**4. Conclusions**

Rottensteiner, F et al. [30] have tried to analyse a few of the grea<sup>t</sup> variety of detection and reconstruction applications from airborne laser scanning (ALS) proposed in the literature by identifying common problems of existing studies and by giving indications about the most promising applications. However, a research demand is still needed for comparing the results of di fferent segmentation methodologies for 3D building reconstruction. Indeed, this study presents an investigation of fully automatic segmentation approaches for 3D building detection and modeling by processing airborne LIDAR point clouds. The first method proposed in this study for the extraction of rooftop patches uses a fuzzy c-means clustering method refined with the separation of planar and coplanar planes, which can be fairly easily accomplished based on planar equations and connectivity, respectively. In a second segmentation approach, a region growing based segmentation combined with RANSAC method was used to detect all significant features on the rooftop. Finally, the boundary regularization approach and the 2.5D dual-contouring method was adopted for the 3D modeling process using the outcome of each of these two segmentation approaches.

The results of both approaches were tested on two case studies that di ffer in their types of urban development and input data characteristics. The (i) benchmark LIDAR point clouds with the related reference data (generated by stereo plotting) over downtown Toronto (Canada) and (ii) the LIDAR data of a complex building in Matera (Italy) with the relative 3D BIM model (Building Information Modelling) (generated though celerimetric survey measurement with a total station) were used to evaluate the geometrical quality of roofs under di fferent operating system of the above described segmentation approaches. Performances were evaluated in terms of computational time but also in terms of mean, standard deviation and Root Mean Square Error of the Euclidean distance (along x, y and z dimension) of each vertex (all the points) of the modeled roof polygons and the nearest neighbors of the corresponding reference point. The results of these two di fferent case studies show that both methods reach good performance metrics in terms of geometry accuracy, demonstrating their transferability in other contexts. However, the approach based on region growing segmentation exhibited slightly better performances than the clustering-based approach and required greater computational time.

**Funding:** This research received no external funding.

**Acknowledgments:** The author gratefully acknowledges GEOCART S.p.A. for providing the LIDAR point clouds of Matera (Italy) and the author would like to acknowledge the provision of the Downtown Toronto data set by Optech Inc., First Base Solutions Inc., GeoICT Lab at York University, and ISPRS WG III/4.

**Conflicts of Interest:** The author declares no conflict of interest.
