Advancing Urban Analysis through Joint Rooftop Segmentation and Building Height Estimation

Accurate segmentation of building rooftops holds significant importance within the realm of urban analysis. The extraction of building structures from high-resolution imagery represents a pivotal task that facilitates insights into urban growth, infrastructure planning, and disaster management. However, the sole segmentation of building rooftops falls short of complete urban planning, as a crucial attribute, building heights, remains absent. For this vital task of height estimation, digital surface models (DSMs) are employed which provide valuable height information about the Earth’s surface. Furthermore, the integration of DSMs with various types of remotely sensed data has shown success across a broad spectrum of applications. Within the context of deriving more profound insights for enhanced urban analysis, this research paper presents a comprehensive approach which utilizes Squeeze and Excite technique along with residual blocks to jointly perform binary rooftop segmentation and predicting the height of building structures. The work leveraged the rich dataset made available by the 2023 IEEE GRSS Data Fusion Challenge. The building annotations provided in the dataset are used to form the building binary masks and normalized DSMs. Another aspect of this work involves utilizing RGB satellite imagery of the used dataset to create binary segmentation masks of building rooftops and normalized DSMs. In terms of statistical results, our approach which uses the adaptive recalibration of channel-wise features when compared with other state-of-art algorithms yields an impressive F1 Score of 0.93 for rooftop segmentation and a Zero Normalized Cross-Correlation (ZNCC) score of 0.88 for height estimation on the test dataset. This approach offers a flexible and accurate solution to address the multiple, interconnected joint tasks of building segmentation and its height estimation simultaneously, thereby reducing the training and testing times proportionally.