Bina bölütlemesi ve yükseklik tahmini için görsel durum-uzayı tabanlı çoklu görevli öğrenme

Accurate building segmentation and height estimation from monocular satellite imagery are critical for urban planning, telecommunications, and disaster management. However, conventional deep learning models often struggle with generalization, class imbalance, and underrepresentation of tall buildings. This thesis introduces BuildMamba, a novel multi-task learning framework designed to jointly perform building segmentation and height estimation using only RGB satellite imagery. The model incorporates a visual state-space backbone VMamba for efficient long-range dependency modeling, complemented by a convolutional path for capturing fine-grained local features. To enhance spatial representation, the proposed Mamba Attention Module (MAM) is introduced. For improved height estimation, a boundary-aware Sobel-based weighting strategy is employed to better predict height values at building edges. Extensive experiments on DFC19, DFC23, and Huawei BHE datasets demonstrate the superior performance and generalization ability of BuildMamba compared to competitive baselines. The results show substantial improvements in both segmentation accuracy and height estimation, particularly in complex high-rise urban environments.