Poverty Detection using Satellite Imagery

Developing countries spend a lot of resources in planning and implementing schemes and policies for poverty eradication. To collect data on poverty, it needs a lot of manual labour and consumes a lot of time and money. To make this process easier, this publication proposes to use satellite images to detect economic activity and, as a result, estimate poverty in a location. A convolutional neural network is trained to learn various developmental parameters like rooftop type, source of lighting and proximity to water sources. The output of this model will help to train another model which helps to understand the relationships between the features learned for the prediction of different developmental parameters. Using this it will be able to estimate the income levels of a particular place and this will be used as a proxy for poverty detection. The results are then compared with the ground truth. It is observed that the results are very close to the ground truth with accuracy of around 99 % and error of 0.77 across all datasets.