Convolutional Expectation Maximization for Population Estimation

There is a fundamental spatial mismatch in the data available for estimate population from satellite imagery. Spectral reflectances are available for each pixel of an image, but ground reference population data are available only for larger zones, therefore satellite imagery has bigger resolution than ground reference images. The general response has been to build models for the average population density of the zones, utilizing spatially aggregated spectral data. This article reports a new approach to solve this problem where per pixel spectral data are used. The already used expectation maximization algorithm (EM) [1] is paired with a convolutional neural network to improve the resolution of a preexistent population ground truth provided by the GHS POPULATION GRID (LDS) [5]. We start with the satellite imagery by Sentinel-2 mission and, the regression model we have built, upscales the LDS dataset to 10 meters resolution, the same as Sentinel-2 images. As you can see in Table 1, we obtained an AvgRelDelta of 50.05 for Uganda’s rural area and 96.31 for the city of Lusaka in Zambia, while our method scored 87.57 in Overall. This results are computed from a test dataset provided for ImageCLEF 2017 Population Estimation Task [8].