Emergent vulnerability to climate-driven disturbances in European forests

European forests cover over 2 million km² and provide key ecosystem services, yet their long-lived tree species make them vulnerable to abrupt disturbances. Anticipated climate change is expected to increase risks from natural disturbances (fires, windthrows, insect outbreaks). Land surface models attempt to represent these disturbances but remain limited by incomplete ecological understanding. Prior large-scale assessments rely on coarse reports or satellite disturbance maps without clear attribution of disturbance agents and often consider limited drivers and assumed functional forms. This study aims to quantify forest vulnerability—defined as relative biomass loss conditional on disturbance occurrence—across Europe (including Turkey and European Russia), identify emergent driver–response relationships and interactions using machine learning, and assess spatial patterns and temporal trends (1979–2018). The work emphasizes that vulnerability is not equivalent to risk (which also includes hazard probability and exposure).

The paper reviews that LSMs increasingly include mechanistic disturbance modules but only partially capture dynamics due to limited ecological knowledge. Compilations of mortality events provide broad coverage but coarse spatial detail, masking environmental controls. Satellite datasets enable high-resolution mapping of disturbances, but attributing agents from remote sensing remains challenging. Previous studies explored climate controls on mortality without attributing to specific agents and typically used limited drivers with regional aggregation and a priori relationship forms, limiting detection of local interactions and compound effects. Emerging machine learning approaches with Earth observations can capture complex, interacting drivers without prescribed functional relationships, but have not been broadly applied to multiple disturbance types at continental scales.

Study domain and disturbances: Europe including Turkey and European Russia. Vulnerability is defined as relative biomass loss (BL) conditional on a disturbance. Disturbance records (2000–2017) were compiled from: fires (EFFIS, 15,818 polygons), windthrows (FORWIND, 89,743 polygons), and insect outbreaks (USDA IDS, 50,777 polygons; used for model development due to lack of European dataset). Records were associated with host plant functional types (PFTs: BrDe, BrEv, NeDe, NeEv). For insects, functional pest groups were merged due to data limitations.

Biomass time series reconstruction (2000–2017): A static 2010 above-ground biomass (AGB) map at 100 m was integrated with Landsat Global Forest Change tree cover loss/gain layers. Biomass density per percent of tree cover lost/gained (P_loss, P_gain) was estimated using 2010 AGB and derived TC_2010 for areas experiencing loss/gain; gain TC was inferred using a 2.5 km moving-window maximum and linear growth assumption to 2010. Spatialized P_loss and P_gain fields (0.1° median) and annual tree cover change produced annual biomass maps B_t at 100 m via a linear adjustment formula.

Computing event-level biomass loss: For each disturbance polygon at year t, BL_t = [max(B_{t−n…t−1}) − min(B_{t…t+m})] / max(B_{t−n…t+m}). Lags: n=m=1 for fires and windthrows; n=2, m=5 for insect outbreaks to capture gradual effects. BL_t was averaged over each polygon and used as the response variable.

Predictors: Three categories—Forest (AGB, LAI, tree density, tree age, tree height), Climate (FWI, MI, precipitation cumulated, snow, short-term anomalies of precipitation and temperature, maximum temperature, long-term average temperature, maximum wind speed, SPEI anomalies), Landscape (slope, elevation, homogeneity, coefficient of variation). Predictors were annual (dynamic) or static climatologies. PFT fractions were retrieved from ESA-CCI.

Modeling framework: Random Forest (RF) regression (500 trees; hyperparameters tuned by Bayesian optimization) was used to predict BL_t from environmental predictors for each disturbance type. Records with BL_t ≤ 5% were excluded. Some wind extremes (e.g., tornadoes; Klaus storm) were excluded due to underestimation of wind speeds in coarse reanalysis. Data splits: for each year, 60% of records formed training and 40% validation sets. To reduce spatial dependence and sampling bias, BL and predictors were binned along the first three principal components (20 equal-interval bins per PC); area-weighted means per bin were used. Only bins with ≥3 records were retained. Missing predictor values were median-imputed.

Feature selection: Initial RF with all literature-based predictors produced importance scores. Highly correlated pairs (Spearman > 0.8): the lower-importance predictor was dropped. Iterative RF runs removed the least important predictor, selecting the set maximizing R². Partial Dependence Plots (PDPs) were inspected for interpretability and boundary consistency. PFT-specific RF models were trained using bins where that PFT fraction >5%. Performance was assessed with R², RMSE, PBIAS, and relative error.

