Predicting household water supply using satellite imagery and deep learning

Abstract

In this study we map Demographic and Health Surveys (DHS) urban household water supply data from 30 African countries and 52 DHS-surveys to Sentinel 2 RGB data and show that modern convolutional neural networks can find a mapping function and predict abstract variables derived from DHS data, like household water supply. In addition, the purpose of this research is to show the ability of such networks to predict data for areas and countries where no survey data are available. Therefore, we use one-year medians of 2×2km cloud removed Sentinel 2 tiles at the surveyed locations in a VGG19 CNN and classify sources of water supply. In addition, we perform a regression analysis for the distance and the first principal component of a PCA. We achieve an F1-score of up to 0.76 for the classification and an r² of 0.76 for the prediction of the first principal component. The prediction of the distance to the water source is less precise with an r² of 0.57, which is potentially due to the extreme skewness of input data. In further studies, we want to prove that these results can also be achieved for rural and mixed models as well as for other food security indicators, such as asset wealth.