Cover Story: Improving OpenStreetMap missing building detection using few-shot transfer learning in sub-Saharan Africa

Current Limitations

The team identified three limitations for this case study.

1. The current work focused on areas where OSM buildings were completely missing and future research is needed to validate the proposed method in areas where buildings are partly mapped in OSM. In those partly mapped area, one could directly consider existing OSM buildings as the FSTL training shots and seek to assess the relative completeness of OSM using a similar DeepVGI layer.

2. As the proposed method was implemented and validated only in sub-Saharan African, a global overview of such a method applied in different geographically remote rural areas (e.g., urban contexts, climate, vegetation, settlement types) is still missing.

3. Both the current Faster R-CNN and the SSD model only detect bounding boxes of buildings but not their footprints. Further research is encouraged to extend the FSTL method toward the automatic extraction of building footprints in vector formats, for example, using the Mask R-CNN architecture (He et al., 2017) and with post-processing similar as (Sirko et al., 2021).

Future Research Avenues

To address the aforementioned limitations, future works are encouraged to adopt or extend the FSTL method proposed in this article and further investigate several research directions.

1. Besides assessing the relative completeness, the availability of ML-generated layers (e.g., DeepVGI) opens new opportunities of revisiting classic OSM extrinsic quality assessment methods (Hagenauer & Helbich, 2012; Hecht et al., 2013), which commonly rely on external reference datasets.

2. Thanks to the contribution of volunteers, the large amount of existing finished humanitarian mapping projects worldwide, for instance from MapSwipe App and HOT Tasking Manager (Herfort et al., 2021), provides a great overview of vulnerable rural areas with a pressing need of mapping in OSM. One can think about scaling up the FSTL method to these remote areas and further investigating the impact of different types (or levels) of urban contexts on the model performance.

3. By thinking about how ML methods can best benefit OSM, especially in rural areas, recent discussion in Mooney & Galvan (2021) and Vargas-Muñoz et al. (2020) shared important insights from either remote sensing or OSM community perspectives. As an intuitive application, one can apply the FSTL method to estimate the overall time and effort required to map all OSM missing buildings, then take these factors into consideration when organizing humanitarian mapping campaigns and volunteer contributions.

4. The previous work in Herfort et al. (2019) reported an inspiring finding that advanced ML algorithms can potentially improve the accuracy and accelerate the speed of current humanitarian mapping method using the MapSwipe App. Therefore, another potential application of the FSTL method could be assisting human mappers in rural areas by giving them mapping hints based on machine-generated results, such as an early attempt with the RapiD tool from Facebook (RapiD, 2019).

In this context, the team aimed to emphasize the necessity of machine-assisted mapping with the lessons learned in this article. Our future work will seek to develop intelligent OSM mapping workflows by keeping humans and machines in a loop.

In A Sentence

In conclusion, the lessons learned in this article sheds light on facilitating and accelerating current humanitarian mapping workflow by efficiently reusing existing ML models via few-shot learning and accurately detecting OSM missing buildings in rural sub-Saharan Africa areas.