Met Office Fellow in Operational Meteorology Steve Ramsdale assesses the relationship between human expertise and AI capability for the future of forecasting.
Weather forecasting has always been about people. Over the decades we have moved from hand-drawn weather charts to global numerical weather prediction systems running on some of the world's most powerful computers.
Throughout this transition however the decisions about what the weather will do and what the weather will mean have always remained with those who dedicate their lives to understanding the complex weather machine.
AI technologies have emerged as a new and potentially game changing approach to weather prediction. This has expanded the range of different sources of weather information, many with impressive claims about forecasting skill challenging the paradigm of weather as the preserve of national meteorological services.
Despite these advances and the hyperbole surrounding AI there is a key one message that can be missed in all the excitement of these new developments.
The real value comes from how that forecast is interpreted, communicated and acted upon, or more simply, a forecast has no impact until a human uses it.
From data to decisions
Modern forecasting systems already generate vast volumes of information with hundreds of possible future scenarios from state-of-the-art ensemble systems. At the same time more and more novel observations are becoming available. This is where machine learning methods may be most valuable, in extracting useful data from these vast and ever-increasing sources.
This leads to one of the main challenges for users, not a lack of information, but an abundance of it.
Forecasters, emergency planners, transport operators and many other decision makers are routinely faced with questions such as:
- How confident should I be in this forecast?
- Which scenario is most plausible?
- What are the consequences if the forecast is wrong?
These questions are not purely scientific. They involve judgement, context and experience. A weather model can calculate multiple evolutions of the atmosphere. It cannot tell you why these may happen. That ability remains human and is a core tool when making decisions about severe weather that could lead to large scale impacts upon people's lives.
Forecasts are about people, not computers
One of the most important lessons from operational meteorology is that users rarely want raw forecast data.
A transport operator wants to know whether high winds could affect a particular route.
An emergency responder wants to understand the most likely impacts and the associated uncertainties.
A member of the public wants to know whether they should change their plans.
In each case, the forecast must be translated from atmospheric information into decision-relevant information.
This translation step is where human expertise becomes invaluable.
Experienced forecasters routinely combine model output with an understanding of local effects, historical events, user requirements and known model strengths and weaknesses. They act not simply as producers of forecasts, but as interpreters and communicators of uncertainty.
The role of trust
Trust remains one of the most important and least discussed aspects of forecasting.
Users make decisions because they trust the information they receive. That trust is built not only through forecast accuracy, but also through transparency and consistency.
When a forecaster explains why confidence is low, users can account for that uncertainty in their planning.
When a forecast changes, users want to understand the reason.
As forecast systems become increasingly complex, ensuring that users maintain trust in the information becomes even more important.
This is one reason why explainability and interpretation are becoming central topics in weather science. Whether forecasts come from physics-based models or newer machine learning approaches, users need confidence that forecasts are physically plausible and operationally meaningful.
New tools, familiar challenges
Recent advances in AI have generated understandable excitement across the meteorological community.
Machine learning systems can often generate forecasts more rapidly than existing approaches and are beginning to demonstrate impressive skill across a range of applications.
However, many of the questions faced by forecasters remain unchanged, such as:
- How should uncertainty be communicated?
- How should conflicting pieces of guidance be reconciled?
- How do we identify unusual or unprecedented events?
- How do we recognise when a forecast may be wrong?
These challenges do not disappear when a new forecasting technology is introduced. In many ways, they become more important.
The role of the forecaster may evolve from generating forecasts to interrogating them, challenging them and explaining them.
Human expertise as a safety system
Operational forecasting has always relied on multiple layers of assurance.
Forecasters routinely identify situations where automated output struggles, whether because of specific atmospheric conditions, rare events or local effects that are difficult to represent.
READ MORE: How do we harness the benefits of AI without losing what matters most?
Human expertise acts as an additional safety system.
The greatest future opportunities may therefore come not from replacing forecasters, but from creating better partnerships between people and intelligent systems.
Computers excel at processing huge volumes of data and recognising patterns.
Humans excel at contextual reasoning, communication and decision support.
The combination of both is often stronger than either alone.
Looking ahead
The future of weather forecasting is unlikely to be a competition between humans and machines.
Instead, it will be a collaboration.
Forecast systems will continue to improve. Machine learning will become increasingly important. New datasets and new methods will emerge.
But the need for human judgement will remain.
The most valuable forecasters of the future may be those who can bridge the gap between increasingly sophisticated prediction systems and the people whose decisions depend upon them.
Because ultimately, weather forecasts are not produced for models.
They are produced for people.
And helping people make better decisions remains the most important part of forecasting.
This blog is written by Steven Ramsdale. The points discussed here can be seen in more detail in this open access article: https://rmets.onlinelibrary.wiley.com/doi/10.1002/wea.70151
Find out more about AI research at the Met Office.
https://www.youtube.com/watch/-YHM-Sy4ihM?start=0