CitySurfaces: City-scale semantic segmentation of sidewalk materials

While designing sustainable and resilient urban built environment is increasingly promoted around the world, significant data gaps have made research on pressing sustainability issues challenging to carry out. Pavements are known strong economic environmental impacts; however, most cities lack a spatial catalog of their surfaces due cost-prohibitive time-consuming nature collection. Recent advancements in computer vision, together with availability street-level images, provide new opportunities for extract large-scale lower implementation costs higher accuracy. In this paper, we propose CitySurfaces, an active learning-based framework that leverages vision techniques classifying sidewalk materials using widely available images. We trained images from New York City Boston evaluation results show 90.5% mIoU score. Furthermore, evaluated six different cities, demonstrating it can be applied regions distinct fabrics, even outside domain training data. CitySurfaces researchers city agencies low-cost, accurate, extensible method collect material which plays critical role addressing major issues, including climate change surface water management.