Emergent vulnerability to climate-driven disturbances in European forests

European forests, covering over 2 million km² (33% of the continent's land surface), provide crucial ecosystem services. While resilient to long-term environmental changes, they are vulnerable to sudden disturbances due to trees' long lifespans, limiting rapid adaptation. Understanding and quantifying this vulnerability is crucial for effective climate impact assessment and adaptation strategies, particularly given expected climate changes that may significantly increase natural disturbance risks. Land surface models (LSMs) are incorporating forest disturbances, but current formulations are incomplete due to limited understanding of ecological processes. While compilations of tree mortality events offer large-scale coverage, coarse resolution masks spatial variability and limits environmental control assessment. High-resolution satellite datasets provide improved spatial coverage for mapping forest disturbances, but attributing disturbance agents remains challenging. Previous studies have explored tree mortality's dependence on environmental controls using satellite data, but typically consider limited drivers and regional-level aggregation, often using a priori knowledge of functional relationships. This study leverages machine learning to overcome these limitations, assessing the vulnerability of European forests to multiple interacting factors at a large scale without assuming explicit functional relationships.

Existing literature highlights the increasing frequency and intensity of forest disturbances in Europe due to climate change. Studies on specific systems provide insights into the complex interactions between disturbances and environmental controls, but extrapolating to larger areas remains unclear. Large-scale analyses using compiled reports of past tree mortality events often suffer from coarse resolution, masking important spatial variability. Satellite data, with its high spatial resolution and global consistency, offers potential for large-scale comparative studies. However, directly attributing disturbance agents to remotely sensed data presents a significant challenge. Recent studies have employed satellite retrievals to investigate tree mortality’s dependence on environmental controls but generally with limited sets of drivers and regional aggregations. They often relied on pre-defined assumptions about functional relationships between vulnerability and drivers, neglecting potential amplification or dampening effects from complex interactions at local scales. The integration of machine learning techniques with readily available Earth observation data offers a promising approach to analyzing ecosystem responses to multiple interacting factors, without pre-conceived functional relationships. However, large-scale implementation for multiple disturbance types remains underdeveloped.

This study investigated the vulnerability of European forests (including Turkey and European Russia) to fires, windthrows, and insect outbreaks from 1979–2018. Random forest (RF) regression was used as a machine learning method to identify relationships between vulnerability (expressed as relative biomass loss following a disturbance, BLrels) and a suite of forest, climate, and landscape metrics. Spatially explicit databases of forest disturbance events were integrated with satellite-based and reanalysis products to retrieve these variables. RF models were developed separately for different plant functional types (PFTs) and disturbances, applied annually across Europe. Factorial simulations isolated key drivers. Model evaluation involved splitting disturbance data into calibration (60%) and validation (40%) sets. Performance was assessed using R², RMSE, and PBIAS. Variable importance metrics and partial dependence plots (PDPs) were used to understand the influence of individual environmental factors and explore ecosystem response functions. Friedman’s H-statistic quantified second-order interactions. Finally, vulnerability was integrated across disturbance types and time to identify forest hotspots of high susceptibility. Vulnerability estimates are conditional on disturbance occurrence and do not reflect risk levels (integrating occurrence probability and exposure). Data sources included the European Forest Fire Information System (EFFIS), the European Forest Windthrow dataset (FORWIND), the National Insect and Disease Survey (IDS-USDA) database, a 100-m above-ground biomass map (2010), Global Forest Change (GFC) maps, and various satellite and reanalysis products. Annual biomass time series were reconstructed integrating static biomass maps with forest cover change data. Biomass loss was quantified as the relative change in biomass before and after a disturbance.

The best RF models explained 34–49% of variance in relative biomass loss (R²). Variable importance analysis revealed key drivers: for fires, increased biomass, tree density, age, low precipitation, high temperature, and low moisture index increased vulnerability; for windthrows, high biomass, tree age, height, wind speed, and saturated soils increased vulnerability; for insect outbreaks, high standing volume, average temperature, and droughts increased vulnerability, while high density and LAI showed lower vulnerability, possibly reflecting good forest health. Second-order interactions amplified vulnerability peaks by 3–7% on average. Spatial vulnerability patterns varied: windthrows were high in Norway, British Isles, Portugal, and southern Europe; fires were high in Sweden, Finland, European Russia, southern Iberia, and Turkey; insect outbreaks increased from south to north and low to high elevations. Local sensitivities varied geographically and across disturbances. Forest characteristics were especially important for insect outbreaks. At the European level, no substantial trend was found for fire and windthrow vulnerability, but insect outbreak vulnerability increased substantially (8.8 × 10⁻²% year⁻¹), particularly after 2000, largely driven by rising temperatures, suggesting a temperature threshold was passed around 2000, reducing plant defenses. Combining all disturbances, over 2009–2018, about 58% of European forest biomass (33.4 billion tonnes) was potentially vulnerable, primarily to windthrows (40%), fires (34%), and insect outbreaks (26%). Overall vulnerability increased by 4.2 × 10⁻²% year⁻¹, driven by insect outbreaks. Hotspots of high vulnerability and increasing trends were found in cold climates (Finland, northern Russia, Alps) and warm-dry forests (Iberian Peninsula).

The findings address the research question by quantifying the vulnerability of European forests to climate-driven disturbances and identifying key drivers. The results highlight the substantial current vulnerability of European forests and the significant impact of climate change, especially the increased vulnerability to insect outbreaks after 2000 due to rising temperatures affecting plant defenses. The large-scale spatial analysis and multi-disturbance assessment offer a novel approach, informing forest management strategies and climate change mitigation and adaptation policies. The findings emphasize the interplay between forest characteristics, climate, and landscape factors, underlining the importance of considering these complex interactions in forest management and climate change projections.

This study provides a comprehensive assessment of the vulnerability of European forests to multiple climate-driven disturbances, revealing significant current vulnerability and increasing trends driven primarily by rising temperatures. Forest management practices that enhance forest resilience and adaptation are essential. Future research should focus on refining disturbance interaction models, exploring the long-term impacts of cascading effects, and enhancing the predictive capacity of LSMs to represent forest dynamics under climate change.

The study's vulnerability estimates are conditional on disturbance occurrence and do not incorporate disturbance probability or exposure. The model does not explicitly account for interactions among multiple disturbances due to data limitations, potentially underestimating the overall impact. Extrapolation outside the range of observed climatological conditions may also introduce uncertainty. The use of a single insect disturbance class could potentially mask distinctive features of different insect groups.