Fig. 2: Visualization of economic development levels predicted by our human-machine collaboration model.

(A) Prediction scores over grid images averaged over 4 years from 2016 to 2019, (B) the yearly aggregated VIIRS nightlight data in 2019 from Earth Observation Group, Payne Institute for Public Policy¹⁵, and (C) the land cover classification map released by the Ministry of Environment (MoE), Republic of Korea in 2019. The zoomed-in views in (D–F) compare predictions for Sepho County in the Kangwon region. From left to right are the Copernicus Sentinel-2 satellite images [2019] (D), model predictions (E), and manually verified buildings colored red from the building footprint data from National Geographic Information Institute (NGII), Republic of Korea in 2014 (F). The land cover classification map shown in (C) uses the ‘North Korea land cover map’ created by MoE. The map is opened to the public as the KOGL first type and can be downloaded for free by directly visiting MoE Informatization Office (Sejong City, Doum6-ro 11, MoE 6th floor, South Korea). The building footprint data shown in (E) uses the ‘the digital map’ created by NGII. This data is opened to the public as the KOGL first type and can be downloaded for free from the National Spatial Data Infrastructure Portal (http://www.nsdi.go.kr/lxmap/index.do).