Until now, it has proved very difficult to make such predictions with current systems based on weather or climate models. With this new machine learning-based prediction technique, Vijverberg's team has now made decisive progress. They analysed soybean harvests in the United States and showed that you can predict crop failure even before sowing. Sem Vijverberg explains: “We tested how well machine learning-based prediction would have performed over the past 25 years and showed that, in the more predictable years, our prediction was correct 75% of the time for February. That’s a big step forward compared to current operational soybean forecasts that are provided only in early August”. The results are published in Artificial Intelligence for the Earth System.
Cubes
Traditional weather forecasts are made using numeric models which split the climate system in cubes. Each cube has its own temperature, humidity, wind velocity, etcetera. The model can calculate all the changes in the variables in function of time. But there are physical interactions which occur on much smaller, spatial and temporal scales than the size of the cubes. With a statistical definition (parametrisation) the impact of these small-scaled processes on the larger cubes are approximated, however these approximations sadly are inaccurate.
Interaction ocean and atmosphere
Parametrisation plays an important role in stimulating the interaction between the ocean and atmosphere, and this interaction is meaningful to the predictability on longer time scales. It seems that inaccurate parametrisation leads to limited predictability of numeric models. To bypass this Vijverberg used data-driven techniques in his research. Vijverberg: “We learned how the interaction between the ocean and atmosphere changes per season, and which factors are needed for a strong interaction between the two. Partially because of these insights we are able to predict the temperature, or for example the failed crops in the United States better than previously.”
Spin-off
Not only growers benefit long-term predictions; also consider adequately filling gas reserves: will there be a mild or extreme winter? This has an impact on the amount of gas that needs to be reserved. Vijverberg: “Through the chaos in our climate system we will never be able to exactly predict the weather forecast at two o’clock in the afternoon in three months. However we could probably predict what will likely happen.”
The developed method is generically applicable to other continents, weather variables or weather dependent variables such as crop yield. Although a sidenote is that for each region and variable a clear limit in predictability exists because of the inherent chaos in our climate system. Profound knowledge on the climate system helps to make proper choices in the development of predictions using data-driven techniques. Vijverberg: “To successfully integrate this innovation in society the engagement of experts is required. That is why we are set out with a spin-off named Beyond Weather, by which we will be developing operational weather forecasts for stakeholders in the following industries: power, agriculture, and humanitarian aid. Currently we have already developed a product which substantially predicts the temperature in Europe one to three months ahead.”