Estimating link-level vehicle density in data-scarce regions using satellite imagery and computer vision: the case of Pyongyang, North Korea

Transportation studies in data-scarce regions are constrained by the lack of conventional data. Using North Korea as a case study, this research integrates high-resolution satellite imagery, deep learning-based vehicle detection, and non-negative matrix factorization (NMF) to estimate link-level vehicle density in Pyongyang. Vehicles are detected using a You Only Look Once (YOLO)-based oriented bounding box model, and NMF imputes missing observations. The results show substantial spatial variation: vehicle density is higher in the urban core and along major corridors, while much of the road network has relatively low density. The analysis also identifies potential mismatches between infrastructure supply and observed vehicle activity. This study offers one of the first quantitative, link-level assessments of vehicle density in North Korea. It also presents a scalable and transferable framework combining satellite imagery, computer vision, and matrix factorization to estimate traffic activity where conventional transportation data are limited or unavailable.