Deep statistical learning-based image analysis for measurement of socioeconomic development in sub-Saharan Africa using high-resolution satellite images, and geo-referenced household survey data
Tackling inequalities and reducing poverty requires timely measurements of socioeconomic development at high spatial resolution. This is a substantial challenge, even in developed parts of the world, due to costly and infrequent data collection methods used for socioeconomic measurements. Availability of such data is even more challenging in low- and middle-income countries (LMICs), particularly in sub-Saharan Africa (SSA).
Over the past decade, researchers have investigated the potential for using emerging sources of large-scale data, such as satellite imagery and deep statistical learning, to advance how we make measurements in data-poor contexts. Studies, however, consist of cross-sectional proof-of-concept studies at a low spatial resolution that are not readily available to be used as input to development research or practice.
This project aims to widen the adoption of deep learning methods and the use of imagery data in development economics. In this project, we will build on existing methods for using imagery data to make measurements at high spatial resolution for multiple years in SSA. These data will then be used in ETH-DEC’s ongoing and future research, as well as shared with the wider community.
PI / Partners
- Dr. Kenneth Harttgen
- Prof. Isabel Günther
- Dr. Esra Suel
This project aims at using geospatial machine learning to measure housing inequalities in urban areas in sub-Saharan Africa for use in development economics research. Better data could lead to better models to extract information to be used to design pro-poor policies. The goal is to set up a data pipeline to be used in accurate and robust machine learning methods, exploiting paired Earth observation data and local census data. As a central case study, deepLNAfrical aims at inventorying and mapping informal settlements in sub-Saharan Africa.
The SDSC leads the implementation and development of techniques to extract information from geolocated data, such as Earth observation satellite images, geolocated natural images, paired census data, and related attributes. We will focus on models which are adapted to extract information from such data and make them easily reusable by the community, therefore supporting research in the same direction.
Measuring socioeconomic status at high spatial and temporal resolution is vital. In rapidly developing low and middle-income countries, population growth and urbanization rates are high, and data scarce. Emerging sources of large-scale data (e.g., satellite imagery, mobile phones, social media), have the potential to significantly improve the frequency and spatial resolution of the measurement of socioeconomic wellbeing. This could help to identify areas of concern earlier, which areas have the greatest needs, and investigate how those needs are linked to social and health outcomes.