"People often ask me — and it's a kind of urban legend — that weather forecasts don't hold up. I cannot confirm that," Ala said. He noted that meteorologists set the bar high. "If the accuracy for the simplest parameters — temperature, precipitation and wind — drops below 90 percent, I would say we have a crisis."

Modern numerical weather prediction, based on physical laws, divides the atmosphere above us into computational layers. A single standard situation often involves more than 50 model outputs, forming a fan of possible scenarios. Behind every grid point is a massive calculation.

In special situations, meteorologists run far more detailed computations. "For example, in France during the Olympics, they produced weather maps with a resolution of a few hundred meters. Essentially, they calculated conditions every half‑kilometer or even every 300 meters," Ala said. Meteorologists do not go that precise on a daily basis. "It's possible, but it becomes extremely time‑consuming and, since everything ultimately has a cost, very expensive."

Despite improvements in machines and models, human forecasters still play a vital role because they can combine model output with past experience. "To be fair: if we take today's model forecasts and add human expertise, we see roughly a 10 to sometimes even 18 percent improvement simply because a human is still involved," Ala said.

Increasingly frequent extreme weather events pose new challenges for both humans and computers. As the climate changes, weather becomes more volatile, and regional‑scale events push current models to their limits.

"In general, it means we already see — and will see even more — extraordinary weather events and phenomena. One characteristic is that they form quickly. They may be short‑lived, like showers, but very intense and very local. Hitting them accurately is quite challenging," Ala said.

Artificial intelligence helps forecasters by processing massive datasets thousands of times faster. Neural networks often perform excellently in short‑term nowcasting. But their ability to predict extremes remains weak because algorithms rely only on previously observed patterns.

As a result, AI often misjudges rare events, usually underestimating the risks. "It doesn't capture extreme weather well. In Estonia this may not be as obvious, but a large part of the world is seriously worried and suffering from extreme weather," Ala said.

He noted that neural networks arrived late to European meteorology. European institutes relied exclusively on classical supercomputers. At a pan‑European summit a few years ago, leaders received a shock: Asian researchers reported successful tests where their AI outperformed traditional forecasting. The unexpected news pushed European weather centers to rapidly develop new solutions. In a short time, Western researchers caught up using graph‑based neural networks.

Despite rapid progress, machine learning cannot replace classical atmospheric physics — at least not yet. Algorithms train on numerical model outputs and past observations. Without the old‑school model, AI loses its anchor. Ala illustrated the dependency: "What would happen if we switched off classical weather prediction and pulled the plug on the machine? Based on what we know today, machine learning would start making mistakes fairly quickly and eventually tell us nonsense."

To enrich regional observations, weather centers increasingly look to home weather stations. The national network of about 100 stations cannot capture every patch of forest or village microclimate. "If home stations are placed correctly and are representative enough, incorporating their data into model forecasts is entirely possible and appropriate. I think it's definitely a topic for the future," Ala said.

He added that the amount of weather‑related information grows explosively every year, changing the nature of specialists' work. The forecaster's role will increasingly be to connect and filter information produced by algorithms. "In the future, it won't be realistic for a person to look at a table and reach a highly accurate conclusion. The human role will be interpretation, but the data volume is simply massive," Ala said.

--

Follow ERR News on Facebook and X and never miss an update!