deepLNAfrica

Motivation

In sub-Saharan Africa, more than 50% of urban residents reside in informal settlements, and are projected to number 1.3 billion people by 2050. Identifying and mapping the locations of informal settlements at scale and tracking their development over time is thus crucial for adequate policies to alleviate urban poverty and inequality. However, given the rapid and often unplanned urbanization dynamics in African cities, existing monitoring systems are insufficient for tracking the development of informal settlements timely and accurately. This project aims to alleviate these problems by introducing a cost-effective data-driven solution based on open-access Earth observation data that allows to accurately map informal settlements across space and time.

Proposed Approach / Solution

‍We use open-access Sentinel-2 satellite imagery with machine learning in order to learn the visual characteristics of informal settlements, as seen from above. First, unlabeled satellite images are used to pretrain a semantic segmentation neural network to become invariant to visual effects irrelevant to the task of informal settlement detection, such as those caused by atmospheric distortions, sun angle and acquisition geometry, weather conditions, and seasonality. In a second step, this model is fine-tuned on ground truth data from 16 cities within 6 different sub-Saharan countries that spatially describe informal settlements. To tackle the difficult challenge of mapping informal settlements in cities and countries where no training data was available, many different models are trained on subsets of the data and their predictions ensembled for increased robustness and decreased bias during inference. Using this setup, informal settlements were mapped in all 529 sub-Saharan African cities with a population of at least 100’000 for the years 2016 and 2022. See Figure 1 for an example prediction in a single location and point in time, and Figure 2 for our aggregated predictions for the share of the urban population living in informal settlements across all sub-Saharan African countries.

Impact

‍The solutions proposed in this project bring great improvements in spatial and temporal coverage of urban informal settlements when compared to established methods. Furthermore, it provides first-time estimates of the share and number of people living in informal settlements across all sub-Saharan African cities exceeding a population count of 100’000. This project demonstrates how the mapping of urban informal settlements could be revolutionized in SSA and across the world but also underlines a great need for standardized high-quality ground truth data.