Mapping Biome Boundaries From Climate Patterns: A Machine Learning Framework

Abstract

Aim

To develop a climate-based, probabilistic framework that maps global biomes as climatic domains and explicitly quantifies the spatial distribution and intensity of biome transition zones.

Location

Global terrestrial biomes.

Methods

We trained a probabilistic machine-learning classifier linking biomes to multivariate climate space. We quantified biome ambiguity for each grid cell using Shannon entropy, identifying ecotones as regions where climate supports multiple biome outcomes, that is, high entropy. We applied the framework to present-day climate and mid-century climate projections to evaluate potential biome reorganization and ecotone shifts under climate change.

Results

The model reproduced the broad geographic distribution of major biomes, although spatially blocked validation revealed variation in geographic transferability among biome classes. High-entropy zones form coherent transition belts that coincide with well-known ecotones, including forest–savanna boundaries, desert margins, temperate forest–grassland transitions, and the boreal–tundra interface. Biome core regions show low entropy, whereas ecotones exhibit diffuse probability structure and heightened sensitivity to climatic variation. Projections of the model under mid-century climatic conditions (2041–2060) identify spatially structured patterns of biome transitions, particularly shifts from tundra to boreal forest and from boreal to temperate systems, while also highlighting regions where uncertainty and model disagreement are greatest.

Main Conclusions

Mapping biomes in climate-probability space reframes biome boundaries as spatially extended, permeable zones shaped by overlapping climatic regimes. By combining probabilistic classification with entropy-based uncertainty quantification, our framework complements traditional biome maps by revealing the gradient structure and stability of biome transitions and by identifying where climatic change is most likely to drive ecological reorganization. The resulting transition and uncertainty surfaces offer globally comparable spatial layers for macroecology, biogeography, and evolutionary research, including studies linking ecotones to biodiversity patterns, trait turnover, and lineage diversification.