Temporal Prediction of Socio-economic Indicators Using Satellite Imagery

Machine learning models based on satellite data have been actively researched to serve as a proxy for the prediction of socio-economic development indicators. Such models have however rarely been tested for transferability over time, i.e. whether models learned on data for a certain year are able to make accurate predictions on data for another year. Using a dataset from the Indian census at two time points, for the years 2001 and 2011, we evaluate the temporal transferability of a simple machine learning model at sub-national scales of districts and propose a generic method to improve its performance. This method can be especially relevant when training datasets are small to train a robust prediction model. Then, we go further to build an aggregate development index at the district-level, on the lines of the Human Development Index (HDI) and demonstrate high accuracy in predicting the index based on satellite data for different years. This can be used to build applications to guide data-driven policy making at fine spatial and temporal scales, without the need to conduct frequent expensive censuses and surveys on the ground.