Questions & Answers - From Historical OpenStreetMap data to customized training samples for geospatial machine learning

DOI: https://doi.org/10.5281/zenodo.3923040 Github:  https://github.com/GIScience/ohsome2label OSM Science Mailing list: https://lists.openstreetmap.org/listinfo/science

Questions

1. [Done] Does ohsome2label also georeference the results (i.e. land use vectors, building footprint vectors)?
2. [Done] Question from René: Thanks a lot for your talk! I’m wondering how your approach performs in areas with lower OSM data quality or with sparser coverage. Have you conducted systematic performance tests in corresponding areas? Would be interesting in support of your ongoing work in the Global South, in particular. Yes, it is :)
3. [Done] How do you ensure the imagery is exactly WGS84 with no skew before you start? Bing: bad, Maxar: often bad.-jidanni
   1. Bing is often up to 10 meters offset in rural areas! Basing work on it will really damage the resulting map.
4. [Done] In some works landuse/landcover information have been derived from geosocial media data, especially from shared photos. Do you see any potential for augmenting your approach taking account also of these data sets (in some areas) especially with regards to possibly achieving fine temporal resolution? []Do you think this could add anything to what you achieve already using OSM data?
5. [Done] Can you discuss more about your future plans for the OSM quality measurement? Would be interesting to know!
   1. Oh no I think it’s kind of choppy, I wasn’t able to hear.
   2. I’ll just discuss through email later, thanks! Thank you too for your participation!
6. Can you compare your work and the work of Development Seed with Label maker? I’m curious to know the similarities and differences.

Answer:

1. Yes, the result is in slippy map tile format.
2. It’s two question, for sparse coverage area, it works well since ohsome API offers a roboust feature query function. For the lower quality  area, some more post processing should be done for your training dataset.
3. We support customed tms, you can change the bad image with the corresponding image.
4. The geosocial media data is varies in different application, so it’s hard to handle it in a unified form. But I will think about it, or you can raise an issue on our github page. Thanks! My question was rather exploratory in nature, indeed. Hopefully inspirational though :)
5. First, we will try to do some tag-specified intrinsic data indicators, like the quality indicator for the building. Second, our next step is to study how the tile-based osm data quality influence the deep learning training progress.
6. For the Label maker, if you have used it before. You will find it’s hard to setup. For ohsome2label, you only need to type `pip install ohsome2label` then you can get it. The second difference is that we support some data quality analysis for your area of interest, and we will extend this function further. Third, we use MS coco format output and Label marker use npz. Fourth, we have the intermidiate output, so you can modified what you want, like add some feature to your geojson. Ah, we also support overpass api aside Ohsome API.

Thank you for your great question. You can connect us by email or on our github page any time. It’s our pleasure. My email is zhaoyan_wu@whu.edu.cn. Many thanks for taking time to address the questions; you were excellent!Thank you!

Comments

1. Award winning session chair. Thanks+1. Many thanks! :)+1 Should say “I’m XX, talking to you from YY (country)”.
2. audio on video is over blown.+1
   1. Noted and thanks, video team will have a look.
      1. (Might be a problem of the recording / microphone rather than the stream.)+1
      2. (had to turn down the audio source as it was too loud complared to live source.)