Using Convolutional Neural Networks on Satellite Images to Predict Poverty

Since there are over a billion individuals worldwide below the international poverty line of less than $2 per day; the first goal of sustainable development is to eradicate poverty. The primary step before poverty can be eradicated is to understand the spatial distribution of poverty. However, the process of going around rural areas and manually tracking census data is time-consuming, needs a lot of human effort, and is expensive. On the other hand, high-resolution satellite images, are becoming largely available at a global scale and contains an abundance of information about landscape features that could be correlated with economic activity. Deep learning with satellite images provides a scalable way to make predicting the distribution of poverty faster, easier, and less expensive, and this helps in aiding organizations to distribute funds more efficiently and allow policymakers to enact and evaluate policies more effectively. This paper focuses on Africa as it is considered the poorest continent. The data, we have used, consist of three datasets which contain satellite images for three countries in Africa with different levels of poverty: Ethiopia, Malawi, and Nigeria. In order to classify the satellite images, two pre-trained Convolutional Neural Networks models (ResNet50 and VGG16) were implemented in addition to our novel structure of CNN. The test accuracy for CNN was 76% for the three countries. VGG16 average accuracy was 79.3% and ResNet average accuracy was 49.3%.