Interactions: Friedman’s H-statistic quantified second-order interaction strength among features. Amplification/dampening (AP) was computed as the relative difference between two-way and one-dimensional partial dependences at peak values, averaged by predictor categories.

Spatial-temporal projections: RF models were run in predictive mode on 0.25° grids for Europe annually (1979–2018) to estimate potential relative biomass loss (PBL). Climate predictors were dynamic over time; forest and landscape predictors were held at their 2009–2018 averages due to lack of time series. PFT-specific outputs were area-weighted by local PFT fractions. Uncertainty per grid cell was quantified as RF standard error across trees. Current vulnerability was defined as the 2009–2018 average. Local sensitivity was computed using Individual Conditional Expectation (ICE) slopes per predictor; category-wise marginal contributions Z_marg were derived from summed ICE slopes by category.

Trend attribution: Grid-cell trends (Mann–Kendall) in PBL (1979–2018) were computed. Factorial simulations isolated the effect of each climate predictor by varying it over time while fixing others to current values, attributing dominant drivers of trends.

Overall Vulnerability Index (OVI): Multi-disturbance vulnerability used inclusion–exclusion to combine disturbance-specific PBLs per year under independence and homogeneous spread assumptions. Space-time integrated OVI combined min–max normalized current OVI and its trend via multiplication to identify hotspots. Interactions among different disturbances and depletion of exposed biomass were not modeled.

Data/code: Datasets are public; models and MATLAB code available via figshare DOIs; additional preprocessing code (R/Python/GEE) available on request.

Model performance: RF models explained 34–49% of variance in relative biomass loss (R²) with RMSE 9–11% (12–15% of observed range). Biases: ~2% overestimation for windthrows; ~2% underestimation for fires; ~10% underestimation for insect outbreaks (PBIAS). PFT-specific R² for insects ranged 0.28–0.53.

Drivers and response functions:

• Fires: Vulnerability increases with higher biomass, tree density, and age (fuel loads). Climatic water stress (low precipitation, high Tmax, low MI, high FWI) increases vulnerability; less homogeneous landscapes and steeper slopes reduce it.
• Windthrows: Higher biomass, age, and tree height increase vulnerability (reduced flexibility, increased bending moment). Higher wind speed, more precipitation and snow (saturated soils and canopy loading), and colder long-term climates (shallower roots, lower stem breakage resistance) increase vulnerability; lower homogeneity and milder slopes reduce it.
• Insect outbreaks: Higher standing biomass, older and taller trees increase vulnerability, whereas higher density and LAI associate with lower vulnerability (indicative of healthier, less water-stressed stands). Warmer temperature anomalies and drought (avg aTavg, avg aPcum, avg SPEI) increase vulnerability via reduced plant defenses and enhanced pest development. Cold-climate forests and high elevations are particularly susceptible (near thermal limits). Higher landscape heterogeneity (CV) reduces vulnerability.

Interactions: Friedman’s H-statistic indicated second-order interaction strengths of ~13–16% depending on disturbance. Forest–climate interactions dominated for fires and insects; landscape interactions were more influential for windthrows. Interactions were on average amplifying, increasing response peaks by 3–7% (up to ~25% for some feature combinations).

Current vulnerability (2009–2018 average PBL_rel): Windthrows 30.2% (29.4–30.3%), Fires 25.6% (25.0–25.8%), Insects 19.9% (19.4–20.0%). Spatial patterns: Windthrow vulnerability is high in Norway, northern British Isles, Portugal, Southern Europe’s mountains (Alps, Caucasus, Carpathians), up to ~40%; lower in southern Sweden/Poland (possibly post-storm biomass reductions). Fire vulnerability is high in Sweden, Finland, European Russia, southern Iberia, and Turkey (>35% locally); lower in wet central Europe and mountainous areas. Insect vulnerability increases northward and with elevation (up to ~30%). Uncertainty (SE) is <1% over most areas; some climate regions are under-represented in training data but sensitivity tests limiting to in-range climates changed European means by <1 pp.

Local sensitivity: Forest structure dominated sensitivity, especially for insects (Z_marg >60%); climate features dominated for fires (Z_marg >49%) and windthrows (>54%); landscape metrics contributed notably to insect vulnerability sensitivity (31%), with elevation important.

