Artificial intelligence finds previously undetected historical climate extremes

There are over 30,000 weather stations in the world, measuring temperature, precipitation and other indicators often on a daily basis. That's a massive amount of data for climate researchers to compile and analyze to produce the monthly and annual global and regional temperatures (especially) that make the news.

Now researchers have unleashed artificial intelligence (AI) on these datasets to analyze temperature extremes in Europe, finding excellent agreement compared to existing results that used traditional methods, and as well have uncovered climate extremes not previously known. Their work has been published in Nature Communications.

With the world's climate changing rapidly, it is important to know how temperature and precipitation extremes are changing, so planners can adapt to the extremes here now and to what's coming.

It is raining heavier in some regions, now "far outside the historical climate" according to a 2021 paper in Nature. Heat extremes are up as well—more than 30% of the global land area now sees monthly temperatures above the two-sigma statistical level in any given year, up from about 1% in 1950.

A significant problem in the analysis of historical temperature averages is the lack of data for some weather stations, especially in the first half of last century.

A manned weather station may have gone unmonitored for years if it is damaged, if its keeper moved or died, if it stopped and was not immediately replaced, or maybe never replaced. New station technologies need to be correlated to previous instruments, and large areas in Africa and the poles offer scant information, if any.

Climate researchers have spent a great amount of time trying to deal with such gaps. A research area known as data homogenization, and different choices of homogenization methodologies largely account for the slight differences seen in the results of the several different groups that publish global temperature averages and trends.

A team led by Étienne Plésiat of the German Climate Computing Center in Hamburg, including colleagues from the UK and Spain, saw extreme temperatures as an area ripe for the application of AI's neural network techniques.

They focused on Europe, which has an especially dense number of weather stations that go further back in time than elsewhere around the world. (For example, the monthly Hadley Central England Temperature data begins in 1659, the oldest record in the world.) Using AI, the group reconstructed observations of European climate extremes—extremely warm and cold days, and extremely warm and cold nights.

Because of the high density of European temperature stations, traditional statistical methods such as Kriging, Inverse Distance Weighting and Angular Distance Weighting perform well in predicting temperature values for any location that lacks a thermometer but has neighboring stations nearby, but they perform poorly when nearby data is scarce.

All are methods to use measured values together with the distance from the point of interest to a neighboring weather station to predict the temperature at the location of interest, the primary difference being how the distances (or angles) are weighted in the calculation.

In the last few years, AI methods have outperformed these traditional ways of infilling to construct missing climate information and quantifying uncertainties.

The AI models used by Plésiat and colleagues were trained on and compared to historical simulations with Earth System Models from the CMIP6 archive (Coupled Model Intercomparison Project, a global collaboration of climate models coupling the atmosphere and oceans that calculate past climate, current climate and future climate).

Their AI's results are evaluated by comparison to such reanalysis simulations, using accepted methods such as root mean square error, the Spearman's rank-order correlation coefficient which indicates the amount of association between an independent variable and a dependent variable (it generalizations the well-known Pearson coefficient R but including nonlinear dependencies), and more.

The researchers found that their deep-learning technique, which they call CRAI (Climate Reconstruction AI), outperformed several interpolation methods such as those described above for calculating warm days (the percentage of days when the daily maximum temperature was greater than the 90th percentile), cool days (the percentage of days when the daily maximum temperature was less than the 10th percentile), and similarly for warm nights and cool nights.

They then applied it to the reconstruction of all fields in the HadEX3 dataset over the European domain—HadEX3 consists of over 80 indices of extreme temperature and precipitation on a gridded Earth surface from 1901 to 2018.

Here, too, their technique showed an ability to reconstruct past extreme events and reveal spatial trends across time intervals not covered by so-called "reanalysis datasets." (Climate reanalysis fills in gaps in observational databases by utilizing a climate model together with what observations are available.)

In addition, their CRAI revealed European extremes previously unknown—for example, cold spells such as one in 1929, and heat waves including a 1911 occurrence. Due to sparse data, such extremes were only hinted at anecdotally.

"Our research demonstrates both the necessity and the potential benefits of applying this approach to the global scale or other regions with scarce data," the team conclude in their paper.

"Indeed, we find that our AI-based reconstruction shows larger accuracy over traditional statistical methods, particularly in regions with pronounced data scarcity," adding that training such CRAI models should enhance accuracy when larger amounts of information are exploited.

"This work underscores the transformative potential of AI to improve our understanding of climate extremes and their long-term changes."