Effect of spatial scale, color infrared and sample size on learning poverty from aerial images

There is a growing amount of literature that focuses on using machine learning algorithms to predict poverty from satellite and aerial images on a low regional level, particularly for countries without a high-quality official statistical system. The data used for annotating images and training an algorithm are generally based on sample surveys. In The Netherlands, statistics on income and poverty are derived from tax registers resulting in a complete enumeration of the Dutch population. In this paper, we use this complete enumeration to simulate to which extent satellite or aerial images can predict poverty on low regional levels. After geocoding these households, aerial images are annotated and a deep learning algorithm is trained to predict poverty. The precision of the predictions is evaluated by comparing it with the true poverty fractions known from tax registers. The effect of different spatial scales (1-ha vs. 25-ha images), spectral bands (RGB vs. CIR), and sample sizes for the training set are compared. It is discussed how this information can be used in the production of low regional statistics on poverty in countries where high-quality official statistical systems are lacking.