Quantifying multi-decadal urban growth using Hexagon spy satellite imagery and deep learning building detection across four global cities

Abstract

Datasets capturing building distribution, size, function, and arrangement are essential for creating sustainable and resilient settlements. This is because building patterns directly affect human well-being, environmental conditions, and climate change. Remote sensing excels at accurately mapping building data. However, large-scale analyses often rely on medium-resolution satellite imagery, which lacks building-level detail, and multispectral high-resolution satellite imagery, capable of detecting individual buildings, is limited by the absence of data before 2000 when many world regions experienced rapid urban growth. Here, we evaluated the potential of high-resolution panchromatic Hexagon spy satellite imagery from the 1970s to map urban growth. We employed a Mask R-CNN deep learning model to detect building footprints in Hexagon imagery from 1972 to 1979 across four urban growth hotspots: San Diego County (USA), Madison (USA), Harare (Zimbabwe), and Hyderabad (India). Our model achieved high precision (0.83–0.91) and detected 73–94 % of the total building area at each site. However, recall, indicating higher false negative rates, was lower in in complex, dense urban environments (0.51–0.57 in Harare and Hyderabad) compared to more standardized US settlements (0.71–0.77). By comparing our data to contemporary building data, we found considerable urban structural changes and urban expansion reaching 350 % in our USA sites and 482 % in Harare. Despite lower accuracy than modern high-resolution analyses, our approach using Hexagon data extends the baseline for historic urban studies by three decades and is available globally, thus enabling mapping up to half a century of urban growth well before the availability of modern high-resolution satellite imagery.