Quantifying urban growth since the 1970s using Hexagon spy satellite imagery and deep learning building footprint detection

Keywords: building extraction, ResNet, convolutional neural networks
Abstract Type: Paper Abstract

Abstract

High-resolution datasets capturing the distribution, size, and arrangement of buildings are essential to creating and planning for sustainable, resilient, and livable settlements. Remote sensing excels at mapping building data accurately. However, analyses of high-resolution satellite imagery that can identify single buildings are limited by the lack of data before 2000, which is unfortunate because many regions experienced rapid population growth before. Here, we evaluated the potential of high-resolution panchromatic spy-satellite imagery from the 1970s as a basis to map urban growth. We used a deep learning Mask R-CNN approach to detect building footprints in Hexagon imagery from 1972 to 1979 across four urban growth hotspots in the USA, Zimbabwe, and India. We achieved good overall mapping results (precision 0.83-0.91) and detected 73-94% of the overall building area. However, we observed higher false negative rates in complex urban environments as compared to more standardized building patterns. We contrasted 1970s building detections with contemporary data and quantified considerable growth in all sites due to sprawl and densification, such as a building area growth of up to 350% in our US-American sites and 480% in Zimbabwe. Our results suggest that analyzing Hexagon spy satellite images, while less accurate than modern multi-band imagery, can provide accurate building data for the 1970s. Since Hexagon is available globally, this extends the baseline of many urban research applications three decades back in time from the advent of multispectral, high-resolution satellite imagery in the early 2000s, and allows to map half a century of urban growth.