The swimming pool capital of Australia is the Perth suburb of City Beach. Just under 60 per cent of properties have their own pool, a remarkable statistic considering most dwellings are minutes from the ocean.
“Right next to the beach. It's phenomenal,” says Michael Bewley, director of AI systems at aerial imagery and data firm Nearmap.
Nearmap has been building taking high resolution photographs of Australia from above for more than a decade, using special camera equipment carried on light aircraft. More than 25,000 square kilometres – home to around 88 per cent of the population – is captured in the imagery, which is updated up to six times a year.
More recently huge swathes of the US and New Zealand are being photographed too.
Organisations use the imagery in a number of ways. Mobile cafe Tasty Fresh Food uses Nearmap’s aerial photographs to locate construction sites (and the hungry builders working on them) to better target sales. South Australia’s NRG Solar Services use Nearmap’s aerial photographs to prospect for leads, and measure roof sizes to calculate quotes quicker. The state’s Environment Protection Agency used the imagery to build a surveillance platform to uncover illegal dumping operations.
To date, scrutiny of the imagery has typically been done manually. But now Nearmap is applying machine learning to its vast troves of data, to automate the process. Earlier this month it launched a beta program for businesses to experiment with various models.
Absolutely killer dataset
The machine learning techniques means answering questions like ‘which suburb has the most pools?’ can be answered quickly.
“At the drop of a hat you can work out, which is the top postcode for solar in New South Wales? Or, where is the most construction happening?” Bewley says.
The answers are Newington, home to the Sydney Olympic Village where “pretty much every house has a solar panel” and a belt on Sydney’s fringes encompassing Box Hill, Marsden Park and Gregory Hill, Bewley says.
“The dataset is phenomenally difficult to build. That took years and team of really bright people. Those stats were literally a few evenings of ‘let's have a poke at that data that we've built and see what you can pull out of it’. Us going ‘I wonder what happens if we sort by solar panel percentage’,” Bewley explains.
“While the dataset it really hard to build. There's a lot of low hanging fruit that very easily pops out of it,” he adds.
While the aerial imagery could potentially be run through off-the-shelf image recognition offerings, Nearmap’s models are optimised for the dataset, and account for all the information attached to the images, such as location and time.
“This works a lot better, all my team does is optimise their results for Nearmap imagery. We don't care about any other sort of imagery. And it's more than just imagery, it's a geospatial representation of the physical world,” Bewley, who was previously lead data scientist at Commonwealth Bank of Australia, says.
The machine learning effort benefits from the sheer size of Nearmap’s datasets, and their high resolution. The imagery is counted in petabytes, and has a resolution of between five and seven centimetres ground sample distance (GSD). That means the centres of an image’s pixels, when measured on the ground, are about six centimetres apart.
“The accuracy relies fundamentally on the volume of data, and also the quality and the consistency and the nuance in it,” Bewley says.
“We have not only many petabytes of imagery, we have consistent high resolution, high quality data that's been captured over many geographies and many time points. And that is an absolutely killer dataset for doing machine learning with,” he adds.
Unleash the beast
Bewley has a team of 20 people dedicated to AI, working closely with the company’s 80-strong technology function to ensure their efforts can scale.
The firm, which listed on the ASX in 2012 and expanded into the US market two years later, is now inviting customers to take part in the beta program to experiment with various use cases.
Nearmap’s executive vice president, technology and engineering Tom Celinski said the company is now considering how to commercialise the AI capability.
“The step towards general availability is going to be; understanding the commercial landscape and then unleashing the beast, the beast being the AI system that can put out a whole raft of other attributes. We've got a collection of 30 plus things in the pipeline right now,” he said.
Potentially this could be in the form of ready-made models, such as a ‘pool finder’ or ‘construction site tracker’.
“Stay tuned, we’re innovating,” Celinski said.