A multi-sensor approach for characterising human-made structures by estimating area, volume and population based on sentinel data and deep learning

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

“…This study is a continuation of the work described in Fibæk, Keßler, et al (2021), which presents a multisensor approach that can reach high levels of accuracy for mapping the area, volume, and population of structures in Denmark. That research, in turn, builds on Frantz et al (2021) and Haberl et al (2021), which showed that Sentinel imagery could be used to predict the height of structures, as well as Corbane et al (2020) and the Global Human Settlement Layers' S2‐NET, which has been used to create global maps of built‐up areas extracted using the Sentinel 2 satellites and Deep Learning.…”

“…Fibæk, Keßler, et al (2021) and Frantz et al (2021) show that it is possible to estimate the building height and volumes of structures in Northern Europe using Sentinel 1 and 2 imagery. Using the volume instead of the area for the population estimates would increase the robustness of the predictions, especially in dense urban areas and clusters with industrial or high‐rise buildings.…”

“…For many areas in Africa, Sentinel data are only available from one orbital direction. Fibæk, Keßler, and Arsanjani (2021) investigated the effects of this on the accuracy of structural predictions and found a significant decrease in accuracy. If both orbital directions become available, it will likely be possible to reduce the number of errors caused by the alignment of structures and natural features (Koppel et al, 2017).…”

“…Predictions on area, structure type, and population were made for the two primary study areas. The choice of including Danish ground truth data is due to the availability of complete ground truth data and the previous application of Sentinel data to map structural characteristics in Denmark (Fibæk, Keßler, et al, 2021; GeoDenmark, 2021). Including data from Northern, Western, and Eastern Africa ensures the applicability of the models in diverse contexts.…”

“…Sentinel 1 and 2 data is well suited for far‐field applications as global data are openly and freely available. The spatial resolution of the Sentinel satellites' data products allow the mapping of individual structures and characteristics, such as area, volume (Haberl et al, 2021) and estimated population (Fibæk, Keßler, et al, 2021; Schug et al, 2021), where structures refer to buildings and other human‐made space‐enclosing structures. However, current approaches to structure mapping from Sentinel data have either not shown applicability in data‐poor regions (Fibæk, Keßler, et al, 2021; Schug et al, 2021) or do not utilize the full potential of multisensor approaches nor all the spectral bands available from the Sentinel 2 satellite (Corbane et al, 2020).…”