Road fragment edges enhance wildfire incidence and intensity, while suppressing global burned area

The study addresses how landscape fragmentation—particularly the edge effects created by roads—modulates wildfire behavior, including burned area (BA) and fire intensity, and why fragmentation can either inhibit or promote fire. Prior work shows mixed correlations between fragmentation and BA across ecosystems, but lacks mechanistic understanding. The authors hypothesize that edge effects alter fuel moisture, wind, and human ignitions in ways that can explain divergent BA responses. Using a process-based land surface model (ORCHIDEE-MICT-SPITFIRE), they aim to mechanistically link road-induced fragmentation to changes in fire probability and behavior, quantify global and biome-specific effects, and benchmark emergent model relationships against observation-based statistics. The work is motivated by extensive land-use change and infrastructure expansion, the importance of fire in Earth system dynamics, and the need for policy-relevant quantification of fragmentation-fire impacts.

The paper synthesizes evidence that fragmentation can decrease BA in grassland savannahs and increase BA in forests, based on meta-analyses and regional studies. Fragment edges are known to influence microclimate, fuel moisture, wind exposure, and biotic processes, which could affect ignition and spread. Road density (RD) has been identified statistically as a strong negative global predictor of BA, suggesting fire-inhibiting effects in some regions, but this association lacks mechanistic attribution. Empirical studies document both positive and negative fragmentation–fire relationships depending on biome and management context (e.g., Amazon interior vs. Cerrado). Prior datasets largely focus on forest fragmentation, leaving gaps for non-forest biomes where most fires occur. The review highlights challenges in directly measuring fragmentation impacts due to counterfactual constraints and data limitations, motivating a modeling approach to isolate edge-effect mechanisms.

Model and proxy: The authors implement fragmentation edge effects in the terrestrial land surface model ORCHIDEE-MICT-SPITFIRE (a CMIP6-participating model with fire module SPITFIRE). Fragmentation extent is proxied by road density/length (RL) using the GRIP-derived global roads dataset (Meijer et al., 2018) at ~5 arc-minute resolution, regridded to 0.5°. They compute an Average Edge Distance (AED), interpreted as the average Euclidean distance from patch edge to interior, assuming each grid cell comprises synthetic circular patches derived from RL and grid area: AED_c = (2 × Area_c × f(Cont)) / ∑(RL_c). To partially correct undercounting of roads, RL is uniformly doubled (noted uncertainty). Urban roads are removed from fragmentation calculations via regression against the grid cell urban area fraction (UAF), producing an ‘effective’ RL (thus AED) that excludes an urban contribution unlikely to affect wildland BA.

Fragment edge-effect processes implemented:

• Fire size limitation by patch size: Maximum individual fire size per event is limited by average patch area derived from AED, unless spread thresholds are exceeded. Multipliers allow limited exceedance of patch size: n_pat(forest)=1.0; n_pat(grass)=1.25. Spread thresholds: (i) Forest crown-fire spread allowed when simulated scorch height exceeds canopy base (flame height > canopy base; conditional FST_TREE = True), lifting size limits. (ii) Grasslands: when area-specific fuel mass exceeds a threshold (~2.4–2.5 t ha⁻¹ wet mass; FST_GRASS = True), fragmentation size limits are lifted.
• Human ignitions enhancement: Human ignition function in SPITFIRE (nonlinear with population density, PopD) is augmented by edge area proportional to the edge-to-patch area ratio, assuming a 1 m ignition edge depth (ED_Humig = 1 m). The added ignition probability scales with number of patches and fractional grid edge area, increasing ignitions especially at high fragmentation.
• Fuel moisture drying near edges: Edge drying implemented via a 20 m edge depth (ED_Moisture = 20 m) with a linear edge-to-interior gradient, applying an average 12.5% reduction (half of a 25% edge vs. interior gradient) to fuel wetness and ignition moisture thresholds for 1- and 10-hr fuel classes in proportion to the edge-to-patch area. The 100-hr class is unaffected.
• Wind infiltration at edges (forests only): Reduced surface roughness near edges increases windspeed within a 16 m edge depth (ED_WIND = 16 m; effective 4 m assuming wind from one of four idealized directions). The model reduces the forest wind reduction factor proportionally to the edge area fraction, potentially increasing spread rate, intensity, and emissions in fragmented forests. This wind effect is not applied to grasslands.

