an end-to-end workflow integrating point-level classification, geometric instance extraction, concave hull vectorization, and independent cadastral validation. The approach was evaluated on a high-density (~3 cm spacing) dataset of 12 million points acquired over Pınarkaya, a rural settlement in Yozgat, Türkiye, utilizing 12% (1.45 million points) for training.

Points were classified into four categories (building, tree, wall, wire) using XGBoost. To prevent spatial data leakage, a 15×15 m spatial block cross-validation scheme across five folds was implemented, yielding an average cross-validation macro-F1 of 0.9675 (OA = 0.9869). On the regional validation set, the model achieved a macro-F1 of 0.9045. Walls exhibited the lowest performance (F1 = 0.69), with error rates reaching 46–88% below a 2 m height compared to 1.5–10% for buildings, reflecting geometric complexity and training sample constraints.

Classified building points were decomposed into roof and facade components to extract 22 building instances, vectorized using concave hull sectioning, and validated against 20 cadastral ground-truth buildings. Object-level matching yielded 16 true positives, 4 false negatives, and 6 false positives (completeness = 0.80, correctness = 0.73, F1 = 0.76). Matched pairs achieved a mean Intersection over Union (IoU) of 0.81, a median centroid offset of 0.31 m, and a pixel-level overall accuracy of 0.97 (F1 = 0.81). These findings demonstrate that high point-level classification accuracy reliably enables cadastral-scale footprint extraction in complex rural settings.