Predicting Health and Living Standards of India using Deep Learning

Abstract

Poverty eradication is an inexorable process in human growth [21], with poverty estimation being the first and most important stage. Identifying strategies for poverty reduction programs and distributing resources appropriately requires determining the poverty levels of distinct places throughout the world. However, trustworthy data on global economic livelihoods are scarce, particularly in poor countries, making it difficult to provide programs and track and evaluate success. This is partly since this information is gathered through time-consuming and costly door-to-door surveys. Furthermore, survey data includes large gaps, especially in densely populated countries like India. Therefore, we use overhead satellite imagery that contains characteristics that make it possible to estimate the region’s poverty level along with the survey data. In this work, we develop deep learning models that can predict a region’s poverty level from both DHS survey data and overhead satellite images. This study makes use of both daytime and nighttime imagery in different combinations and analyzes the performance. Poverty prediction studies are mostly focused on datasets from Africa, and very few studies have used a dataset from India. Therefore, in this, thesis, we train a Single Frame model with two deep CNNs having ResNet-18 architecture to predict the average cluster wealth index which is an indicator of poverty given a satellite image of the cluster using DHS survey data and satellite imagery.