NSU MA Graduate Research Helps Assess Poverty in Zambia

Owen Siyoto, an MA graduate of the Big Data Analytics & Artificial Intelligence Program at the NSU Mechanics and Mathematics Department, completed a thesis applying the learning transfer and ridge regression concepts to satellite imagery to predict poverty in Zambia. He assessed the Zambia Welfare Index using survey data, satellite imagery, and machine learning methods and yielded results comparable to similar work done in other countries.

Siyoto talked about his dissertation and its relevance,

The idea for the project came when I realized that in many developing countries poverty is still one of the most important problems. For example in Zambia, where I come from, 54.4% of the population lives below the poverty line. Another problem is that these countries cannot collect reliable information about where the poor are located. Collecting this information requires a large amount of resources, both time and money. That is why my supervisor and I began to explore the use of machine learning technologies to measure poverty

The learning transfer concept is gaining popularity for those working with Google satellite images with day and night light. Siyoto used a Convolutional Neural Network (CNN) that was previously trained to predict the well-being of an area based on the intensity of light in the image. In addition, the neural network made it possible to identify the types of buildings and roads to highlight the characteristics of the rich and poor areas of Zambia. Using this data, another algorithm was trained to estimate the poverty level of an area according to a set of characteristics.

Siyoto continued,

The work was very difficult, especially in the first year when I started the project and did not know many methods. For a long time I worked in a team with other students who helped me understand and implement the methods in this project. This research addresses a real problem so it has a very practical application. The United Nations is trying to solve the problem of poverty, and to do this it is necessary to find a cheap and reliable way to collect information. My research provides one of them since the prediction accuracy of the algorithm I applied is very high.

With the successful application of learning transfer and ridge regression, the MA graduate was not only able to assess the level of well-being of the country as a whole, but he demonstrated that the positive results from his model were not a statistical accident. In addition, the research provides statistics that governments and other stakeholders can use to solve the problem of poverty. Siyoto argues that this data would be nearly impossible to obtain at lower administrative levels due to the cost and machine learning technologies make the data available almost free of charge.