Recently a new paper about Estimating OpenStreetMap Missing Built-up Areas using Pre-trained Deep Neural Networks (DNNs) has been presented at the AGILE GIScience conference 2019 in Cyprus.
Although built-up areas cover only a small proportion of the earth’s surface, these areas are closely tied to most of the world’s population and the economic output, which makes the mapping of built-up areas a vital challenge. Thanks to the generous contribution of volunteers, OpenStreetMap (OSM) shows great capability in addressing this challenge, while the missing of maps is still a major concern. In this study, we propose a built-up areas mapping method by fine-tuning pre-trained deep neural networks (DNNs), which aims to estimate OpenStreetMap missing built-up areas in a large-scale humanitarian mapping scenario, as e.g. for the MissingMaps project (Scholz et al 2018). Specifically, we train an object detection network using very high resolution satellite images and corresponding OpenStreetMap building features. Then, we employ task-level labeling algorithms to produce the built-up estimation results and compare their accuracy performances with state-of-art baseline data sets. Considering the model transferability during scaling up to larger areas, we select two geographical independent areas in north Tanzania, Africa, for training and testing, respectively, where finished MapSwipe projects are available. Experiment results confirm that the pre-trained networks could yield high quality built-up maps and competitive estimation performances, which lead to over 75% of missing areas detection and over 92% of estimation overall accuracy.
In detail, we propose the Faster R-CNN+TLA method to estimate OSM missing built-up areas in Tanzania. Firstly, fine-tuned on very high resolution satellite images and corresponding OSM training samples, the proposed method could generate accurate built-up areas maps. The preliminary results in Tanzania show that our method could significantly reduce the missing building tasks by over 75% and achieve around 85% F1 score. Next, we evaluate the estimation performance of the proposed method regarding OSM missing built-up areas, which leads to competitive estimation results (over 92% OA) in comparison to the crowdsourced MapSwipe baseline. It is worthwhile to develop a machine volunteer collaborating workflow by combining both methods, especially when considering large-scale estimation in heterogeneous regions.
More importantly, we intentionally select independent train and test areas in order to develop robust and transferable DNNs, which could be easily implemented to unmapped areas. Nevertheless, the transferring capability of DNNs deserves further study. In the future, we would include more study areas and adopt different pre-trained DNNs. We would also investigate the factors that may affect the model transferability, such as population density, land use land cover, and geographical terrain.
Towards integrating deep learning methods into the crowdsourcing applications for more intelligent workflow, our DNNs based method provides important insights into various applications. For instance, the accurate estimation of missing built-up areas could help project managers in humanitarian organizations to make better plans before starting a new campaign. Moreover, the integration of machine-generated and crowdsourced data might significantly accelerate the mapping procedure while achieving similar or even higher accuracy as crowdsourcing workflow. The detailed workflow would be investigated in future work.
Li, H., Herfort, B., Zipf, A. (2019): Estimating OpenStreetMap Missing Built-up Areas using Pre-trained Deep Neural Networks. Proceedings of the 22nd AGILE Conference on Geographic Information Science, Limassol, Cyprus.
Acknowledgement: This work has been supported by the Klaus Tschira Stiftung (KTS) Heidelberg.
Selected earlier related work:
Chen, J., Y. Zhou, A. Zipf and H. Fan (2018): Deep Learning from Multiple Crowds: A Case Study of Humanitarian Mapping. IEEE Transactions on Geoscience and Remote Sensing (TGRS). 1-10. https://doi.org/10.1109/TGRS.2018.2868748
Chen, J., Zipf, A. (2017): Deep Learning with Satellite Images and Volunteered Geographic Information. In: Karimi, H. A. and Karimi, B. (eds.): Geospatial Data Science: Techniques and Applications. Taylor & Francis.
Scholz, S., Knight, P., Eckle, M., Marx, S., Zipf, A. (2018): Volunteered Geographic Information for Disaster Risk Reduction: The Missing Maps Approach and Its Potential within the Red Cross and Red Crescent Movement. Remote Sens., 10(8), 1239, doi: 10.3390/rs10081239.