Deep learning: population estimation with Sentinel 1 & 2

Abstract

High resolution gridded population maps represent an integral part of the efforts to monitor and implement the Sustainable Development Goals of the United Nations. However, censuses often lack spatial resolution, temporal consistency and are prone to systematic omission. Due to the increasing spatial and temporal resolution, as well as the availability of Earth Observation data, satellite data information could overcome deficiencies in population data. To estimate population in Uganda, this study proposes and evaluates three different Convolutional Neural Network (CNN) architectures which only rely on satellite data. These architectures include a novel combination of Sentinel-2 (RGB, NIR, SWIR) and Sentinel-1 SAR backscatter, as well as interferometric coherence. The CNNs are assessed twofold: With a regression analysis testing the CNN’s capability to predict population within an administrative unit as well as with a disaggregation analysis to determine the ability of the CNNs to map population counts beyond the reported census to a finer gridded resolution. To further assess the quality of the products, results are compared to WorldPop’s population estimations. The proposed CNN Uganda 3- Branch-ConvNet was able to keep up with WorldPop’s estimations by feeding the satellite data input sources through three separate branches into the architecture. Further, the results indicate that SAR data can enhance population estimations, but a CNN solely relying on Sentinel-2 data can achieve similar results as well. The study’s results suggest that satellite data could complement censuses. The study itself contributes to a free and automated framework, which is required to estimate population frequently.