Predicting Poverty Index on Satellite Images & DHS Data using Transfer Learning

Abstract

For economic livelihood, reliable information remains skimp in the developing world. Poverty is one of the seventeen sustainable development goal to be get removed as mentioned by the United Nations.  In the earlier times, the primary source of data to measure poverty is ground-level survey data such as household consumption and wealth. With the advancement of technology, among different approaches that are available the one adopted by the developed or developing countries is to estimate the poverty index of an area using remote sensing data such as satellite images using machine learning technique. The source of data used in this approach is highly structured, inexpensive and easily available. This machine learning technique not only estimates the poverty index of a year, but also establishes the relationship of index between the years. Our proposed approach uses pre-trained Inception Net_V3 to predict the nightlight intensity corresponding to input daytime satellite images. The proposed model also predicts the cluster wealth score and established the correlation between wealthscore obtained from Demographic and Health Survey (DHS) data and Satellite Images i.e. r value (Pearson Correlation Coefficient)