Proxying economic activity with daytime satellite imagery: Filling data gaps across time and space

GAICIL Insights

Huge stretches of the world are economic dark matter: places where activity clearly happens, yet official statistics barely register it.¹ Proxying economic activity with daytime satellite imagery hands researchers a flashlight. The authors apply machine-learning techniques to a historical time series of daytime satellite imagery dating back to 1984 and use it to proxy economic activity across time periods and spatial units for which reliable data are otherwise unavailable.

The benchmark it beats is the famous night-lights map. Satellite data on night light intensity have become a common proxy for economic activity. The daytime proxy developed here more precisely predicts economic activity at smaller regional levels and over longer time horizons. Light at night shows where the bulbs burn; built-up land in daylight shows where the economy was physically constructed. The showcase is Germany, where East German data on economic activity are unavailable at detailed regional levels and for historical time series. A reunified country with a statistical hole in its own past becomes the test bed.

The engine is quietly clever. The technique classifies land cover from satellite images, combining data from four Landsat satellites to produce a time series of land cover types. Built-up surfaces become an economic signal. Abundant, freely observed imagery turns into measurements that no survey team could ever collect retroactively. The authors stress that their procedure is generalizable to any region in the world and has great potential for analyzing historical economic developments, evaluating local policy reforms and controlling for economic activity at highly disaggregated regional levels in econometric applications.

Think of it as AI doing archaeology on the present. The archive was already sitting in orbit, decades deep. What was missing was a reader. Machine learning supplies one, and in doing so it lets researchers ask questions about places and periods that had been written off as unmeasurable.

That changes policy evaluation. Judging a local reform requires knowing what happened before and after it, in the places affected and in comparable places that were spared. Where official data are thin, evaluation becomes guesswork dressed up in regressions. A credible proxy for local activity, stretching back decades, allows governments to measure the effects of infrastructure, regulation and institutional change with far greater precision, and to drop policies that fail.

The authors sketch a next frontier that is itself a story of AI progress. Image segmentation procedures based on convolutional neural networks, such as U-Net or ResNet, could deliver even more accurate land-cover classification by using contextual information from neighboring pixels. Such networks could help distinguish building types, such as stores or industrial buildings, and evaluate housing quality at a global scale. The trade-offs are named plainly: higher computing-power demands and possibly additional region-specific calibration of ground-truth data, with hyperparameter tuning and model pre-training proposed as remedies.

Sharper sensors add more. Satellite data with finer spatial resolution than Landsat, such as those from the ASTER instrument or the Sentinel mission, could support more sophisticated metrics of economic activity. These cover shorter time series and help less with history, yet they are potentially very valuable for recent events in areas where reliable data are otherwise unavailable. The research program compounds: every improvement in models, compute and sensors raises the resolution at which economies can be seen.

Public debate tends to cast AI as the watcher to be feared: surveillance, error, displacement. This paper shows the other face. Machine learning applied to openly available imagery produces public knowledge, sharpens policy evaluation, and gives regions without strong statistical systems a way to measure their own development. Once the procedure exists, extending it to new regions costs little. That makes it an equalizer in access to economic data.

The conditions for that payoff are worth naming. Access to satellite archives, computing resources and training data decides how far such methods can go. The authors describe their paper as laying first methodological foundations. Those foundations stay valuable only if data remain open and researchers remain free to try new models. Keep both, and AI becomes a lens for seeing the economy clearly, which is the precondition for governing it well.

The striking implication: some of the most valuable economic datasets of the coming decades may already exist, unread, in public archives. The bottleneck is the method, and methods like this one are how the archive starts to speak.