Decoding urban complexity: Deep learning-based terrain-specific building segmentation for Indian cities

Abstract

Accurate building segmentation from satellite imagery is essential for urban planning, disaster management, and environmental monitoring. This paper presents a novel approach utilizing the UNET architecture for deep learning-based building segmentation, focusing on diverse terrains in Indian cities. Indian cities are uniquely complex in their urban complexity because of a highly densely packed urban landscape and patterns of buildings, thus making segmentation a challenging task. Our method includes meticulously performed data preprocessing and exhaustive validation to achieve high accuracy and adaptability in our trained terrain-based model. Very high-resolution satellite imagery with a 0.5-m spatial resolution was utilized for model training. Specialized models were developed for different terrain types—urban, coastal, and hilly—resulting in significant improvements in segmentation performance compared to generalist models. We reduce human effort and increase efficiency as the proposed system automates segmentation. Such research will, therefore, scale the solution very well in building segmentation. Its application will be practical to aspects of urban planning and disaster response while developing the smart city, and further work will be oriented towards expanding the dataset and generalizing and further developing the model.