Satellite monitoring of war urban damage with a temporal knowledge-guided deep learning scheme

Abstract Since 1990, between 40 and 68 countries, home to 46–79% of the population in the world were involved in armed conflict per year [1–3], which have caused huge socio-economic costs and severe casualties. Geotargeting is a central challenge for comprehensive assessments of wars and humanitarian assistance: it remains a difficult task to rapidly identify severe destructive zones with traditional or satellite-based methods. To overcome the challenge, we propose the first tempo- ral knowledge-guided detection scheme (TKDS) for satellite urban destruction monitoring. By integrating the proposed pixel-based Transformer model (PtNet) into the efficient TKDS scheme (TKDS-PtNet), we achieve the best destruction detection performance in the Syrian civil war and Russia-Ukraine war with high or medium resolution satellite imagery. Even using medium-resolution satellite images, the performance of the TKDS-PtNet model outperforms the state-of-the-art methods by 44.4 (more than twice as high as the latter, 72.4 vs 28.0) in terms of F1 score in six Syrian cities, and 36.1 (84.6 vs 48.5) in four Ukrainian cities. Our findings located 1,652 severely damaged buildings in Mariupol, Ukraine, including various urban infrastructures such as 81.5% (1,346) houses/apartments, 4.0% (66) industrial, as well as 11 garages, 2 schools, 2 kindergartens, and 1 sports center. Our scheme provides a low-cost, reliable, highly flexible, and generalizable solution for urban destruction monitoring in data-sparse environments. It can be applied to any urban context to near real-time monitor damage caused by wars, earthquakes, or extreme weather events, and approximate the spatio-temporal distribution of the damages to aid relief efforts, in addition to the estimation of potential rehabilitation costs.