Could AI give us rapid-fire forecasts?

Could we forecast precisely enough to be warned of local severe weather within an hour? Photo / File

Could we forecast accurately enough to be warned of local severe weather within an hour?

That's the goal of a new $500,000 study, drawing on artificial intelligence to predict where and when extreme convective weather events will happen, in a matter of hours or even minutes.

Often bringing heavy rain, lightning and strong winds, convective weather events are usually localised in time and space, developing rapidly and without warning.

Conditions are always associated with cloud masses, such as large thunderstorm cloud, which can be spotted from earth observing satellites.

But the main way we detect and forecast them remains to be looking at a combination of satellite imagery, models and rain radars.

While trained forecasters are skilled at this, it's a resource-intensive job that can't be fully automated - and we can't look everywhere all of the time.

That's where machines come in.

Recent advances in machine learning have proven exceptional at identifying and labeling features in images, prompting researchers to ask whether artificial intelligence could be steered toward weather-watching.

"The basic idea is to train a machine learning algorithm to analyse satellite imagery, possibly combined with other inputs such as numerical model guidance or rain radar, and predict where and when heavy rain, lightning or wind squalls might occur," MetOcean Solutions technical director Dr David Johnson said.

"As this is a machine process it can potentially be fully automated and then used to send alerts on a phone app."

In the new study, supported through the Ministry of Business, Innovation and Employment's Endeavour Fund, MetOcean Solutions experts will collaborate with machine-learning researchers from Auckland University of Technology's Knowledge Engineering and Discovery Research Institute (KEDRI) and MetService.

Johnson felt there was a "reasonable chance" the algorithm would be better than a human at consistently making the correct predictions.

Further, an automated machine approach allowed for better customisation.

"For example I might request an alert if heavy rain is likely - so I can take in my washing - my phone already knows where I am," he said.

"If I was coaching a kids soccer game, I might want to be able to get everyone under shelter if lightning or hail was likely.

"Human forecasters could never manage to serve all the myriad of end user needs at different locations and times."

Machine learning had already been applied to a similar problem of precipitation nowcasting from radar, and has shown to be better than current state-of-the-art algorithms.

The technology was also being increasingly used to integrate information from multiple numerical models and local observations to provide optimal "consensus" forecasts.

"I am convinced that machine learning will play a key role in the future of weather forecasting," Johnson said.

"Humans simply cannot process all of the data coming from earth observing platforms and traditional computer forecast models.

"As with many other domains, machine learning or 'artificial intelligence' is all about automating certain aspects of human workflows which a computer can do better, faster or more efficiently."