Boreal tree species diversity increases with global warming but is reversed by extremes

The study addresses how boreal tree species diversity has changed in recent decades under climate change, and how diversity relates to productivity and temporal stability of boreal forest ecosystems. Boreal forests, covering about 30% of global forest area, are experiencing rapid warming, increasing fire activity and drought, yet large-scale, time-varying assessments of tree species diversity have been lacking. Plot-level studies suggest diversity underpins carbon sink functionality and stability, but their relationships may vary across environmental gradients, limiting generalization. The authors aim to generate spatially continuous, temporal maps of boreal tree species diversity for 2000, 2010 and 2020, evaluate climatic and environmental determinants of diversity and its change, and quantify associations between diversity and multiple independent carbon flux, stock and stability indicators. This work is important for understanding the resilience and sustainability of the boreal carbon sink under ongoing climate change.

Prior work demonstrates forests as a persistent carbon sink and highlights the role of tree diversity in productivity and stability at plot scales. Global changes in forest extent and climate-driven trade-offs (growth, mortality) have been documented, along with plot-level shifts in tree species. Global statistical models have produced static, coarse-resolution diversity maps, while satellite-based spectral heterogeneity approaches have shown promise for predicting tree diversity with improved spatial resolution and temporal dynamics. However, continuous, large-scale mapping of tree species diversity and its temporal changes remains limited, impeding generalized assessments of how diversity contributes to ecosystem functioning. Boreal forests, with relatively low species richness and high sensitivity to warming and disturbance, may undergo tipping points and biome shifts, emphasizing the need for a high-resolution, time-varying diversity assessment.

The authors developed a satellite remote sensing and field-validated framework to map boreal tree species diversity (Shannon index, H′) at 30 m resolution for epochs centered on 2000, 2010 and 2020. Data: 5,312 field plots (190,516 trees, 254 species) from national inventories and public datasets across boreal regions; Landsat surface reflectance imagery (30 m) from Landsat-5/7/8, composited for growing seasons (May–October) over 1999–2001, 2009–2011 and 2019–2021; environmental variables (temperature, precipitation from ERA5; soils: sand fraction, organic carbon, cation exchange capacity; disturbances: fire frequency from MODIS burned area, human population density; topography: ASTER DEM); stand age map. Preprocessing and feature engineering: Forest masking (tree cover ≥30%); object-based image segmentation via a simple non-iterative clustering algorithm to delineate forest segments; computation of 217 predictors per segment including spectral heterogeneity metrics (coefficient of variation, spectral angle/gradient, texture dissimilarity and entropy), spectral and temporal statistics for bands and vegetation indices, monthly bands and six vegetation indices (May–October). Training data and augmentation: Segment-level spectral metrics matched to plot H′; data augmentation via averaging within 1×1, 3×3, and 5×5 windows; after filtering, 20,100 samples (each with 217 metrics); split into 70% train, 10% validation, 20% test; ten-fold cross-validation. Predictive modeling: InceptionTime deep learning regression (modified architecture with removed max pooling, reduced inception modules, added dropout, adjusted residuals); hyperparameter tuning via grid search and cross-validation. Model performance assessed with R² and RMSE; best model used to predict H′ maps for 2000, 2010, 2020. Post-processing: Uncertainty reduction using standard error thresholds derived from repeated plot measurements; pixels with predicted H′ standard deviation above 0.20 (95th quantile) flagged as uncertain (~10% area, ~1.53 million km²) and excluded. Analyses of change and drivers: Per-pixel H′ differences (2020 minus 2000) classified as gain (>0.01), loss (<−0.01) or no distinct change (between −0.01 and 0.01). Boosted regression trees (BRT) used to attribute spatial variability in H′ to predictors; 92,416 samples with ≥50 km spacing; ten model runs; partial dependence analyses. Trend analyses related H′ changes to trends in temperature, precipitation, fire frequency and stand age. Carbon associations: Multiple independent indicators of carbon fluxes (MODIS NPP, kNDVI) and stocks (VOD Ku-band, AGB from SMOS L-VOD as AGB_1, and integrated LiDAR/optical/microwave AGB_2), plus temporal stability of biomass (mean AGB/SD over five-year windows), were related to H′ (levels and trends) using multiple linear regression with standardized coefficients, controlling for climate, disturbances, soil, topography and stand age; sampling with ≥50 km spacing to reduce spatial autocorrelation; significance at P<0.05. Analyses conducted in R, Python, Google Earth Engine and QGIS.

