Advanced Deep Learning Models in Earth Observation for Urban Applications: Bologna and Turin Case Studies

This thesis investigates the application of deep learning models for building segmentation and footprint extraction using pansharpened satellite imagery, with the aim of supporting or optimizing urban data management and enhancing environmental resource management. Building footprint extraction is critical for identifying the exact spatial extent of structures, which can be used to inform various urban planning and resource management efforts. The study utilized panchromatic and multispectral satellite images from the WorldView-2 platform to assess multiple deep learning architectures, including UNet with attention mechanisms, DHAUNet with hybrid and multi-head attention configurations, and several iterations of DEEPLabV3 with ResNet backbones. Among these, DEEPLabV3 with ResNet-50, tested across different tile sizes (1024, 512, and 256), demonstrated the highest accuracy in extracting building footprints, with optimal results achieved using a 256×256 tile size. The refined model was applied to datasets from two cities - Turin and Bologna - representing distinct urban roof architectures. Additionally, a combined dataset facilitated the comparative analysis of segmentation performance across varying urban environments. The results emphasize the model’s adaptability and precision in extracting building footprints, demonstrating its robustness across diverse architectural contexts. This research contributes to the advancement of deep learning integration with satellite remote sensing, providing valuable insights for future applications in sustainable urban development. Furthermore, the findings underscore the potential for leveraging accurate building footprint data to support renewable energy initiatives, disaster risk management, and improved urban planning practices globally.