Paved with Good Reflections: Road Quality Mapping with Multispectral Satellite Imagery

The advent of multispectral satellite imagery with high spatial and temporal resolution has opened new avenues for monitoring and mapping infrastructure, including electricity, building, and transportation systems. This is particularly crucial in the developing world, where researchers and development practitioners face challenges in obtaining accurate, timely, and comprehensive infrastructure data necessary for various studies, such as monitoring development indicators and conducting impact assessments. The difficulty in accessing ground truth data also complicates the development of models required for mapping infrastructure accurately in these regions. In this study, we address these challenges by leveraging OpenStreetMap road data in conjunction with satellite imagery from Sentinel-2 and Landsat 8 to develop machine learning models for classifying road surfaces across Sub-Saharan Africa. We achieved classification accuracies of approximately 80% with Landsat 8, and 92% with Sentinel-2 when differentiating between paved and unpaved roads in the region. With the developed model, we estimated the share of paved roads in countries in the region and demonstrated the correlation of this road quality metric with other widely used metrics for comparing road quality across countries, achieving correlation coefficients as high as 0.71. We further show the capability of the proposed approach for conducting multitemporal road surface assessment over time from 2019 to 2023 with Sentinel-2 satellite imagery. These findings highlight the potential of using remotely sensed data in conjunction with crowdsourced data for comprehensive infrastructure analysis in data-poor regions.