For many years now, experts have studied how changes to climatic patterns are leading to major shifts in vegetation around the world. That’s because as various regions experience unprecedented changes to temperature and precipitation, plants will gradually migrate to places where they were not able to grow before, or even disappear altogether, resulting in serious impacts on both wildlife and humans alike, in addition to setting the stage for systemic feedback loops that will accelerate and worsen the climate crisis further.

Typically, scientists now collect data on these vegetation patterns which is analyzed with what is known as a dynamic global vegetation model (DGVM), a kind of computer-simulated model that helps them predict what kind of changes might lie ahead.

But the interactions between plants and climate — especially on such large scales — can be quite complex, and existing techniques like dynamic global vegetation models have limitations that might prevent them from accurately predicting future outcomes.

To address this gap, a team of researchers from the University of Oslo, University of Helsinki, and the Norwegian Institute for Nature Research propose using machine learning as a way to improve these computer simulations, with a particular focus on investigating how climate extremes impact future predictions.

“Variation in climate is the major factor determining the distribution of vegetation around the world,” explained the researchers in their paper, which was recently published in the journal Global Change Biology. “As the world is facing climate change, large-scale future dynamics in vegetation distribution are expected, which in turn may exert strong biophysical and biochemical feedback on the climate. Predicting future vegetation distribution in response to climate change, however, is particularly challenging, requiring a detailed understanding of how vegetation distribution on a large scale is linked to climate.”

As the team also points out, the use of machine learning in the sciences is nothing new. Machine learning has become increasingly popular in the biogeosciences, the interdisciplinary field that studies the interactions between biological and geological processes, and such models that are built upon observational data present a simultaneously granular and big-picture approach.

“Models built upon observational data offer the potential to combine a higher resolution while keeping investigations at the largest possible scales,” said the team. “In this study, we employed a decision tree approach from machine learning to explore available climate and vegetation data, and to systematically re-examine long-lasting and reappearing scientific questions regarding climate — vegetation relations. This approach enabled us to analyze whether any novel climate thresholds affecting the large-scale distribution of vegetation types could be detected, particularly climate extreme thresholds that have been overlooked in previous studies.”

The team chose to focus on climate extremes because it is these types of conditions that are now becoming more frequent and severe, due to climate change. These extremes stray statistically from average climate records, and are now occurring more often, which can have huge impacts on how the dynamics of vegetation patterns ultimately play out. For instance, extreme lack of precipitation or extreme cold are crucial to the proliferation of savanna and deciduous needle-leaf forest.

“The predictive performance of species distribution models increases when mean climatic predictors are complemented by climate extremes,” noted the team. “A changing climate influences the duration, frequency, intensity, timing and spatial extent of climate extremes. For instance, daily temperature and precipitation extremes, in particular, have been observed to increase in frequency and intensity due to global warming with distinct spatial pattern from average climate changes.”

The team’s method involved the use of decision tree models, also known as classification trees or regression trees, which is a form of machine learning that features a hierarchically structured series of queries or tests. The decision that is made on one level of query or test will influence which test is taken next, and these progressive series of tests will affect the ultimate outcome.

Because the mechanisms and reasoning behind AI-powered predictions is often unclear, the researchers chose to use decision tree models because they are easily interpretable, meaning that it is easy to determine why a certain classification was made. The team then trained their decision tree models with publicly available, present-day global climate and vegetation data, and tested their ability to predict what type of vegetation would be dominant in a region, given its climatic variables.

Intriguingly, the team’s work found that their AI-assisted approach made significantly more accurate predictions about future vegetation distribution than baseline models. In addition, the team emphasized the limitations of “hard-coded climate thresholds” that baseline models typically rely upon.

“To the best of our knowledge, no attempts have yet been made to use machine learning for understanding threshold conditions that govern and separate dominant vegetation types at a global scale,” said the team. “Our decision tree results, however, emphasize the importance of using climate extremes, especially extremes on a daily scale, in defining the climate thresholds of different vegetation types.”

Ultimately, the team’s findings indicate that it will be critical for next-generation DGVMs to incorporate varying climate thresholds, gleaned from mean climate conditions, rather than static thresholds for the whole globe, as is typically done now. Such a shift will help experts improve their tools, and in doing so, humanity will be better informed — and therefore also better equipped — to face the dire challenges of climate change.

Read more over in the team’s paper.