Satellite data and machine learning tools for predicting poverty in rural India

Abstract

Increased availability of satellite imageries and rapid development in algorithms to process imagery data has spurred interest amongst economist to use high frequency imagery data for meaningful economic interpretations. One such application is to use satellite night light data as an indicator of poverty. As poverty statistics in India is released once in five years, high frequency night lights data can be used to predict poverty in the years where official poverty statistics is not available. In this paper, we explored use of satellite night light data and machine learning algorithms (Artificial Neural Network) to predict rural poverty at sub-national level, i.e. state. We compared night light data with per capita domestic product as a predictor for the model. We find night light data as a better predictor of poverty than of per capita domestic product. Such predictions using satellite data can be used as a complement to the existing data-sets. This will facilitate economist for modelling the economic relationships in understanding poverty and provide more frequent and local estimates for policy makers.