Building Footprint Reconstruction from Satellite Imagery Using a Deep Learning Framework with Geometric Regularization

Since the launch of Landsat-1 in 1972, Earth observation satellites have significantly evolved, now capturing vast amounts of high-resolution imagery. These satellites continuously transmit data that helps us monitor urban expansion, plan infrastructure, and respond to natural disasters. Satellite imaging plays a crucial role in creating accurate spatial maps, offering detailed insights into the built environment. However, reconstructing maps from satellite images is a complex challenge. The field of computer vision has made impressive progress in object recognition and representation, but accurately modeling buildings as geometric primitives remains difficult. Unlike traditional pixel-based approaches, primitive-based object representation requires understanding the spatial structure and relationships that existing methods often struggle with. In this thesis, we explore four different network architectures aimed at improving our baseline model, R-PolyGCN, by integrating novel modules that enhance building footprint reconstruction. The final chapter presents Decoupled-PolyGCN, our most advanced deep learning model, that leverages Graph Convolution Network (GCN) to enhance building footprint reconstruction. By incorporating geometric regularity, multi-scale feature fusion, and Attraction Field Maps (AFM), the model generates more structured and precise building outlines from a single satellite image. Evaluations on the Wuhan University (WHU) and SpaceNet-2 datasets show that Decoupled-PolyGCN outperforms existing approaches, improving Average Precision (AP) by Average Recall (AR)% and AR by 10%. These improvements enable more accurate and reliable mapping, benefiting applications in urban planning, disaster management, and large-scale spatial analysis.