Measuring urban socio-economic disparities in the global south from space using convolutional neural network: the case of the City of Kigali, Rwanda

The rapid global urbanization, projected to be 68% of the world's population residing in cities by 2050, is accompanied by socio-economic disparities, especially in the Global South. Despite the acknowledged necessity for detailed, granular socio-economic data to comprehend these disparities, many cities in the Global South lack such data. This study addresses this data gap by supporting satellite imagery as an alternative data source, offering opportunities to investigate urbanization challenges at a granular spatial scale previously inaccessible through conventional census and survey statistics. Focusing on Kigali, the capital of Rwanda, as a representative city in the Global South, the study demonstrates the limitations of relying on Demographic and Health Surveys (DHS) wealth index data for predicting neighborhood-level wealth. Subsequently, it proposes an approach that combines human intelligence for labeling satellite images with DHS wealth index data and a less computationally based Convolutional Neural Network (CNN) technique. This approach achieves a notable 73% explanation of variations in neighborhood-level wealth. To enhance the interpretability of the model's predictions, Gradient Class Activation Mapping was used to identify features in the images that contributed most to model's basis for making decisions for prediction. This sheds light on visually interpreting the model's basis for prediction and could facilitate understanding how the model works for users who do not necessarily have machine learning skills. This study advances methodologies for socio-economic mapping using satellite imagery by underscoring the significance of combining human intelligence with machine learning in areas lacking reliable ground truth data and computing infrastructure.