In the second and final part of our Q&A with software engineers Jim Clark and Miles Carlsten, share more on how machine learning is changing the way Trimble MAPS improves map automation by leveraging leading-edge data.
Q: How has machine learning changed the way you build maps?
Miles: There’s so much we can learn from the satellite machine learning model alone and all the output we can get from that to apply to maps. Some examples are finding missing roads, bogus links, road width measurements, turn radius data, and driveway detection. Machine learning can help us do a better job of addressing existing problems, like the best way to find missing roads.
But it can also allow us to do things that weren’t possible before, like detecting driveways and turn radii. Without machine learning, there’s no road width data, and you need that to analyze sharp turns. Driveway detection lets us automatically calculate extremely accurate arrival locations for addresses.
Jim: Machine learning is helping us address issues like roads that were planned but never built. That’s a pretty common occurrence that can create a bogus link in map data. Humans can find that information, but they’d have to search through thousands of maps to flag a bogus link for removal. We’d rather automate that sort of process and all “easy” tasks that we can solve algorithmically so our GIS operators can use their specialized knowledge on more complex tasks.
Q: What kind of automated map improvements are we already using? What can we expect in the future from machine learning?
Miles: We already have some tools created for our GIS team that utilizes a huge GPS breadcrumb database, our road network, and a simple machine learning model based on satellite imagery that help to identify missing roads, bogus links, incorrect one-ways, and missing or incomplete sites. Our GIS team has made hundreds of thousands of map edits based on the feedback from automatically generated lists. We've got lots of other projects in the works, including a new model for satellite imagery that will improve the lists we generate for missing roads and bogus links. Our road network data is already very good, but we’re working on a new model to make it great with better output.
This improved satellite imagery model will be a cornerstone we can apply in so many ways, including enabling new tools like analyzing road widths to identify roads that are potentially in the wrong class, and looking at intersection shapes for the turning radius. We've also got the CoPilot Data Analytics project, which lets drivers opt to send us data. That allows us to identify the areas where they collectively ignore our guidance and figure out why. Sometimes this can reveal routing issues on our end or data issues such as missing restrictions.
Jim: One thing to keep in mind about machine learning is that it can allow us to make improvements that just aren’t feasible otherwise because you’d have to hire 10,000 humans to do it. There are many improvements we can make using machine learning that is completely impractical in terms of labor without it.
In addition to the projects that are already underway, we could take a look at new models that analyze street view images and identify important objects like stoplights and stop signs. Machine learning models could analyze speed limit or weight/height restriction signs on the side of the road and other tasks that could add to the dataset to help us operate more efficiently and deliver more accurate and up-to-date data to customers.
Q: How do you see the process for keeping Trimble MAPS data leading-edge as we move forward?
Miles: We’re already providing lists to our GIS team to help them update roadway information, and we’ll do more of that in the future. For example, road data from counties and other sources may be out of date or incorrect, and it’s a manual, labor-intensive process for GIS to update that. Our team provides lists, like the missing roads or bogus links list, to assist them.
But pointing out where problems are is just the beginning — eventually, we’d like to build on those lists to help solve the problems with automation. For a missing road, to cite one example, we'd want to automatically draw a road that GIS edits, as opposed to alerting them that a road is missing and having them do all the work. At some point, we'd also want to propose road attributions like name, class, etc. And then, when we become more confident in our results, we could eliminate manual review.
The idea is to use automation to free up the GIS team so they can work on more complex data initiatives. When the easier tasks are automated, the GIS team can focus on improving the information that already sets Trimble MAPS apart from competitors and delivering even more value to customers.
Jim: Another way machine learning can help us stay on the cutting edge is by using sources like satellite imagery to make edits even faster and proactively find and deliver relevant information to customers. For example, we might use technology to update map information before it’s corrected in county data. We can also find things like housing developments as they are being built, which can help customers who are delivering things like building supplies to an address before it even appears in a county source.
Special thanks to Jim and Miles for taking the time to share their insights.