Press "Enter" to skip to content

Polar ice and land masses: How AI should improve climate forecasting

The World Health Organization warns of the serious consequences of climate change for individuals and public health. There would be increasing cases of malnutrition due to soil and drinking water pollution, as well as decreased crop viability. Air pollution causes breathing difficulties.

Elsewhere, climate change is causing the polar ice to melt and the habitat of animals to be endangered. But the calculations of when the Arctic should be ice-free are different. In order to concretize consequences like these, researchers are working on AI systems that are supposed to make forecasts more concrete – be it for solid land masses or for ice surfaces.

For example, the British Antarctic Survey (BAS) and other institutions have the KI IceNet developed. It is intended to forecast the decline in ice sheets in the Arctic. According to one Arctic Council report As of May 2021, the air and water in the Arctic have warmed three times as fast as the rest of the planet since 1971. Polar bears and marine mammals rely on the ice as a hunting ground and resting place. Indigenous peoples move on routes that are no longer reliable.

How can Germany become climate neutral? How can AI make climate models better? And: what is behind negative emissions? The current climate special from MIT Technology Review (now available in well-stocked newsagents) revolves around these and other questions.

IceNet uses a deep learning method that learns from computer simulations of ice changes since 1850 to derive six-month forecasts. Another team of scientists at the Johns Hopkins University Applied Physics Laboratory has developed a predictive model that uses satellite images to predict how quickly the ice will form each week.

On the other hand, an international research team led by Max Callaghan from the Berlin Mercator Research Institute on Global Commons and Climate Change has focused on the effects on solid land masses. With the help of artificial intelligence, he had classified and analyzed more than 100,000 published studies on climate change. The result: climate change is already having an impact on at least 80 percent of the global land area and thus affects areas where around 85 percent of the world’s population live. The studies examined came from different specialist areas, such as agricultural research, biology, geophysics, migration and conflict research.

There is a good reason for the laborious work: Knowledge about climate change is growing rapidly. Since the first status report of the Intergovernmental Panel on Climate Change (IPCC) was published in 1990, the number of specialist articles published annually has increased a hundredfold. This year even the actual report of the IPCC had almost 4,000 pages to show, based on 14,000 studies. This multitude of studies can hardly be evaluated without artificial intelligence.

Max Callaghan and his team first filtered 102,160 out of over 600,000 studies they found in the Web of Science and Scopus databases on the topic of climate change. The voice recognition model BERT (Bidirectional Encoder Representations from Transformers) developed by Google was used for their evaluation. BERT recognizes the meaning of individual words in a text and relates them to the others. BERT learned this through training with countless texts from Google Books and Wikipedia, for example. The AI ​​had to fill in the blanks. BERT can therefore predict missing words in texts and therefore also grasp the content of a text faster and more reliably than many other models.

“We had a metadata-provided dataset with around 3,000 studies for our task,” says Callaghan. “We then made use of the implicit knowledge of the Google model, so that we only had to train it in this application with a much smaller data set”. It’s like a cook preparing an unknown recipe – he doesn’t have to learn his craft again – he still has to practice the dish before serving it to the guests.

With such a broad spectrum of studies, however, it was not easy. For example, a study can look at the effects of a change in temperature on apple blossom in a particular area. “It’s a very specific question, but the pattern behind it repeats itself in different contexts,” says Callaghan. In the end, BERT initially succeeded in identifying whether a study was about climate impacts – the system then also identified the location, the type of impact and the driver behind it, such as changes in temperature or precipitation.

More from MIT Technology Review

More from MIT Technology Review

More from MIT Technology Review

More from MIT Technology Review

The researchers combined the results of the analysis with geophysical data from climate research. Using geoparsers, they decoded the information and transferred it to an interactive map. This consists of grid cells, the color of which indicates the number of studies in the corresponding region. The study and its graphic evaluation is intended to provide decision-makers from politics, society and business with a realistic overview of the overall situation – so that they can make better decisions

Callaghan’s researchers now want to equip their interactive maps with direct links to the studies. “We don’t necessarily differentiate between good and bad effects in the studies we identify,” he says. So not every study is about catastrophic floods or fatal heat waves. “Some of them are about shifts in the migration patterns of a particular butterfly or something similar,” he says. “Ultimately, though, all of this points to the huge impact climate change is having around the world. And I think the scale of that is pretty daunting.”


(jle)

Article Source

Disclaimer: This article is generated from the feed and not edited by our team.