Institutional Repository

A machine learning-guided data integration framework to measure multidimensional poverty

Show simple item record

dc.contributor.advisor Mnkandla, Ernest Musungwa, Tapera 2021-07-15T09:59:18Z 2021-07-15T09:59:18Z 2021-03
dc.description.abstract As developing nations like South Africa chart a path of socio-economic development, the spatialisation of progress, opportunity, and neglect is a critical antecedent to policy-making and regional interventionism. Efforts to capture meaningful data using household surveys and censuses face a diluted accuracy due to sampling, surveying, and quantification errors. The reliability and regularity of these traditional methods is also constrained since the processes are costly and time consuming. Recent investigations in the field of machine learning and satellite imaging have presented a viable proof-of-concept technique to exploit specific economic indicators to demonstrate economic development patterns across regional areas. The current study adopts several interrelated approaches encompassed within the field of remote sensing in order to evaluate and model poverty in the South African landscape. By adopting publicly accessible information for classification to indicate the intensity of poverty, this study proposed an inexpensive solution to poverty estimation. Concretely, the solution combined satellite imagery and geospatial data with regional poverty data exploiting an ensemble approach to poverty diagnosis. The solution is based upon multidimensional indicators and multi-layered insights that can be extrapolated from overlapping models to bolster them and help with socio-economic well-being estimations. Through machine learning techniques and object-oriented training of a convolutional neural network, this study revealed that a naïve combination of distinct data sources shows patterns of socio-economic well-being in South Africa by achieving an R2 of 0.56 wealth estimation compared to 0.54 from satellite imagery. This outlined variability and incongruity within landscapes that not only reflect the persistent subdivisions of apartheid-era enclavisation, but indicate critical gaps in domestic social services, infrastructure, and developmental pathways. This study is applicable to policy makers in low- and middle-income countries that lack accurate and timely data on economic development as an important precursor to public support, policy making, and planned expenditures. en
dc.format.extent 1 online resource (xiv, 138 leaves) : illustrations, color graphs, color maps, color photographs
dc.language.iso en en
dc.subject Poverty prediction en
dc.subject Machine learning algorithm en
dc.subject Satellite imagery en
dc.subject Deep learning en
dc.subject Convolutional neural network en
dc.subject Computer vision en
dc.subject Geospatial information en
dc.subject Night-time lights en
dc.subject Data for development en
dc.subject.ddc 339.460285631
dc.subject.lcsh Poverty -- South Africa -- Forecasting en
dc.subject.lcsh Machine learning en
dc.subject.lcsh Poverty -- South Africa -- Data processing en
dc.subject.lcsh Information storage and retrieval systems -- Poverty en
dc.subject.lcsh Computer vision -- South Africa en
dc.subject.lcsh Spatial data mining -- South Africa en
dc.title A machine learning-guided data integration framework to measure multidimensional poverty en
dc.type Dissertation en
dc.description.department School of Computing en M. Tech. (Information Technology)

Files in this item

This item appears in the following Collection(s)

Show simple item record

Search UnisaIR


My Account