Atlas of Urban Change via AI

Our cities are changing profoundly, experiencing increasing urbanization and more frequent extreme weather events. To understand how these drivers impact our cities, we need precise tools to measure and track urban change over time. However, existing census and survey data have constraints in spatial and temporal granularity, failing to capture real-time physical changes, especially at the building level. In this dissertation, we developed a framework to construct a fine-grained, scalable, and dynamic atlas of urban change based on street view and satellite imagery.

First, we introduce an artificial intelligence-based mapping toolbox, including (i) Siamese-based change detection models pretrained on the largest benchmarks to date, (ii) attention-based aggregation methods to pass building-level signals into the neighborhood level, and (iii) encoder-based foundation models to support various mapping tasks at scale. Next, we develop a building-level disaster recovery dataset covering 12 extreme weather events in the US. Our analysis reveals that in the aftermath of extreme weather events, lower-income neighborhoods are less likely to rebuild and do not return to their pre-disaster state. In contrast, higher-income areas rebuild and tend to improve compared to their pre-disaster state, highlighting the increased disparities. Overall, we aim to inform stakeholders, support future decision-making, and contribute to making cities more resilient and sustainable.

Speaker: Tianyuan Huang, Stanford University

Attend in person or online