Machine learning helps identify climatic thresholds that shape the distribution of natural vegetation

Changing climate brings more frequent and more intense climatic extreme events. It is unclear, however, exactly how climate extremes will affect vegetation distribution in the future. This is an acute question for research in order to be able to mitigate coming extremities and their impact on vegetation.

A study published in Global Change Biology explores large-scale relationships between vegetation and climatic characteristics using machine learning. It demonstrates that combining climate and remotely-sensed land cover data with tree-structured predictive models called decision trees can effectively extract the climatic thresholds involved in structuring the distribution of dominant vegetation at various spatial scales.

The findings of this study highlight the importance of climatic extremes in shaping the distribution of several major vegetation types. For example, drought or extreme cold are essential for the dominance of savanna and deciduous needleleaf forest.

“One of the most important questions left to answer in the further research is whether the climate thresholds recognized in this study are static or changing with the climate changes in the future,” says researcher Hui Tang from the department of Geosciences of the University of Oslo.

Collaboration between machine learning and vegetation experts

Predicting future vegetation distribution in response to climate change is a challenging task which requires a detailed understanding of how vegetation distribution on a large scale is linked to climate. The research team consisting of computer scientists, vegetation modelers and vegetation specialists examine the rules coming from the decision tree models to see if they are informative and if they can provide any additional insights that could be incorporated into mechanistic vegetation models.

“It is a difficult task to validate whether a data-based model is informative and robust. This study highlights the importance of interpretable models that allow such meaningful collaboration with the domain experts,” says doctoral researcher Rita Beigaitė from the department of computer science of University of Helsinki.

“The major climatic constraints recognized in the study will be valuable for improving process-based vegetation models and its coupling with the Earth System Models,” says Hui Tang.