Mapping socioeconomic conditions using satellite imagery : a computer vision approach for developing countries

Pakistan is a developing country with more than a quarter of its population living under the poverty line. Data relating to the socioeconomic conditions of the population is scarce and sporadic. Government, NGOs, and other international organizations perform door-to-door surveys to collect data, but these can be expensive and time-consuming to conduct. Currently, statistics about poverty in Pakistan exist only as estimates from field surveys, which are often unofficial and limited. This lack of reliable information leads to ineffective policy decisions. This is one of the biggest challenges in fighting issues such as poverty. Reliable information is thus needed for the unified and concerted direction of development activities. This thesis aims to build a computer vision system that can automat- ically extract information about poverty from publicly available high-resolution satellite imagery. The proposed system output is heat maps indicating poverty and development levels. These heat maps can be overlaid on digital maps for easy visualization. We propose to use transfer learning techniques to extract indicators such as geographical and man-made structural features from high-resolution satellite imagery. The trained model learns to filter and distinguish between various terrains and man-made features, such as highways, buildings, and farmlands. We show that these learned traits are quite useful for mapping socioeconomic variables and even come close to matching the prediction data with field survey data. For a developing country like Pakistan, this thesis is useful because the approach uses publicly available data and is a scalable and inexpensive alternative to traditional surveys. The results from our thesis can aid policymakers and NGOs in distributing their funds to areas that are most deserving and enacting and assessing policies more effectively.