Trends (1979–2018): Europe-wide trends were negligible for fires (−4.9×10⁻³ % yr⁻¹) and windthrows (+1.4×10⁻³ % yr⁻¹) with mixed regional patterns. Insect vulnerability increased significantly (+8.8×10⁻² % yr⁻¹), with widespread significant increases and local trends >0.2 % yr⁻¹ in NE Fennoscandia and northern European Russia, driven by temperature in 91% of areas. A temperature anomaly of ~+0.5°C (vs. 1970–1990) around 2000 corresponded to a threshold beyond which insect vulnerability rose steadily, suggesting a tipping point linked to weakening plant defenses.

Overall Vulnerability Index (OVI): Current European forest OVI corresponds to ~58% (57.0–58.4%) of biomass vulnerable, or ~33.4 billion tonnes. Contributions: windthrows 40%, fires 34%, insects 26%. OVI trend: +4.2×10⁻² % yr⁻¹ (4.1–4.3×10⁻²), dominated by insects. Hotspots (high current OVI and increasing trend): cold climates of Finland, northern European Russia, the Alps, and warm-dry interior Iberia. Spatial–temporal correlation is largely controlled by insect vulnerability (R²=0.55).

The study demonstrates that forest structural, physiological, and mechanical properties, together with climatic context, govern European forest vulnerability to major disturbance agents. Machine learning revealed emergent, nonlinear, and interacting driver effects that align with ecological expectations while quantifying their relative influence and interactions. The absence of strong Europe-wide trends for fire and windthrow vulnerability suggests interannual climate variability dominates, while a clear warming-driven increase in insect vulnerability—exceeding a threshold around 2000—indicates reduced plant defense capacity and enhanced pest performance under warming. The identified hotspots at climatic envelope edges highlight regions where current conditions and recent warming jointly elevate susceptibility. These results provide actionable insights for forest management (e.g., managing stand structure, promoting species and structural diversity) to enhance resilience, and offer benchmarks to improve LSM disturbance modules and climate impact projections. The amplification effect of interacting drivers underscores the potential for future warming to exacerbate disturbances even without large changes in single drivers, with implications for carbon sequestration and biodiversity.

By integrating disturbance records, satellite-derived biomass dynamics, and machine learning, the study quantifies spatial–temporal patterns of European forest vulnerability to fires, windthrows, and insect outbreaks over 1979–2018, identifies key drivers and their interactions, and reveals a significant warming-driven increase in vulnerability to insect outbreaks since ~2000. On average, 58% of forest biomass (≈33.4 Gt) is potentially vulnerable to the combined disturbances, with windthrows contributing most, followed by fires and insects. Hotspots in northern/cold climates and some warm-dry regions are both highly vulnerable and experiencing increasing susceptibility. These findings emphasize the importance of adaptive forest management to modulate structural drivers and the need to incorporate dynamic, interacting processes into models and policy planning. Future research should integrate compound disturbance interactions, dynamic forest and landscape attributes, improved high-resolution wind and insect datasets for Europe, and species composition/diversity metrics to refine vulnerability estimates and link them with risk by incorporating hazard probabilities and exposure.

• Vulnerability (biomass loss conditional on disturbance) is not risk; occurrence probability and exposure are not included.
• Interactions between different disturbance agents and cascading effects are not modeled due to lack of compound-event observational data; OVI assumes independence and homogeneous spread within grid cells.
• Climate predictors vary over time in backward runs, but forest and landscape predictors are held static at 2009–2018 means, omitting stand dynamics, management, and adaptation over 1979–2018.
• Some climate–region combinations are under-represented in training data (e.g., cold-wet, warm-dry for windthrows; cold-dry for fires), increasing extrapolation uncertainty despite boundary checks and sensitivity tests.
• Insect outbreak models were trained on U.S. data (IDS-USDA) and transferred to Europe at PFT level; species-specific processes and regional differences may not be fully captured.
• Reanalysis wind speeds at coarse resolution underestimate localized extremes (e.g., tornadoes; storm Klaus), and such events were excluded.
• Potentially relevant predictors (e.g., species composition/diversity) were unavailable at scale and not included.
• Only events with BL >5% were modeled, possibly biasing toward more severe disturbances.