An Uncertainty Aware Method for Geographic Data Conflation

With a significant amount of spatial data archives online, data conflation is becoming more and more critical in the domain of Geographical Information Science (GIScience) because of its broad applications such as detecting the development of road networks and the change of river course. Existing conflation approaches usually rely on the vector data of corresponding features in multiple sources to have an approximate location. However, they commonly overlook the uncertainty produced during the vector data generation process in the data sources. In previous work, we presented a Convolutional Neural Networks (CNN) recognition system that automatically recognizes areas of geographic features from maps and then generates a centerline representation of the area feature (e.g., from pixels of road areas to a road network). In this paper, we propose a method to systematically quantify the uncertainty generated by an image recognition model and the centerline extraction process. We provide an end-to-end evaluation method that exploits the distance map to calculate the uncertainty value for centerline extraction. Compared with methods that do not consider uncertainty value, our algorithm avoids using a fixed buffer size to identify corresponding features from multiple sources and generate accurate conflation results.