AirWorks is a service that uses AI (machine learning) and computer vision technology to take the drudgery out of making CAD drawings from aerial and satellite images. Anyone who has had to create CAD drawings from aerial photos—or worse, Google Earth imagery—should appreciate the assistance that AirWorks provides.
We had to find out how much assistance AirWorks can actually provide, so we zoomed in on JP Buckley working from his Boston-area home office. Was AirWorks magically able to extract details out of fuzzy photos, guess correctly where roads and lane marking were even if obscured by trees, and so on, and roll out DWG files with little or no human intervention?
Not quite, we are to find out. AirWorks is more of a full-service shop and not so much push-button magic. There is an AI-assisted pass over your satellite, aerial photo or LiDAR data that does a commendable job interpreting pixels as vectors—the stuff of CAD—but to finish the job will require some human effort after all.
“We like to stay away from using ‘magic,’ said Buckley, a rare director of sales who is able to resist overselling. He explains how AirWorks may be considered an aid, an assistant with AI—one that will save considerable manual work—but not quite all of it.
“What we’ve done is used machine learning and computer vision technology that allows our clients to import a variety of remote sensing data sources, whether that’s from aerial or satellite images,” explained Buckley. We take their datasets, and we make sense out of them and we output a CAD file for them. We’ve built autonomy around feature extraction, so our AI will take the imagery or the LiDAR, for example, and it will extract certain features like roads, buildings, pavement … and create objects in a CAD file format. We also are able to process point clouds and create surface files for surveyors and engineers.
“This campus [shown above] took us 5 days,” continued Buckley. “It would have taken our client five months. The machine learnt algorithms take about two hours of cloud-based number crunching. AirWorks uses Amazon Web Services.”
The rest of the five days is spent on what Buckley calls “quality control” but what sounds more like manual post-processing of all the imagery the AI could not identify and classify. AI is like the politician running for office: full of promises that go unfulfilled. Still, AirWorks delivers assistance, if not a complete, hands-off solution, that can give back days or weeks to surveyors and engineers, most of whom are inexperienced with and ill-equipped for the overwhelmingly massive sets of data that a camera-equipped drone gathers.
The lack of manpower available to meet massive datasets has created a bottleneck, says Buckley. Compounding the problem is an aging workforce.
“The average age surveyor is 60 years old and there’s not many coming out of school going into surveying,” said Buckley. “The young engineers are eager, hungry, very skilled. Should they be bogged down connecting the dots on aerial or satellite images? No, they will want to use technology, use AI to really speed up the process.”
“We’re good at interpreting the continuation of a curb underneath tree cover,” said Buckley. Where it cannot interpret features, because the tree coverage is too dense, AirWorks marks off the area so that a request for further inspection by a ground crew can be issued.
The company spent its first two years doing the machine learning, mapping pixelated images to objects, creating a library of tens of thousands of common shapes, says Buckley.
“The model will make a prediction, or it will learn what we’re trying to predict,” explained Buckley. “We will take those corrected images or what the AI predicted and compare that to what it should be, then we feed that back into the model. We’re constantly retraining our models.”
The accuracy of the interpretation varies from feature to feature. Each feature, such as a building or vegetation, will have its own layer, its own algorithm. The ability to detect and work around vegetation is the most accurate (“90th percentile”). Other features, like buildings, solar … they can be between “70 percent to low 90 percent accuracy.”
The more datasets that AirWorks is able to work with, the more accurate it gets, according to Buckley.
Dismissing Google Earth images as not detailed enough is less a case of looking a gift horse in the mouth than it is the reality of being unable to distinguish one detail from another. “Google Earth is coming in at around a 60-centimeter resolution. That’s pretty low,” explained Buckley. “You’re not going to be able to identify things like fences, signs, mailboxes, water, gates, manholes, drains, etc. Even the items you can see—the buildings, the vegetation, the roads—the pixels are so large that tracing over the imagery is going to be inaccurate. Surveyors and engineers need centimeter, even millimeter accuracy. A building could be three feet off if you have traced off of pixels that represent 23 inches each. The high-resolution imagery we use has pixels that represent half an inch each. Our AI is able to trace accurately, and we can deliver to with one-tenth of an inch. You can build off of that. Working off LiDAR, we can get even tighter.”
AirWorks takes Google Earth images, orthorectified photos, photogrammetry, and using machine learning and computer vision, tries to figure it all out, that is, interpret the images into curbs, buildings, mailboxes, roadways, power lines and so on.
AirWorks charges a base amount of $32.67 per acre for the AI pass. A more densely packed acre will cost more. The client can select only the area of interest to be processed to keep the costs down. If additional work is required, such as extra layers with geometry that must be manually determined, the extra cost is instantly calculated per layer.
AirWorks was founded in 2017 by David Morczinek and Adam Kersnowski. The German-born Morczinek is an aerospace engineer from the manufacturing side of Airbus’ large cargo aircraft who is also a trained pilot. He went on to get an MBA at MIT, where he met Kersnowski, who had a 13-year background in construction and a passion for drones.
Drones were used in proposals for buildings,” said Buckley. “He really started using them all over the place—in his personal life and then for a company. That segued into a venture with his own drone network of drone pilots that operated all across the country. At that time, you needed pilots to fly drones.”
When people needed a drone, they would call on Kersnowski’s company. Most of his customers were surveyors and engineers. They were quick to adopt the new technology but just as quickly found that the that datasets were unmanageable.
“There were tons of data, hundreds of millions of points,” related Buckley. “Engineering firms were not equipped to handle the data, so they would ask him to process the files. Their computers were not powerful enough, their drives not big enough. Their systems would crash. They would ask, ‘Can’t you just handle it? All I need is the line drawing. Or the surface.’ In other words, just give me the deliverables. ‘I don’t need the millions of points.’”
And so, the mission started to form: to build an AI-based solution for this purpose, make AI do all that it can to free civil engineers and surveyors from having to process the big datasets, the images, and deliver, quicker than humanly possible, the line drawings or surface models that are the end product, the deliverables for which they get paid. They saw the sensors in drones and satellites were getting better, cheaper and becoming more plentiful. All that was going to generate a lot of data and require a lot of manpower to process.
“The trend was clear, but no one had cracked the problem,” said Buckley. “We were drowning in data and manpower was getting scarcer. Here’s where AI would fill the labor shortage by making up the missing piece, to make raw outputs into CAD deliverables.”
The company has raised $5.3 million in funding, according to Crunchbase.
The Future of Data
At present, if a surveyor or engineers needs detailed imagery, they must procure an aerial service for it. But AirWorks sees a future where there is abundant imagery widely available in an immense, publicly available source.
We may have been spoiled by Google, which very generously provides images of the Earth, stitching together a mosaic of all the Earth’s surface from satellite images. Thanks, Google, but your images lack detail, say surveyors or civil engineers. Google also maintains an enormous fleet of ground vehicles with 360-degree cameras that have filmed practically every street on Earth. Thanks, Google. But a street view is insufficient. You can’t see a green or brown field from the street. For that, we’ll have to fly drones.
AirWorks has high hopes for Amazon, which may very well darken the skies with delivery drones equipped with cameras and possibly choose to share its images with the rest of the world—despite Amazon having no history of generosity such as we take for granted from Google. Perhaps late model cars, with all their cameras, radars and sensors, will be willing to share? Or from on high, the air taxis that are sure to come. Hope springs eternal, and one can certainly dream of abundant, clear, detailed images from an increasing number of cameras and sensors our technology promises.