Observatory of Poverty: using machine learning and satellite images to measure global living conditions

Potential supervisors: 
Description: 

About 900 million people—one-third in Africa—live in extreme poverty. Operating on the assumption that life in impoverished communities is fundamentally so different that it can trap people in cycles of deprivation (‘poverty traps’), major development agencies have deployed a stream of development projects to break these cycles (‘poverty targeting’). However, scholars are currently unable to answer questions such as in what capacity do poverty traps exist; to what extent do these interventions release communities from such traps—as they are held back by a data challenge: a lack of geo-temporal poverty data.

This project will produce such data by using existing machine (deep) learning algorithms or develop new ones to identify poverty from satellite images (e.g. NASA’s Landsat mission, European Space Agency’s Sentinel II) between 1980s to 2020. The project (1) gathers census and survey data (2) assemble satellite image data from Google Earth Engine using javascript and Python (or R programming), and (3) experiment on evaluating the predictive power of existing algorithms to predict poverty from satellite images. The ultimate goal of this project is to build a new database of poverty trajectories for the African continent.

Machine learning and deep learning offer methods for knowledge representation, prediction, and decision making by learning from complex data. One of the challenges of learning algorithms is to tailor predictive models from high-dimensional (image) data to lower-dimensional interpretable data to be used in other domains such as for poverty research. These models and data are of great value for other domains in which they can be used for causal inference and policymaking. For example, with such data, scholars can predict where to allocate the next intervention to alleviate poverty.

Examples for project research questions

  1. To develop, train and evaluate deep-learning algorithms one set of countries (e.g., Nigeria, Ghana) and predict on another (e.g., Malawi, Egypt).
  2. To develop, train and evaluate deep-learning algorithms in one period (e.g., 1990s) and predict on another (e.g., 2000s)
  3. To develop, train and evaluate deep-learning algorithms using different dimensions of poverty (e.g., material living conditions and health).
  4. Evaluate the use of self-supervised models on satellite images.
  5. Compare machine learning models using non-satellite-image data with models that use satellite image data.
  6. Develop cross-validation techniques for geo-spatial predictions.
  7. Apply existing models to predict living conditions in one continent (Africa) and predict on another (India).

Requirements

  • Studies computer Science, physics or mathematics
  • Courses in machine learning and AI
  • Courses in statistics/math
  • Good programming skills (preferable in Python, Javascript, or R programing)
  • Being motivated, creative, focused.

The project is suitable for 2 students.

Contact: Adel Daoud, daoud@chalmers.se, Fredrik Johansson, fredrik.johansson@chalmers.se - AI Global PovertyLab

Date range: 
June, 2021 to December, 2023