Measuring Quality of Life in Gauteng using satellite images and machine learning

Abstract

Quality of life is a multidimensional concept that accounts for various socioeconomic factors, including wealth, safety, and access to education, healthcare, and basic services. Understanding and improving the quality of life for people around the world is a global endeavour supported by several of the United Nations Sustainable Development Goals. Governments and organisations rely on socioeconomic data to inform the quality of life of a particular country or region. Traditionally, socioeconomic data is collected via surveys and censuses. However, these are expensive to run and require significant work and coordination. Prior research investigates using machine learning and big data sources like satellite images to predict various socioeconomic indicators. In recent years, studies on several sub-Saharan countries have shown the potential to measure poverty via indicators such as asset wealth and consumption expenditure. However, these studies often focus on the narrow scope of economic wellbeing. Limited work has been undertaken in estimating proxies of multidimensional quality of life from satellite images, especially in a South African context. This research investigates the question: “how well can a machine learning model estimate a quality of life index from satellite images in Gauteng, South Africa?” By adapting the existing gold standard methods for predicting poverty from satellite images, this research investigates predicting a multidimensional quality of life index from daytime satellite images. The quality of life index for wards across Gauteng, South Africa is obtained from data provided by the Gauteng City-Region Observatory. This dataset is used to train a convolutional neural network to extract features from daytime satellite images (obtained from Planet Labs). From these features, the selected quality of life index is predicted. Results obtained are suboptimal with a negative Coefficient of Determination, which is below baseline performance for poverty prediction and worse than predicting the sample mean of the test set. A critical analysis of the methods and results concludes that the methods applied are not suitable for a multidimensional index such as the one collected by the Gauteng City-Region Observatory. Primary conclusions suggest that this result is due to the subjective nature of factors in the index, the appropriateness of the machine learning model selected, and the physical disparities of South African urban environments within small spatial areas. A primary recommendation is further exploration of the boundaries of using machine learning methods with satellite data for measuring quality of life.