A global typological framework for mapping neighborhood variation using machine learning and remote sensing.

Humanitarian organizations face logistical challenges when trying to reach impoverished or marginalized communities. These people can be hard to locate and are often poorly accounted for in public records. One way to improve access to these communities is to better map their location and extent. Despite an increasing demand for this information, neighborhood map data is still scarce across most of the developing world.

To address this problem, researchers at Oak Ridge National Laboratory use machine learning - principally, factorization-based texture segmentation - to map neighborhoods using remote sensing at the city-scale. This method successfully differentiates settlement patterns within a given metropolitan area based on specific texture classifiers and has been successfully implemented in multiple countries, including DRC, Ethiopia, Iraq, Kenya, Nigeria, Senegal, South Africa, Tanzania, Yemen, and Zambia. Still, this modeling approach has practical challenges - most notably, its reliance on city-by-city model training - and is therefore difficult to administer across varying contexts.

Efforts to broaden the scale of this project using more advanced computational methods require a more generalizable typology for categorizing neighborhoods. This study presents a global framework for typing settlement patterns that could be used to dynamically inform machine learners without the need for context-dependent model training. This framework categorizes buildings within a given satellite image among 8 generalized neighborhood types based on observed patterns in building size, density, material, orientation, and land-use. In practice, this framework's application still suffers from constraints imposed by availability of adequate imagery and variability in settlement patterns. When available, supplemental data such as building footprints, road networks, or LiDAR may help overcome these issues.

KEYWORDS: machine learning, neighborhoods, typology, population, land use, remote sensing