Prediction of demographic indicators from remote sensing images

Abstract

This thesis explores the use of satellite images to assist in the analysis of data from demographic surveys in the context of Sub-Saharan African countries. In these countries, demographic surveys are rare and provide little updated environmental information, limiting the study of the relationships between environment and population at a fine scale. Satellite images, available at a high temporal frequency and with high spatial resolution, allow for the addition of environmental factors to the analysis, but require specific preliminary methods that meet the constraints determined by survey data. Deep learning methods can be used to train models to characterize the environment. The thesis examines the link between environmental characterization and population at two study scales: at the city level and the country level. At the local scale, our first case study examines mortality in Antananarivo, the capital of Madagascar. Health authorities record all deaths in the city and collect information on the deceased, such as the cause of death and their residential neighborhood. This geographic information allows for the linking of deaths to neighborhood environments. The environmental characterization uses a map based on the "Local Climate Zones" (LCZ) classification system. This network is trained via a domain adaptation method, using the physical definitions of the LCZ system. In addition, usual indicators such as vegetation index, altitude, and building detection are integrated. After attributing each death to a neighborhood, the influence of the environment on the causes of death is studied through these indicators, controlling for socio-economic data to ensure the independence of the results. At a larger scale, the second case study examines this relationship at the national level. International organizations, such as the Demographic and Health Survey Program, organize surveys to evaluate demographic indicators in low- and middle-income countries. A "Mala