Towards a digital twin of Bologna: features extraction and semantic classification using LiDAR

The development of an Urban Digital Twin has been become a shared objective for many cities around the world. Airborne LiDAR has been one of the most used technologies to reconstruct the urban environment, but an unified univocal method to analyse its point cloud data has not been developed yet. The techniques differ on selected features, algorithms and input data. Within the bounds of Bologna Digital Twin, the aims of my thesis are: 1) analyzing LiDAR data from a physical-geometrical point of view; 2) to select features that extrapolate useful information in Bologna LiDAR from all the features found in literature; 3) to develop a model to classify point into defined classes, i.e., buildings, cars, grass, rails, roads and trees. It does not have to be the best algorithm or scalable in other contexts, but it must serve as a starting point for future developments. I discovered how much the three-dimensional distribution of point cloud encompasses information about the analyzed objects, therefore spatial features should be included in the inputs of classification algorithm. The regulation of the feature extraction process was conducted and the importance ranking of all the analyzed features was calculated. A random forest model was developed to classify LiDAR points and it achieves 95% of accuracy.