Using maps to predict economic activity
We introduce a novel machine learning approach to leverage historical and contemporary maps to systematically predict economic statistics. Remote sensing data have been used as reliable proxies for local economic activity. However, they have only become available in recent years, thus limiting their applicability for long-term analysis. Historical maps, on the other hand, date back several decades. Our simple algorithm extracts meaningful features from the maps based on their color compositions. The grid-level population predictions by our approach outperform the conventional CNN-based predictions using raw map images. It also predicts population better than other approaches using night light satellite images or land cover classifications as the input for predictions.
READ FULL TEXT