Flooding in Southland. Photo: Guy Dowding/High Country Helicopters
The National Institute of Water and Atmospheric Research has developed an AI model which can forecast flooding within minutes.
NIWA manager Nava Fedaeff said staff currently used another form of physical modelling to predict flooding, but this take 24 hours to develop forecasts.
"By the time that time is done, those areas are probably already flooded. So, it's too long," Fedaeff said.
The new machine learning technology can be leveraged to save time.
She said they have been comparing the AI results to existing physical models and it was very accurate.
"Inundation forecast accuracy, compared to that other model, is 95 percent. It is very good, is what we are saying."
When most people think of AI, they often think of the large language models used for example by ChatGPT.
But Fedaeff said machine learning AI was quite different
"It is around pattern recognition - you just say [to the model] 'look at all these pictures, this is the flooding, this is the river flow and sea level that caused the flooding. Can you learn that relationship?' And it does it really well."
Once the AI model has been trained, NIWA can feed in other sets of river flows and sea level heights and the AI model can predict what flooding might happen.
Fedaeff said AI technology was a very rapidly developing technology.
"NIWA's trying to stay at the front edge of that science and we're always looking at ways we can leverage the science and technology to make things either faster or more accurate."
At present, the AI model was trained for Westport only. This was because NIWA found they had a depth of data from Westport to use to train the AI and compare results to reality.
"Obviously we'd love to be able to do this for the whole country," Fedaeff said.
NIWA is now focusing on whether the model needs to be retrained for each area of the country or if the model can use what it has learnt from Westport to develop forecasts for other areas.
Sign up for Ngā Pitopito Kōrero, a daily newsletter curated by our editors and delivered straight to your inbox every weekday.
