AI for remote sensing image and LiDAR data classification within Bologna’s digital twin project

This thesis, developed within the framework of the Bologna Urban Digital Twin project, explores the effectiveness of three Deep Learning models for semantic segmentation of 2D remote sensing images and evaluates their adaptability for classifying vegetation in 3D LiDAR point clouds. The primary goal was to assess the potential of Vision Transformer based foundation models, however, the most suitable solution proved to be a Convolutional Neural Network (CNN) based model. Among the models tested, SAMGeo, MTP, and DeepForest, only the latter demonstrated useful to be applied to lower resolution images and to support individual tree crown detection, which is essential for vegetation mapping. DeepForest model was integrated into a custom pipeline to classify LiDAR points, reaching a precision of 71 % and a recall of 67 %, showing promising results despite limitations due to underfitting. All evaluations were conducted without model training or fine tuning, relying on its zero-shot inference.