NATIONAL PHD COURSE IN EARTH OBSERVATION

Thesis title: Artificial intelligence for human settlement characterization using Earth Observation

Urban environments are among the most dynamic and complex landscapes on Earth, continuously evolving under the influence of human activity, population growth, and environmental pressures. Monitoring these changes is essential for effective urban planning, disaster management, and sustainable development. A fundamental component of urban analysis lies in the accurate extraction of building footprints and the estimation of building heights, which together provide the geometric foundation for modeling urban structure and growth. In this context, the increasing availability of Very High Resolution (VHR) Synthetic Aperture Radar (SAR) data has opened new opportunities for deriving both two- and three-dimensional (2D/3D) urban metrics, offering unique advantages in terms of all-weather and day-night imaging capabilities, as well as sensitivity to surface geometry and roughness. This research presents an integrated framework for 2D building footprint extraction and 3D building height estimation using single-acquisition VHR SAR images. The work is developed through two complementary phases. In the first phase, a U-Net-based convolutional network is proposed for footprint segmentation using COSMO-SkyMed (CSK) imagery. In the second phase, two object-based deep learning methodologies are introduced to estimate the height of buildings using the same VHR SAR source. The overarching goal is to design a scalable and computationally efficient system capable of learning robust representations of urban structures from a single SAR acquisition, minimizing reliance on multi-temporal or multimodal datasets while maintaining competitive accuracy. The first stage of the study was implemented using CSK HIMAGE (HI) data in StripMap (SM) mode, which provides a spatial resolution of approximately 2.5 m. The city of Milan and the Piana del Sele region were selected as the study areas, encompassing roughly central and suburban regions characterized by both historic and modern architectural typologies. The combination of complex geometry, dense building clusters, and heterogeneous scattering made these sites ideal for evaluating SAR-based footprint extraction. The second part of the research addresses the estimation of building heights using two different object-based deep learning approaches, ResNet-101 and Masked Autoencoder. Unlike pixel-based regression or height-from-shadow methods, the proposed frameworks treat each building as an independent entity, integrating SAR backscatter features and footprint-derived geometric attributes. This study demonstrates the feasibility of using single VHR SAR images for both 2D footprint segmentation and 3D building height estimation in complex urban environments. The U-Net segmentation framework effectively identifies building extents from COSMO-SkyMed imagery, while the ResNet-based and Masked Autoencoder-based models successfully infers height using integrated geometric and radiometric cues.