Detecting Human Interpretable Features to Support Poverty Mapping

ENHANCING THE EFFECTIVENESS OF SOCIAL SAFETY NETS 

THROUGH AI and MACHINE-LEARNING INNOVATION

Abstract

In response to worsening food security, developing countries rely on poverty mapping to identify impoverished communities and provide social transfers. Recently, deep learning-extracted satellite features have been used as proxies for household wealth but with a minimal understanding of local poverty. To address these challenges, the proposed study will develop human-interpretable and context-aware wealth metrics to enhance the responsiveness of social safety net programs, specifically the Food Security Pack (FSP), in Zambia. Our AI/ML innovation aims to provide an intuitive, locally relevant, and fine-grained picture of the region’s built environment. We will automatically extract features related to poverty, map their locations, and quantify the contribution of each measured variable. The project is composed of four steps: 1) community engagement to identify visual signs relating to poverty; 2) acquisition and labeling of high-resolution satellite imagery; 3) development of semantic segmentation models to obtain human-interpretable features, such as rooftops and roads; and 4) mapping these features to estimate a granular/community-level of poverty. In doing so, we can increase the poverty-reducing impact of the FSP distribution. This project empowers local residents to conceptualize poverty and assist the government in reaching communities in need.

​

​

​

​The team consists of computer/data scientists

and social scientists:​

​

​​

• Woojin Jung (PI) - Rutgers, School of Social Work

• Dmitris Metaxas - Rutgers, Computer Science

• Quentin Stoeffler - University of Bordeaux, Economics

• Saeed Ghadimi - University of Waterloo, AI Institute

• Dimitrios Ntarlagiannis - Rutgers, Earth and Env. Sciences

• Arunesh Sinha - Rutgers, Mgmt Science and Information Systems

• Min Xu - Rutgers, Statistics

• Tawfiq Ammari - Rutgers, Communication and Information

• Andrew Kim - Rutgers, School of Social Work

• Maryam Hosseini - MIT, Urban Studies and Planning

• Melissa Sartorius - Rutgers, School of Social Work

Research Duration
.................................................
 
July 2021 - Ongoing

The study will conclude with the dissemination of findings to policymakers.

           4 STAGES        ................................................​​

1. Community engagement to visualize poverty

2. Acquisition and labeling of satellite imagery

3. Development of deep learning models to obtain intuitive features

4. Creation of high-resolution poverty map and index

• Sajedeh Goudarzi - Rutgers, Global Urban Studies

• William Benjamin - Allstate Corp., Data Science

• Rofaida Benotsmane - Istanbul Technical Univ., Economics

• Yuxiao Lu - Singapore Management Univ., Computer Science

• Vatsal Shah - Rutgers, Computer Science, Statistics

• EJ Knopf - Rutgers, School of Social Work

• Jamie Cramer - Rutgers, School of Social Work

• Tamara Billima - Local NGO

• Bernard Tembo - Local NGO

• Owen Siyoto - Local NGO

• Cheelo Mwiinga - Local NGO

The research team represents diversity across disciplines, countries/continents (Africa, Asia, US), languages, and cultures. Our study can contribute to broadening the reach of social protection programs by identifying vulnerable and poor populations in the Republic of the Congo and other countries where there is very limited, georeferenced wealth data.