AI aerial imagery solving tomorrow's challenges
Dec 2021
Mike Bewley talks about artificial intelligence and machine learning for aerial maps, looking at how Nearmap integrated models deliver efficiencies.
Dec 2021
)
Dr. Michael Bewley, Nearmap Senior Director of AI Systems met with The MapScaping Podcast in November 2021 to discuss Collecting and Processing Aerial Imagery at Scale.
Mike shared the Nearmap approach to Artificial Intelligence (AI) and Machine Learning (ML), looking at how an in-house, integrated model delivers greater efficiencies.
Here’s a snapshot of the podcast episode.
The Nearmap distinction
Mike explained what sets the Nearmap system for AI and machine learning apart: “We do AI at the same scale that we do imagery — whenever we process our surveys, we’re processing it through our AI system.”
With every survey, Nearmap customers can access 200+ AI facts about an address — such as counts of objects, areas and other metadata about vegetation; solar panels; swimming pools; asphalt; lawn grass; construction sites and construction vehicles; roof type, shape, material and condition; light and power poles, and more.
While the Nearmap system is fully in-house — from the systems that capture and process the imagery, to the custom labelling tool, machine learning model architectures, data processing, and tools to extract and interact with the data — Nearmap APIs allow easy integration with other platforms.
)
“We can optimise for every single thing we know about our own data — which is a lot — we’ve been working with our aerial imagery for a decade or more,” said Mike.
“It’s quite amazing the level of R&D that we’ve leveraged from the world of deep learning on imagery.”
Michael Bewley,Senior Director, AI Systems, NearmapMike explains that the advantage of automating the processing of billions of features on millions of square kilometres worth of 5.5—7.5cm pixel imagery at scale (instead of having to run algorithms one at a time) is that the model learns faster and data is available to customers sooner.
Looking for anomalies
Any machine learning model is only as good as the data it’s fed. “We run at such incredible scale that it gives us advantages that wouldn’t otherwise be possible,” said Mike.
Because Nearmap processes vector maps at a massive scale, anomalies can be either validated or rejected, to label more and feed into the deep learning system.
One example Mike described was the machine learning model identifying a swimming pool overlapping with water. At first sight, this would seem to be an error. On further investigation the item turned out to be a swimming pool, on a cruise liner, at sea — which enabled verification of the ‘water on water’ anomaly.
)
“We have focused on an incredibly high degree of accuracy in our models, so you get results you just can’t get by hand-labelling, especially with more than a million images worth of labelled data,” said Mike.
The future of AI feature classes
Enhanced feature classes help guide AI detection, with each AI pack considering multiple layers. For example, to label swimming pools the model may include attributes such as above- and below-ground swimming pools, empty pools or pools under construction, spas, hot tubs and paddling pools.
“If you do it right, the machine learning model can actually learn all those nuances,” said Mike. “But it can only do that if the humans [labellers] are consistent with those nuances.”
To see how Nearmap defines feature classes — such as swimming pools — visit our Help Center.
Consistency and alignment
Because Nearmap AI is aligned with Nearmap aerial imagery, the AI attributes align perfectly with the vector map, removing many of the challenges associated with getting different geospatial data sets to line up — both spatially, and for a particular point in time.
“A lot of local governments are trying to move from hand-digitised maps, that are really accurate, but are hard work and quickly become out of date,” said Mike.
For local government agencies and councils — like the City of Ryde in Sydney — the ability to see how areas have changed over time, comparing variations in vegetation cover and non-porous spaces (asphalt, concrete), helps inform better planning decisions to predict and manage urban heat islands.
To keep up to date and identify what’s happening in a local area, at scale, Mike says the only way to achieve that is with automation.
AI and automation
The focus of machine learning is that instead of a human adjusting the algorithm and writing a program, it’s programmed through the data. “My solution is not to fiddle with the algorithm,” said Mike. “The solution is to label more data in that scenario so that the model can learn from it.”
For those who have been working with algorithms for a long time, that is a big shift in thinking.
Emphasising the importance of good data, Mike summarises: “Data at massive scale with machine learning always beats a hand-rolled algorithm. Clean labelled data is the absolute key – and then you can train a model and run that at scale.”
Want to learn more about Nearmap AI? Watch the Nearmap AI Workshop presented by Dr Michael Bewley at Nearmap NAVIG8 2021.
Learn more about accessing deeper insights with Nearmap AI.