This project aims to enhance our understanding of the pace of agricultural and rural transformation in contemporary sub-Saharan Africa, its poverty and distributional impacts and drivers. This is done by mixing methods – household surveys, satellite imagery and machine learning – in human development studies. It is a collaborative project between AI researchers at Halmstad University and Social Scientists at the Department of Human Geography at Lund University.
The research project Mixed Methods addresses a longstanding and to date unresolved theoretical question in agrarian studies, namely whether agricultural transformation is poverty driven as proposed by neo-Marxist perspectives or if it enables inclusive growth as propounded by advocates of the pro-poor, agricultural growth model.
Building on the methodological advances in machine learning and artificial intelligence, the researchers will combine training data from an existing survey-based database (Afrint), covering around 3 000 households from six African countries, distributed over sixteen different regions, 54 villages and spanning fifteen years of development. Drawing on this unique combination of data sources and methods, they will be able to provide new insights into the distributional effects of agricultural transformation, using a variety of established welfare indicators from the field of rural development studies, but breaking new ground in development research by collecting them through remote sensing techniques. This innovative mixed methods approach will also provide a real-life contribution to addressing a practical problem of collecting statistics on the ground in developing countries that lack infrastructure or administrative resources.
In the mid-1990s the National Aeronautics and Space Administration (NASA) approached the research community in an effort to realise the potential of satellite imagery – specifically addressing the social sciences. High hopes were expressed in People and Pixels: Linking Remote Sensing and Social Science (Liverman et al. 1998). But the results have been meagre and their added-value questioned (Hall 2010, Longley 2002). At that time remote sensing was difficult to combine with socio-economic data; at most it could provide maps and quantitative appraisals of surface features.
The prospect has changed as suggested by recent advances in artificial intelligence (AI) and nearby disciplines. For example Jean et al. (2016) demonstrate that satellite imagery combined with machine learning can document poverty in Ugandan villages with a precision similar to that achieved with traditional household questionnaires. This has sparked interest in the policy community and even suggestions to abandon surveys as the workhorse of development research (Mullainathan 2016). To date, however, only three independent studies using remote sensing and machine learning to study poverty broadly exist to substantiate such a claim. This dearth of studies calls for additional research focused on methodological development that couples remote sensing methods with conventional survey techniques, potentially enabling the addition of remote sensing to the toolbox of social scientists not only in the field of development studies, but broadly in the social sciences as a whole.
Data and Methods
Sub-Saharan Africa is characterised by large variations in agronomic conditions and farming systems, even over short geographical distances. This poses a challenge in terms of generalising from localised data to wider areas or whole countries. Here satellite based fine-grained remote sensing is crucial for supporting decision makers with up-to-date information on people’s livelihoods and environmental conditions. However, satellite imagery is highly unstructured, requires costly ground-truthing (due to the mentioned variations) and does not easily translate into meaningful information about socioeconomic conditions on the ground. This issue can however be resolved through the use of new promising techniques in the remote sensing area that belong to the broader field of AI and through deep and transfer learning. Although we have data on household level, these data will be used to develop indicators at village level, like rates of poverty, intensification, commercialisation and pluriactivity. This implies that villages are our de facto data units. This is necessary to achieve in combination with remote sensing imagery and machine learning techniques. We will map the spatial distribution of village development over time. How have poverty and living standards more generally evolved in different localities? What is the role of drivers like intensification, commercialization and pluriactivity in this development? What is the role of policy? The latter questions will be resolved through multi-level regression analysis.
Image recognition and deep learning is a set of machine learning techniques that attempt to mimic human learning (i.e. from examples) by using already labeled images. Classical examples of applications are face detection and spam filtering. To be successful, massive amounts of training data is required and performance generally increase with more training data. For that reason large training databases are developed, such as AlexNet, with millions of labeled images. Here, an algorithm is trained to recognise target indicators, in this case from satellite imagery. However, training data of the form and volume required for this new task is not available. To compensate for this lack of training data, we will use transfer learning to combine knowledge from different sources to develop indicators of structural transformation and poverty. We build on the works of Xie et al. (2015) and Jean et al. (2016) but differ in the scope of living standards indicators used. As the first study of its kind, we include vegetation indices and night time lights in the training process.
- January 1, 2020, to December 31, 2022
- The Swedish Foundation for Humanities and Social Sciences (Riksbankens Jubileumsfond)
- Lund University
Project group at Halmstad University: