[HOT] Building Detection using Machine Learning
phunt123 at icloud.com
Mon Dec 19 14:17:55 UTC 2016
I attended my first Humanitarian OpenStreetMap Team (HOT) mapping event a few months ago and was interested to see how successful machine learning would be at detecting buildings in satellite images. The results look promising but I wanted to know if it could be useful to the community and if it’s worth pursuing further. I thought I would post a sample of the results and then quickly explain the process and issues.
These are the results of a test I ran on project 2101 (Rongo, Kenya - PMI/USAID) on 1 November 2016. These images show the buildings detected by the algorithm on the first six unstarted tasks from the project. Potential buildings are marked with green rectangles:
As you can see the initial results look promising - most of the buildings have been detected and the false positive rate is pretty low.
I’ve been using the Viola–Jones machine learning algorithm, which requires training to know what is and isn’t a building. Once the algorithm is trained, it can be used to detect buildings in new images in a few seconds.
The whole process looks like this:
- Get the HOT project and task data using the HOT API
- Get the satellite imagery of the area from OSM
- Get the nearby existing buildings from the OSM API
- Find the existing buildings in the satellite imagery and use these to train the algorithm
- Run through each incomplete task in the HOT project and detect buildings
- Output the results as OSM XML
- Load the output into JOSM, validate and upload to OSM
I loaded the output of the algorithm into JOSM and completed tasks 1 and 2 of project 2101. However it still took a bit of work to make sure the data is good enough for OSM and I think an experienced mapper would have taken roughly the same amount of time starting from scratch.
The main issue is the algorithm can’t rotate the detected rectangle to fit the building shape (as you can see from the example images above, none of the rectangles are rotated). I’ve tried using methods such as line detection to detect the building and rotate and crop the rectangle around the edges - this worked well some of the time and other times went horribly wrong.
The second issue is false positives. While the examples above we’re generally clean, sometimes the algorithm would think a field was a building. Because data uploaded to OSM needs to be accurate it can take some time checking each potential building in JOSM.
Another potential issue could be training samples. When testing I trained a new algorithm for each project, using local existing building data from OSM as training data. The assumption here is that nearby buildings will look like buildings in the project area and that nearby building data is available and accurate.
The Viola–Jones objection detection research paper was first published in 2001 so the algorithm has been around for a while. Machine learning has improved since then and neural networks are showing a lot of promise - using these could increase the reliability and also allow the detected rectangle to be fixed around the edge of the building - meaning a lot less editing in JOSM. I’m also aware of similar projects, but I haven’t found anything that’s able to detect buildings or ready for use yet:
https://github.com/trailbehind/DeepOSM (find misconfigured roads in OSM)
https://github.com/patrick-dd/landsat-landstats (predicts population size)
I'm not suggesting this could replace volunteers (since algorithms will never be completely accurate), but maybe this could help speed things up or be used to quickly estimate building locations over large areas.
Please let me know your thoughts. Could this could be a useful tool for HOT or any other volunteer organisations and is it worth taking any further?
More information about the HOT