Poverty prediction by using deep learning on satellite images

Abstract- As the universe finds it challenging to define poverty, the World Bank consider poverty as anyone living below $2 per day. Governments and international organizations are working to eradicate this poverty. In this study, some research on satellite imagery on poverty prediction through the concept of CNN (Convolutional Neural Network) is reviewed. The study considered satellite images of Nigeria, Mali, Malawi and Ethiopia from Kaggle. While 90% of the dataset was used for training, the remaining 10% was used for testing. Datasets were analyzed using CNN, VGG16 and ResNet50 and it was observed that the VGG16 model outperformed the other two models with a validation accuracy of 94%. CNN had the second with 91% and ResNet had the lowest validation success with 62%. The rise of high-resolution satellite imagery containing comprehensive data of patterns, features and landscapes of regions or countries can be applied to determine the economic lifestyle of people or nations. Applying satellite imagery to poverty estimation would be easier, faster and cheaper. This study suggests that large satellite images should be available for each region or country. In future studies, researchers should focus on satellite imagery to apply it to the prediction and detection of poverty and crime, road traffic, agricultural land and similar practices.