Transforming weather forecasting through AI

By Treena Hein

AI is in the process of transforming most aspects of our lives, whether that is helping us directly write or research something, or helping professionals in health care, air travel or many other fields do a better job meeting our needs.

In agriculture, AI is already crunching large amounts of data (from individual farms or collective data from many farms) to provide a wide range of recommendations that improve crop and livestock production each year. But now, it’s poised to alleviate what’s probably the biggest stressor for farmers — weather. Or more precisely, weather forecasts.

“When you look at the variables that farmers face, there are some that they can control but others they can’t control, and top among those is the weather,” says Bruce Burnett, director of markets and weather information with Glacier FarmMedia. “If farmers have access to accurate longer-term forecasts, it will help with operational decisions, mitigate a lot of risk and reduce a large amount of stress.”

Even if it’s not perfect, AI generated forecasts would give farmers a lot more control with their planning throughout the growing season, says Burnett. “If you look at an operation like haying, you need dry weather over two to five days,” he notes. “If you know there is going to be dry weather, great. But even if you know there’s going to be two solid weeks of wet weather, that’s still going to be stressful, but at least you know what you’re facing.”

Artificial Intelligence: what’s happening?

It’s important to keep in mind that AI grows in power exponentially and that it learns as it grows. This is true in all areas where AI is being used, including weather forecasting.

For instance, the traditional method of predicting weather uses equations that represent the behaviour of various elements of atmosphere. While these equations have improved over time they are, by definition, simply equations. AI, on the other hand, intelligently compares its models to the actual weather that we experience and, with the aid of human experts, understands how to improve those models — very quickly.

So, to understand this better, imagine a current forecasting model where subsets of data are analyzed in a prescribed way and those results contribute to the accuracy of the model. Now imagine an AI forecasting model where each data subset can be analyzed in a thousand different ways in a matter of hours or even minutes, then each of those many results is fed back into the main model to see which ones make the forecast more accurate and which ones don’t. This repeats over and over with various subsets of data, allowing the AI to steadily chip away at inaccuracy — and even understanding why.

Another aspect of AI is that it can make real-time adjustments. In weather forecasting, this means that if a storm over the Pacific is forming, the AI will update the forecast as needed for, say, farmers in southern Alberta.

And the more data you give AI, the better. “Weather forecasting is very data intensive and there is already an enormous amount of data for AI to use its computational power on,” says Burnett. “And there will be more of it.”

Rainbow and Aardvark: two real-life examples

Poland-based Rainbow AI has a proprietary forecasting engine that integrates satellite imagery, radar data and ground-based sensor inputs, along with other data sources, to make weather predictions — mostly rain forecasts. “These data sets are refined through deep learning-based computer vision techniques, which enable sophisticated noise filtering, bias correction and resolution enhancement,” explains Dzmitry Danilchuk, head of partnerships with Rainbow AI. “Continuous validation against certified meteorological observations worldwide ensures a closed-loop feedback system, allowing real-time performance monitoring and rapid system optimization.”

Rainbow focuses on super-precise precipitation forecasts up to four hours ahead, with a six-plus hour demonstration model available for select clients. “We outperform major industry providers, as validated by open-sourced, transparent methodology benchmarking,” says Danilchuk, adding that people can see it for themselves online at WeatherIndex.ai. “We offer standard spatial resolution at one square kilometre, in line with industry benchmarks. But for select geographies, we can provide (resolution) down to 100 to 250 square metres.”

Another weather AI system, Aardvark Weather, developed by researchers at the University of Cambridge in the UK, forecasts the entire weather picture, not just precipitation, with up to 10 days’ lead time.

Aardvark is not only 10 times faster than other AI weather forecasting systems, it uses a thousand times less computing power to deliver its results. Developed with the support of the Alan Turing Institute, Microsoft Research and the European Centre for Medium Range Weather Forecasts (ECMWF), a paper about Aardvark’s groundbreaking work was recently published in the journal Nature (March 20, 2025 online edition).

“Aardvark reimagines current weather prediction methods offering the potential to make weather forecasts faster, cheaper, more flexible and more accurate than ever before,” said Professor Richard Turner in in a University of Cambridge announcement. Turner, part of Cambridge’s Engineering Department and the Alan Turing Institute, is the lead researcher for weather prediction.

Why Aardvark is so much faster and efficient comes down to its ability to simplify the heavy computational lifting that AI weather forecast systems are normally required to do. It can take many hours, require powerful super computers and large teams of experts to oversee the data set amalgamation and data processing needed to generate forecast results. In recent years, some AI companies have solved for part of this process — that is, they have found a way to replace one component of the weather forecasting pipeline (numerical processing) with AI, and that is helping to make forecasts faster and more accurate.

What Aardvark has done is completely replace that entire multi-stage forecasting process with a single, simple AI machine learning model that directly processes observational data from many sources, such as satellites, weather stations, planes and ships, and eliminates the need for numerical processing. The result is global and local forecasts generated in minutes and requiring the computational power of a mere desktop computer.

It is a massive streamlining of what was a multi-stage, multi-model, computer power intensive forecasting system into one, simple, efficient, flexible and smart system. Indeed, when using just 10 per cent of the input data of existing forecasting systems, Aardvark outperforms the U.S.’s national Global Forecast System (GFS).

Patience is required

So, how soon until Canadian farmers can benefit from these new systems? For now, University of Cambridge researchers can only say that ECMWF is in the process of rolling out AI forecasting methods into operational practice — and this is across many industries, not just agriculture.

For its part, Rainbow AI has plans to enter agriculture “through integrating into leading platforms such as Bayer’s Climate FieldView,” says Danilchuk. “We believe collaboration with existing ecosystem leaders will bring maximum value to farmers by embedding hyperlocal, high-precision precipitation insights directly into their operational workflows.”

When it is eventually available to farmers, Burnett believes AI weather prediction tech will have to show its value, like any other new tech, to be fully adopted. “That typically happens like it always does,” he says, “farmer by farmer.”