BUILDING TYPOLOGY EXTRACTION EMPLOYING MULTISOURCE REMOTE SENSING DATA AND DEEP LEARNING

Abstract

Building typology refers to the classification of buildings based on shared characteristics, which can be both qualitative, such as architectural style or functionality, and quantitative, such as height and number of floors. This classification is crucial for various applications, including urban planning, disaster management, governance, and the development of smart cities. Traditional methods for extracting building typology are often manual, resource-intensive, and limited in scalability. This dissertation addresses these challenges by employing multisource remote sensing data and advanced deep learning techniques, specifically Conditional Generative Adversarial Networks (cGANs) and Vision Transformers (ViTs). The initial approach using cGANs was challenging and didn’t produce expected results. However, optimizing the input patches and incorporating self-attention mechanisms in the generator network significantly improved performance, demonstrating the model's ability to perform human-like decision-making in distinguishing buildings from surrounding elements. For quantitative typology extraction, monocular depth estimation techniques, typically designed for mobile video data, were adapted for aerial and satellite imagery. This adaptation presented challenges due to the model's initial poor performance, requiring careful input data processing. Results improved substantially when the tallest buildings were centered in the input patches, leveraging the model's ability to better estimate depth. Additionally, errors arising from the spatial resolution of Digital Elevation Models (DEMs) and tree cover were identified as major sources of inaccuracies in height estimation. The research demonstrated the integration of qualitative and quantitative typology extraction into a 3D model with embedded typological characteristics, providing a comprehensive solution for urban analysis. This approach holds potential for improving smart city initiatives in India by offering detailed and scalable methods for building typology extraction, contributing to more informed urban planning and policy-making. Ultimately, this work highlights the potential of deep learning techniques in overcoming traditional limitations, offering scalable and accurate solutions for building typology extraction using remote sensing data.