Automatic detection of human settlements in rural Sub-Saharan Africa from satellite imagery with convolutional neural networks and OpenStreetMap

Abstract

Spatial data play a critical role in both long-term planning and sustainable development as well as immediate emergency response. In Sub-Saharan Africa, a region struggling with extreme poverty, they are essential for effective humanitarian action. The advancements of deep neural networks in the past decade have opened great possibilities for generating this data in an automated way. However, neural networks require large amounts of training data. In regions where the data is not provided by national agencies, crowd-sourced mapping platforms such as OpenStreetMap have been proposed as an alternative.

In this thesis, convolutional neural networks were trained on satellite imagery and OpenStreetMap data to detect buildings in rural Sub-Saharan Africa. Multiple models were trained on different data sets and with different hyperparameters. Performance of all models was assessed and compared. Additionally, a method for fine-tuning trained models to new geographic areas that requires only a small amount of additional data was proposed and tested in multiple settings.

Using the presented methodology, buildings were detected in 16 test areas across Tanzania, Zambia and Malawi with average f1 score over 0.7. Small buildings and densely populated areas presented a challenge to all models. The fine-tuning method was successfully used to adapt a model to a region in Cameroon more than 2600 km away from where it was trained. Additionally, fine-tuning a model improves its performance in other areas as it benefits from learning from new data. The findings of this thesis can be used to assist in humanitarian mapping, especially in identifying areas where human settlements are missing in maps and estimating with high accuracy the amount of missing buildings.