Predicting Poverty of a Region from Satellite Imagery using CNNs

Abstract

Poverty in a socioeconomic context can be defined as the inability of individuals to meet their basic needs. Measuring poverty is important to target efforts in places that need aids the most and evaluate the effectiveness of government programs. However, it is difficult and expensive as it requires the collection of detailed data from the households. The development of machine learning-based techniques has enabled the use of big data such as social media, mobile phone, and satellites for poverty measurement. In this thesis, Convolutional Neural Network (CNN) models are evaluated to directly predict poverty from daytime satellite imagery. Two approaches, naive and semantic segmentation, are proposed and compared with the multistep learning approach that uses nighttime lights image. We perform experiments using publicly available daytime and nighttime satellite images from Google Maps and NOAA. The best model is achieved by combining the semantic segmentation approach and the night lights data. Moreover, we test the generalizability of the models using higher-level administrative and out-of-country data. The test reveals that we can use the models to estimate poverty in higher level administrative region, but they are not robust to be used to predict poverty in other countries.