Predicting the quality of life from satellite imagery

Description

In large developing countries like India, it’s tough to base policy decisions on data. Of a population of 1.3 billion, 70% reside in rural areas with minimal digital trails. Large governmental initiatives such as the national census are collected once every 10 years. Despite the best of intentions, they aren’t comprehensive and quickly get out of sync with reality.

Ganes Kesari and Soumya Ranjan explore how satellite imagery offers an alternate ground truth that’s accurate at high resolution, available across periods as a time series, easily accessible, and is relatively inexpensive. While this is a rich source of visual information, the challenge has been in processing images and generating useful insights. The advances in deep learning have helped solve this last hurdle, placing enormous power in people’s hands and paving the way for socioeconomic data analytics.

This information helps answer basic questions about India’s patterns by applying deep learning on satellite imagery, enriched with census data, and using advanced analytics approaches. It can show the extent and pace of urbanization over time, provide a comparison of census-based indicators of poverty and visual indicators of development from aerial imagery, and identify anomalies observed in census-measured factors such as literacy, employment, and healthcare when viewed from the lens of satellite imagery.

Ganes and Soumya explain the inspiration of their work, Stefano Ermon et al, who used night light as a proxy to detect poverty in Africa. Deep learning requires a lot of labeled data; in this case, they used transfer learning to get over the initial hurdle. Planet.com releases some open data, and they used SpaceNet data for training the model’s weights, which was then adapted to the specific satellite imagery of villages and cities in India.

Three networks were trained to do three different tasks, extract building footprints, extract land use and road patterns, and extract night lighting patterns. The output of these three networks was structured and combined with the district-level census data to be used as an input for the socioeconomic analytics. Statistical techniques and machine learning algorithms were shed light on demographic insights that can drive potential policy decisions.