Identifying climate thresholds for dominant natural vegetation types at the global scale using machine learning:...

Area of Science:

• Ecology
• Climate Science
• Machine Learning

Background:

• Global vegetation distribution is driven by climate and influences climate.
• Predicting vegetation changes under climate change requires understanding vegetation-climate coupling.
• Dynamic Global Vegetation Models (DGVMs) use climate thresholds, but optimal thresholds are challenging.

Purpose of the Study:

• To use machine learning to identify large-scale relationships between vegetation distribution and climate characteristics.
• To compare the importance of climate extremes versus averaged climate variables in determining vegetation distribution.
• To improve the representation of climate impacts on vegetation in DGVMs.

Main Methods:

• Applied machine learning, specifically decision trees, to analyze remotely sensed vegetation and climate data.
• Extracted relationships between dominant vegetation types and climate extremes (e.g., extreme cold days).
• Compared model performance using climate extremes versus averaged climate variables for predicting vegetation distribution.

Main Results:

• Climate extremes, particularly extreme cold days, more accurately describe vegetation distribution and eco-climatological space than averaged climate variables.
• Vegetation types in cold environments show temperature thresholds that vary with moisture.
• Future vegetation change predictions incorporating climate extremes align better with DGVMs and are less pronounced.

Conclusions:

• Climate extremes are crucial for accurately modeling vegetation distribution and predicting future changes.
• DGVMs need more explicit representation of climate extreme impacts on vegetation.
• Machine learning offers a powerful tool for uncovering complex vegetation-climate relationships.