Developing an Automated Algorithm to Detect Rural House Roofs in Kenya using Fine Spatial Resolution Images

The success of the United Nations Sustainable Development Goals relies on temporally frequent socioeconomic data collection. High-resolution satellite data can be used to detect roofs and provide an insight into local poverty dynamics and the progress of development interventions. This study developed two algorithms to detect metal and thatched roofs in high resolution imagery of Sauri, Kenya; one fully unsupervised, the other texture-optimised, which were tested for applicability across sites, years and platforms. Accuracy assessments show an unsupervised approach can achieve an accuracy of 84.8% and utilizing texture increases this to 85.7%. When applied to a different year, accuracy of metal roofs was 84.0%, providing opportunity for change detection analysis, however thatched roofs were more difficult to detect. When applied to a different region in Senegal, accuracy fell, but methods to increase this applicability were explored. Ultimately, the algorithms require only the statistics of the image to be widely applicable and provide data for development monitoring however customizations are necessary to achieve the highest accuracy