Exploring the use of satellite imagery and computer vision-based machine learning method to improve the spatial granularity of poverty statistics

Abstract

Spatially granular poverty statistics can enhance the efficiency of targeting resources to improve the living conditions of the poor. Previous studies suggest that the use of high-resolution satellite imagery may be an alternative approach in generating granular poverty maps. This study outlines the methods in improving the spatial granularity of government-published poverty estimates using convolutional neural networks and ridge regression applied on publicly available satellite imagery, household surveys, and census data from the Philippines and Thailand. A convolutional neural network (CNN) was used to extract features of satellite images that are correlated with the intensity of nightlights. These features were then aggregated at the same level for which government-published estimates were available to estimate a prediction model for poverty rates. Results suggest that the adopted methodology performed satisfactorily in predicting lower levels of nightlight intensity for the specific years considered in this study. Additional preliminary numerical assessment also reveals that prediction accuracy may be enhanced by using random forest as an alternative to ridge regression. The use of proprietary satellite images with higher resolution may also improve prediction accuracy.