Simulations and benchmarking:

• Main simulations at 0.5° for 2000–2013 with all fragmentation functions activated, and a control (CTRL) without fragmentation. Model spun up for vegetation equilibrium (historical forcing loops). Climate forcing: CRU-NCEP v8. Vegetation: imposed from ESA-LUH2 PFT distributions.
• Sensitivity ensemble (AEDF2): Ten global simulations homogenizing AED across all grid cells at factor-of-two levels (AED = 39.06 to 20,000 m), run for 2001–2003, to assess monotonicity and biome-specific sensitivity to fragmentation doubling.
• Biome attribution uses the PFT with maximum fire CO2 emissions per grid (since BA is not PFT-disaggregated) to categorize grids into tropical/temperate/boreal forests and C3/C4 grasslands.
• Statistical comparison: The model’s emergent relationship between logit-transformed monthly BA and sqrt(RD) is compared to the observation-based global regression (Haas et al., 2022). Additional diagnostics isolate grids where fragmentation explicitly constrains individual fire size.

Data handling highlights: RL (Meijer et al., 2018; adjusted), urban fraction (Gao & O’Neill, 2020/2021), fire records (MODIS, FireCCI, FRYv2.0), and regional BA/fragmentation trend products (e.g., Amazon/Cerrado; Rosan et al., 2022). The AED input is static (~2017), acknowledging limitations for temporal analyses.

• Global burned area impacts: Fragmentation caused both increases and decreases in simulated mean annual BA. Gross decreases summed to −30 MHa yr⁻¹ and gross increases to +25.6 MHa yr⁻¹, equivalent to roughly −3.25–6.5% and +2.25–5.5% of satellite-observed annual BA, respectively. Net effect is a global BA reduction consistent with the abstract’s ~−1.5%.
• Spatial magnitude: BA changed by more than ±10% in about 17% of burned grid cells and by more than ±25% in about 7% of burned grid cells.
• Regional patterns: Largest proportional BA decreases occur in highly fragmented, densely populated regions (e.g., NW Europe, California, NE USA). BA increases are simulated in moderately fragmented, moderately populated areas (e.g., Indonesia, eastern Brazil, parts of the Mediterranean, and Eastern Congo Basin). Mediterranean regions prone to summer fires show substantial BA increases due to enhanced ignition, drying, and wind effects.
• Validation with observations: The emergent model regression of logit(BA) vs. sqrt(RD) reproduces observed slope and intercept from Haas et al. (2022). Overall R² is low for all cells (Spearman’s rho = −0.07; R² = 0.05; p<0.001) but improves when focusing on grids where fragmentation explicitly limits individual fire size and when excluding urban roads in RD.
• Tropical frontiers: In Amazonia, simulated BA increases align along the Trans-Amazonian highway and match observed patterns of increased BA with increased fragmentation in rainforest interiors; Cerrado shows BA decreases or weak response, consistent with observational trends. In Indonesia/Malaysia, 58% of grids with observed BA increases are matched by simulated BA increases; among these, 67% coincide with significant deforestation/plantation inception. Mismatches cluster in peat regions not represented by the model.
• Population density modulation: Fragmentation-driven BA response vs. PopD is biome-specific. Tropical forests show a hump-shaped response peaking near ~0.5 individuals km⁻², with increases at low-to-moderate PopD followed by declines. Temperate forests peak near ~50 individuals km⁻²; boreal forests show little PopD sensitivity; temperate grasslands increase slightly at low PopD and decline strongly at high PopD.
• Intensity and emissions: Despite net global BA reduction, fragmentation can increase area-specific fire CO₂ emissions (a proxy for intensity) in many forested regions, particularly boreal and tropical forests, indicating decoupling of BA and intensity. Globally, fragmentation reduces total fire CO₂ emissions by ~1% (−0.02 PgC yr⁻¹).
• Sensitivity (AEDF2): Doubling fragmentation reduces BA on average across biomes: −7.5% (tropical forest), −15% (temperate forest), −19% (boreal forest), −30% (C3 grasslands), −22% (C4 grasslands). Tropical forests show greater resistance to BA reduction; achieving a −5% BA decrease in an average tropical forest grid may require ~+2.5 km km⁻² of additional road length (~6,000 km per 0.5° grid). Areas with the smallest average BA decreases across AED levels are most susceptible to fragmentation-driven BA increases.

