Towards Global Slum Mapping From Space: Detecting Urban Poverty Using a Transfer Learned Fully Convolutional Network

Informal settlements are the result of the continuing rapid growth of mega cities, especially in the global south where people migrate from rural areas in hope for a better future. Poverty is considered one of the major challenges to our society in the upcoming decades, making it one of the most important issues in the Sustainable Developments Goals as defined by the United Nations. These settlements, however, often lack basic sanitation and access to clean water. While many urban agglomerations of the global south are prone to large slum areas, still, the exact location and size of these settlements is often unknown. Remote sensing methods have improved tremendously in their capabilities of mapping informal settlements and its morphological features, which can be described by their high building density, small building sizes or its building orientation. But the challenge of large scale slum mapping still remains open, due to fuzzy feature spaces between formal and informal settlements, as well as a significant imbalance of slum occurrences where slums only account for 1% in the data. To tackle this issue we propose a fully convolutional xception network (XFCN). With its 34 convolutional and five dilated convolutional layers including four skip connections during the up-sampling phase the XFCN is capable of detecting poor urban areas in a coherent transfer learning approach using high resolution satellite images. This proves to be an ambitious task, differentiating between formal built-up structures and informal settlements at high resolutions. We train the network on a large sample of globally distributed slums (Cape Town, Caracas, Delhi, Dhaka, Lagos, Medellin, Mumbai, Nairobi, Rio de Janeiro, São Paulo and Shenzhen), greatly heterogeneous in its morphologic feature space and transfer the XFCN to map informal settlements. The XFCN is trained from scratch using 5 input channels and rigorous regularization. Using this approach we are able to reach an overall accuracy of up to 95%