A Decade of Three-Dimensional Urban Change: Unsupervised Changepoint-Driven Post-Processing for Satellite-Based Building-Height Predictions

This work demonstrates how archives of daylight satellite data can be utilized to inform urban studies about the micro-regional, annual building-stock evolution over longer time spans. We predict building footprint and height on a 5-meter resolution to generate an 11-year panel data set for Beijing. In order to achieve this, we employ a U-Net algorithm on multiple scenes of the same region per year from the RapidEye legacy archive. We introduce a post-processing pipeline to address prediction variation over time with a changepoint detection algorithm, Pruned Exact Linear Time (PELT). The method is straightforward to use and does not require any longitudinal labels that are hard to obtain as input. We also compare our results with well-known data sources such as the Global Human Settlement Layer (GHSL) and show our strengths and contributions in modeling annual building volume change.