• Model performance: The InceptionTime model predicted H′ with R² = 0.77 and RMSE = 0.12 across boreal forests.
• Overall diversity change: Mean H′ increased by 12% from 2000 to 2020 (0.41 ± 0.14 to 0.46 ± 0.16). Increases were 5% ± 2% (2000–2010) and 7% ± 3% (2010–2020).
• Spatial extent of change: H′ gains occurred in 53% of boreal forest area (~8,165,000 km²); losses in 17% (~2,684,000 km²); no distinct change in 20% (threshold ±0.01 in H′). Gain-to-loss pixel ratio ≈ 3:1; frequency of higher H′ (>0.4) increased over time.
• Regional patterns: High diversity in the Okhotsk–Manchurian taiga (mean H′ 0.56 ± 0.16) and Scandinavian–Russian taiga (0.55 ± 0.18); low in Northeast Siberian taiga (0.28 ± 0.07). In North America, highest in the eastern forest–boreal transition (0.77 ± 0.20), then Central Rockies (0.49 ± 0.16) and Mid-Continental Canadian forests (0.44 ± 0.12); lowest in Alaska–Yukon lowland taiga (0.33 ± 0.09) and Northern Canadian Shield taiga (0.31 ± 0.08).
• Magnitude of regional changes: Gains concentrated in Scandinavian–Russian taiga (+35% ± 16%, +0.18 ± 0.08), Okhotsk–Manchurian taiga (+33% ± 17%, +0.17 ± 0.09), and eastern forest–boreal transition (+27% ± 11%, +0.19 ± 0.08). Losses in Kamchatka Mountain forest (−25% ± 16%, −0.13 ± 0.08) and West Siberian taiga (−26% ± 10%, −0.13 ± 0.05). Gains tended to occur in warmer regions, losses in colder regions.
• Drivers of spatial variability (BRT, average R² ≈ 0.62): Temperature was the dominant predictor (relative importance ~53%), followed by precipitation (~21%); elevation (~9%), population density (~6%), fire frequency (~4%), cation exchange capacity (~3%), topsoil organic carbon (~2%), sand fraction (~2%). Partial dependence showed: diversity increases with mean seasonal temperature up to ~12 °C then saturates; precipitation positive above ~100 mm, saturating near ~400 mm; elevation weakly positive up to ~700 m then weakly negative; population density slightly positive up to ~1.5 people km⁻² then slightly negative.
• Drivers of temporal change: H′ trends were negatively related to temperature trends (Spearman ρ = −0.51, P < 0.001); positive diversity trends occurred with modest warming, but warming rates >0.065 °C yr⁻¹ were associated with diversity declines (notably in northeastern Siberia). H′ trends were negatively related to increasing fire frequency (ρ = −0.26, P < 0.001) and weakly positively related to precipitation trends (ρ = 0.10, P = 0.002), with saturation under extreme precipitation changes. Stand age showed a negative relationship with diversity change; young forests exhibited larger diversity gains than mature forests.
• Richness vs evenness contributions: In eastern forest–boreal transition and Canadian Shield forests, richness and evenness contributed roughly equally to H′ changes; in Northern Canadian Shield and Central Rockies, evenness contributions exceeded richness (e.g., β 0.47 vs 0.21; 0.56 vs 0.28).
• Carbon associations: Across spatial and temporal domains, tree diversity showed significant positive associations with carbon fluxes (NPP, kNDVI, VOD Ku) and stocks (AGB_1, AGB_2), and with temporal stability of biomass. Diversity trends were positively associated with trends in NPP, kNDVI and AGB_2. Climate variables had the largest effects on carbon indicators, followed by diversity, disturbances and stand age; fire impacts were generally negative and stronger than population density effects. Stand age was negatively associated with flux trends and positively with stock levels. Elevation tended to correlate positively with most carbon indicator trends.

The study demonstrates that boreal tree species diversity has generally increased over two decades, largely in response to climate warming, but diversity gains diminish and reverse under rapid or extreme warming. This clarifies the large-scale temporal response of boreal diversity to climate change and highlights nonlinear temperature effects with a threshold beyond which diversity declines, compounded by increasing fire activity. The high-resolution, time-varying diversity maps link rising diversity to enhanced forest productivity, biomass stocks and stability, indicating that diversity preservation aligns with maintaining the boreal carbon sink. The object-based, spectral-heterogeneity-driven, deep learning approach captures fine-scale heterogeneity across environmental gradients better than prior coarse, static models. However, warming-driven biome transitions and changes in species dominance may alter future productivity and stability trajectories, and young forests are particularly dynamic, showing larger diversity gains.

The authors provide the first spatially continuous, 30 m resolution maps of boreal tree species diversity for 2000, 2010 and 2020 and quantify their changes and drivers. They show an average 12% increase in H′ across the boreal zone, with gains in over half of the forest area, but identify a critical warming-rate threshold (~0.065 °C yr⁻¹) beyond which diversity declines, especially where fire activity increases. Diversity is positively associated with carbon fluxes, biomass stocks and temporal stability, indicating co-benefits for ecosystem functioning and carbon sequestration. Future research should incorporate shifts in dominant species, structural and functional trait diversity, and improved spectral methods (e.g., unmixing, radiative transfer, data fusion) to reduce uncertainties. Monitoring potential biome shifts at ecotones and assessing interactions among warming, drought, and disturbance regimes will further clarify how diversity underpins the resilience of the boreal carbon sink.

Uncertainties arise from satellite data quality and environmental conditions (lighting, shadows), and influence of understory vegetation on reflectance in sparse forests. Although segmentation and multiple indices were used to minimize greenness and tree cover effects, residual biases may remain. Approximately 10% of the area was excluded due to high prediction uncertainty. The analysis did not explicitly account for shifts in dominant species composition, which may affect productivity and stability responses. Associations with carbon indicators may be influenced by differences among data sources and retrieval methods. Incoming solar radiation was excluded due to collinearity with temperature in boreal regions. Threshold-based classification of change (±0.01 H′) may miss subtle ecological shifts.