The results support the hypothesis that fragmentation edge effects mechanistically explain divergent BA responses: physical limitation of fire size by patches and human suppression correlate with BA reductions, while increased human ignitions, edge drying, and enhanced winds at edges can elevate BA and intensity, especially in forests and tropical deforestation frontiers. The model reproduces the observed negative global BA–RD relationship without tuning, lending credibility to the implemented processes. Importantly, fragmentation can reduce BA while increasing intensity and emissions in forests, implying more severe but spatially constrained fires, consistent with recent observations of rising intensity despite declining global BA. The sensitivity experiments reveal that while average BA declines with increasing fragmentation, tropical and boreal forests remain relatively susceptible to BA increases under certain fragmentation levels and human contexts. These insights are policy-relevant: rising road infrastructure and land-use change, particularly in the tropics, could shift fragmentation from a net global fire inhibitor to a net driver of wildfire activity, with implications for carbon emissions, suppression planning, and risk management. However, interpretation must consider data and model uncertainties (road undercounting, static AED, unrepresented peat fires, and parameter assumptions).

The study introduces a parsimonious, mechanistic representation of fragmentation edge effects in a global land-fire model using road density as a proxy. It reproduces key observation-based relationships between fragmentation and burned area, quantifies how edge-driven processes (ignitions, drying, wind) modulate BA and fire intensity, and reveals that fragmentation can simultaneously suppress BA and increase intensity in forests. Globally, fragmentation modestly reduces BA and fire CO₂ emissions, but in tropical deforestation frontiers it increases BA, suggesting that continued degradation could tip fragmentation toward a net global fire driver. The framework enables identification of vulnerable biomes/regions, informs mitigation (e.g., managing edge effects), and supports scenario-based forecasting of future fire regimes under expanding infrastructure. Future work should incorporate dynamic road/fragmentation time series, distinguish road types, improve process resolution (e.g., peat and soil burning, fuel-load changes), refine parameterizations with more empirical data, and better account for cultural fire stewardship and land-use interactions.

• Proxy and data: Road length/density is an imperfect proxy for fragmentation; GRIP underrepresents small and unofficial roads (potentially by factors up to ~8 regionally), and other linear features (rivers, topography) are omitted. AED is static (~2017), limiting temporal attribution for 2000–2013 simulations.
• Process simplifications: Single-shape (circular) average patch assumption; uniform road-type effects; simplified edge depths (1 m for ignitions, 20 m for moisture, 16 m for wind) and linear gradients; limited fuel classes affected by drying; wind-edge effects applied only to forests.
• Model scope: Does not represent peat/soil burning (notably important in SE Asia), nor the interior land-use effects (e.g., biomass depletion) that may reduce fuel and BA; BA not PFT-disaggregated (biome assignment via emissions proxy); small-fire detection biases in observations complicate validation.
• Statistical comparison: Low global R² due to model-observation differences (e.g., undetected small fires, urban road treatment), and potential confounding with suppression and non-fragmentation correlates of RD.
• Generalizability and culture: Cultural and management-specific fire practices are not explicitly represented; parameter values carry uncertainty and may vary by biome and context. Absolute magnitudes are uncertain; relative signs/patterns are more reliable.