US Sports Fields Import 2018
Phase 1 - January 2018
Data Creation In November of 2017 all of the mapped baseball, basketball, and tennis courts in Massachusetts, New York, Connecticut, Rhode Island, and Pennsylvania were used to train a neural network. The network identifies baseball, basketball, and tennis courts in Bing satellite images.
The import data was created by
- Running the neural network over the training satellite images to identify the missing sports fields from OSM. Fields found by the network but not in OSM are passed to the next step.
- A visual QA script shows the newly mapped features over the Bing satellite images, the user approves or rejects the edits. I rejected any way that wasn't an excellent match to the satellite images.
- The final OSM files are created from the ways selected by the user in visual QA script.
This initial pass successfully identified 2,800 new sports fields.
Tagging The ways are tagged with leisure=pitch, plus the sport=* tag.
Changeset Tags will be comment="Imported data, phase 1, https://wiki.openstreetmap.org/wiki/US_Sports_Fields_Import_2018", and source="Bing", import=yes
Uploading - The jremillard-import account will be used to upload the 2 osm files. The OSM files are already split to be under the 10,000 item limit on a changeset. They will be uploaded with josm.
License - Prior to starting the project jremillard emailed Microsoft bing licensing group describing the project and received approval to use the Bing satellite images for this purpose. Also, the QA script qualifies as a visual editor in the existing OSM Bing license. The two final output files are in the public domain.
Source Code - All of the source code used for the import is on Github Images To OSM.
Import Files - The Data OSM Files.
Post Import Cleanup - Some imported ways may overlap with existing OSM data. Specifically, any sport fields added since November, sports fields that are tagged incorrectly, and sport fields that is a relation might overlap with the uploaded data. After the files are uploaded, all of the ways, and relations with the sport= tag will be checked for overlaps and manually fixed. It is expected that just a handful of issues will be found in the post cleanup.
Phase 2 - Spring 2018
Phase 2 will be repeating the same process as phase 1, but with a retrained network using the additional 2,800 features added to phase 1.