Combining satellite imagery and machine learning to predict poverty.

Area of Science:

• Remote Sensing
• Machine Learning
• Development Economics

Background:

• Economic livelihood data is scarce in developing nations, hindering policy and research.
• Accurate poverty assessment is crucial for effective development interventions.

Purpose of the Study:

• To develop and validate an inexpensive, scalable method for estimating economic outcomes using satellite imagery.
• To assess the effectiveness of machine learning in analyzing satellite data for poverty tracking.

Main Methods:

• Utilized high-resolution satellite imagery and survey data from five African countries.
• Trained a convolutional neural network to identify image features correlated with economic indicators.
• Leveraged publicly available satellite data for scalability.

Main Results:

• The convolutional neural network model explained up to 75% of the variation in local-level economic outcomes.
• The method proved accurate, inexpensive, and scalable across diverse African settings.
• Demonstrated successful application of machine learning with limited training data.

Conclusions:

• Satellite imagery combined with machine learning provides a powerful tool for estimating economic livelihoods.
• This approach can significantly enhance efforts to track and target poverty in developing countries.
• The methodology shows broad potential for scientific applications with data limitations.