A deep learning method for creating globally applicable population estimates from sentinel data

Recent research has shown promising results for estimating structural area, volume, and population from Sentinel 1 and 2 data at a 10 by 10-m spatial resolution. These studies were, however, conducted in homogeneous countries in Northern Europe. This study presents a deep learning methodology for population estimation in areas geographically distinct from Northern Europe. The two case study areas are Ghana and Egypt's Mediterranean coast, with supplementary ground truth data collected from Uganda, Kenya, Tanzania, Palestine, and Israel. This study aims to answer the question: How can we use Deep Learning to map structural area and type to derive population estimates for Ghana and Egypt based on Sentinel data? At 10 by 10-m resolution, the accuracy of the presented area predictions is similar to the Google Open Buildings dataset. An intercomparison of the presented population predictions is made with global state-of-the-art spatial population estimates, and the results are promising, with the proposed methodology showing comparable or better results than the state-of-the-art for the study areas.

Automated summary

Recent research has shown that estimating structural area, volume, and population using Sentinel 1 and 2 data at a 10 by 10-m resolution is promising. This study presents a deep learning methodology for population estimation in geographically distinct areas from Northern Europe. The results demonstrate that the proposed methodology shows comparable or better results than the state-of-the-art for the case study areas of Ghana and Egypt.