Modelling land values in African cities with open data and machine learning

Understanding the geography of land values in cities is essential for urban science, revenue mobilization and public policy. Yet most African countries lack accurate and up-to-date information on urban land values and rents. An accessible, affordable and scalable method of land value estimation could enhance both research and urban development policy in African cities, supporting achievement of Sustainable Development Goals 9 and 11. Here, we develop and validate an approach to estimating unit rents per square meter at high spatial resolution using open data and machine learning in two large African cities: Accra, Ghana, and Nairobi, Kenya. Hedonic price models traditionally combine data on location (e.g. proximity to amenities and transportation infrastructure), neighbourhood characteristics (e.g. socioeconomic conditions) and intrinsic property characteristics (e.g. size of unit, age of building) to estimate rental prices. In the absence of reliable and accessible data on these features at the building or unit scale, we use OpenStreetMap (OSM) data and features extracted from freely available satellite imagery with a convolutional neural network to predict rents per square meter. For target data, we collected rental prices for over 8,600 properties across Accra and Nairobi. We use these data to train and validate random forest models for each city, successfully explaining approximately 60% of the variation in rental prices across samples from each city. This low-cost, scalable approach could be used to advance urban science and improve revenue mobilisation across the continent.