Classifying satellite images for socio-economic activity
This project proposes a methodology for predicting subnational economic development using high-resolution, daytime satellite imagery.
- Klaus Ackermann, SoDa Labs and Department of Econometrics and Business Statistics
- Nandini Anantharama, SoDa Labs and Faculty of Information Technology
- Alexey Chernikov, Faculty of Information Technology
- Paul Raschky, SoDa Labs and Department of Economics
- Miethy Zaman, Deakin University
Project background and aims
The goal of this project is to go beyond nighttime light luminosity as a measure for subnational economic development or social activity and use daytime imagery.
High resolution daytime satellite images are becoming relatively more available for the entire world and compared to night lights, it contains more information about the landscape which are reflective of economic activity.
As mentioned before, night lights have the limitations of distinguishing between the poor and densely populated areas and in these cases, the daytime images can fill in the gap. However, with more information in the daytime images, they are highly unstructured and thus makes it rather difficult to extractinformation which can be scaled to some economic measure.
We first employ a convolutional neural networks (CNN) approach extract physical features from the daytime images (e.g., roads, railways, buildings). The predicted values then will be used, to build a second model, to predict economic indicators (e.g. GDP) from the aggregated OSM predictions, from grid cells to regions.