Improving OpenStreetMap missing building detection using few‐shot transfer learning in sub‐Saharan Africa

Abstract

AbstractOpenStreetMap (OSM) has been intensively used to support humanitarian aid activities, especially in the Global South. Its data availability in the Global South has been greatly improved via recent humanitarian mapping campaigns. However, large rural areas are still incompletely mapped. The timely provision of map data is often essential for the work of humanitarian actors in the case of disaster preparation or disaster response. Therefore, it has become a vital challenge to boost the speed and efficiency of existing humanitarian mapping workflows. We address this challenge by proposing a novel few‐shot transfer learning (FSTL) method to improve the accuracy of OSM missing building detection. We trained two popular object detection models (i.e., Faster R‐CNN and SSD) in a training area in Tanzania and transferred the model to target areas in Cameroon and Mozambique. The FSTL method significantly improved the base model performance even with only one training shot. Moreover, we successfully produced a grid‐based OSM missing building map (DeepVGI) of 10 m spatial resolution with over 96% overall accuracy (ACC) and 0.85 Matthews correlation coefficient (MCC) in both Cameroon and Mozambique. Such maps show great potential to assess and estimate the overall completeness of OSM buildings to support humanitarian mapping activities, especially in places where other (e.g., buildings, roads) datasets are not available.