What matters in the new field of machine learning and satellite imagery-based poverty predictions? A review with relevance for potential downstream applications and development research

This paper reviews the state of the art in satellite and machine learning based poverty estimates and finds some interesting results. The most important factors correlated to the predictive power of welfare in the reviewed studies are the number of pre-processing steps employed, the number of datasets used, the type of welfare indicator targeted, and the choice of AI model. As expected, studies that used hard indicators as targets achieved better performance in predicting welfare than those that targeted soft ones. Also expected was the number of pre-processing steps and datasets used having a positive and statistically significant relationship with welfare estimation performance. Even more important, we find that the combination of ML and DL significantly increases predictive power by as much as 15 percentage points compared to using either alone. Surprisingly, we find that the spatial resolution of the satellite imagery used is important but not critical to the performance as the relationship is positive but not statistically significant. The finding of no evidence indicating that predictive performance of a statistically significant effect occurs over time was also unexpected. These findings have important implications for future research in this domain. For example, the level of effort and resources devoted to acquiring more expensive, higher resolution SI will have to be reconsidered given that medium resolutions ones seem to achieve similar results. The increasingly popular approach of combining ML, DL, and TL, either in a concurrent or iterative manner, might become a standard approach to achieving better results.