Remotely measuring rural economic activity and poverty : Do we just need better sensors?

It is difficult and expensive to measure rural economic activity and poverty in developing countries. The usual survey-based approach is less informative than often realized due to combined effects of the clustered samples dictated by survey logistics and the spatial autocorrelation in rural livelihoods. Administrative data, like sub-national GDP for lower level spatial units, are often unavailable and the informality and seasonality of many rural activities raises doubts about accuracy of such measures. A recent literature argues that high-resolution satellite imagery can overcome these barriers to the measurement of rural economic activity and rural living standards and poverty. Potential advantages of satellite data include greater comparability between countries irrespective of their varying levels of statistical capacity, cheaper and more timely data availability, and the possibility of extending estimates to spatial units below the level at which GDP data or survey data are reported. While there are many types of remote sensing data, economists have particularly seized upon satellite-detected nighttime lights (NTL) as a proxy for local economic activity. Yet there are growing doubts about the universal usefulness of this proxy, with recent evidence suggesting that NTL data are a poor proxy in low-density rural areas of developing countries. This study examines performance in predicting rural sector economic activity and poverty in China with different types of satellite-detected NTL data that come from three generations of sensors of varying resolution. We include the most popular NTL source in economics, the Defense Meteorological Satellite Program data, whose resolution is, at best, 2.7 km, two data sources from the Visible Infrared Imaging Radiometer Suite (VIIRS) on the Suomi/NPP satellite with spatial resolution of 0.74 km, and data from the Luojia-01 satellite that is even more spatially precise, with resolution of 0.13 km. The sensors also vary in ability to detect feeble light and in the time of night that they observe the earth. With this variation we can ascertain whether better sensors lead to better predictions. We supplement this statistical assessment with a set of ground-truthing exercises. Overall, our study may help to inform decisions about future data directions for studying rural economic activity and poverty in developing countries.