Drivers of tree growth across Europe: An explainable AI analysis of tree- and site-level influences

Abstract

Climate change is increasingly impairing forest ecosystems in Europe, reducing tree vitality and increasing mortality. Crown defoliation and stem growth are widely used indicators of early stress responses, yet their drivers remain difficult to disentangle due to complex, non-linear interactions among climatic, edaphic, and biotic factors.

Here, we apply explainable AI (XAI) to model annual diameter growth of individual trees from four dominant European tree species (Norway spruce, Scots pine, Common beech, and Oak) using long-term data from the ICP Forests Level II network. Gradient-boosted decision trees trained on tree-level attributes (e.g., defoliation, social class) and plot-level variables (e.g., soil solution chemistry, atmospheric deposition, and topography) outperform linear baselines across all ablations and grouping strategies. Ablation experiments show that plot-level features account for most of the predictive power while defoliation contributes only marginally.

XAI analyses reveal strong non-linear, species-specific response regimes, including marked growth reductions at high defoliation levels (∼35–60%) and optimal regimes of nitrogen and sulphate deposition, illustrating the capacity of XAI to identify candidate growth-relevant regimes and interactions.

However, temporally explicit validation using tree-wise cross validation results in a 54% reduction in R² score whereas spatially explicit validation based on plot-wise cross-validation leads to a near-complete collapse in predictive performance, indicating strong reliance on both temporal and spatial autocorrelation and context-specific patterns. Moreover, ablating defoliation features causes attribution to shift toward correlated environmental variables, highlighting the role of proxying and statistical confounding.

Overall, our results illustrate both the potential and the limitations of XAI for forest ecology: while effective for screening large observational datasets and generating hypotheses, XAI outputs require cautious interpretation, mechanistic understanding and spatially robust validation.