Fire may prevent future Amazon forest recovery after large-scale deforestation

The Amazon basin, containing about 40% of global tropical forests, underpins climate stability and ecosystem services by storing substantial carbon, sequestering CO2, and recycling rainfall. Intensifying land-use change, drought, heat stress, and fire have increased the risk of an Amazon forest die-back and potential tipping from evergreen tropical forest to savanna/grassland. Feedbacks between climate change, moisture recycling decline due to deforestation, and fire may accelerate this transition, while potential CO2 fertilization benefits are uncertain. The research question is whether fire, interacting with climate and land-use change, can prevent Amazon forest recovery after large-scale deforestation and thereby induce irreversibility. The study aims to quantify the role of fire in multi-century recovery trajectories under different atmospheric CO2 levels using a fully coupled Earth system model, to assess the risk of fire-controlled lock-in to a treeless state and implications for resilience and restoration.

Prior work identified the potential for Amazon forest die-back and highlighted fire as a stabilizer of grass-dominated states through positive feedbacks, as well as the importance of moisture recycling loss from deforestation reducing rainfall. Many studies relied on remote sensing, conceptual models, or reduced-complexity frameworks, often neglecting simultaneous fire–vegetation–climate interactions and longer-term recovery dynamics. Existing analyses typically span past decades or up to ~100 years, whereas tropical forest biomass, hydrological networks, and ecosystem functions may require centuries to recover. There remains uncertainty about the magnitude of CO2 fertilization effects under nutrient and thermal constraints. This study addresses gaps by using a fire-enabled, fully coupled Earth system model to examine multi-century recovery and feedbacks across a range of atmospheric CO2 concentrations.

Model: The Potsdam Earth Model (POEM; configuration CM2Mc-LPJmL v1.0) couples the GFDL CM2Mc atmosphere–ocean general circulation model (AM2 with MOM5; ~3° × 3.75°) to the dynamic global vegetation model LPJmL (0.5° × 0.5°) including the process-based fire module SPITFIRE. Coupling exchanges canopy humidity, soil/canopy temperature, roughness length, and albedo at hourly temporal resolution via GFDL’s Flexible Modeling System, with conservative interpolation between grids. LPJmL simulates establishment, growth, competition, and mortality of plant functional types (tropical evergreen tree, tropical raingreen tree, tropical grass), water and energy fluxes, and fire-induced mortality. Fire occurrence depends on fuel, moisture, vegetation structure, and ignition sources (human and lightning) and feeds back on vegetation and climate.

Experimental design: Three phases per experiment: (1) Spin-up to equilibrium (5000 years stand-alone LPJmL, then 1500 years fully coupled under pre-industrial, no land-use) to equilibrate soil carbon, vegetation, ocean, and climate. (2) Grassland phase: complete deforestation of the Amazon basin; only grass allowed (tree establishment disabled) for 250 years at constant atmospheric CO2 levels. (3) Recovery phase: tree establishment re-enabled for 250 years under the same constant CO2 levels. CO2 concentrations tested: 284 (pre-industrial), 450, 750, and 1200 ppm. Each scenario was simulated with fire enabled and with fire disabled. Control experiments (no deforestation) were run for 250 years under the same CO2 and ignition forcings, with and without fire, to assess equilibrium states and bi-stability.

Ignition forcing: To emulate a deforested Amazon socio-environment, human ignition potential and population density for each Amazon grid cell were randomly drawn from cells in the Cerrado biome; lightning frequency was reduced by prescribing average lightning from the low-convection Caatinga region. These ignition forcings were applied identically to scenario and control simulations with fire.

Key process representations and adjustments: Potential evapotranspiration via Penman–Monteith; canopy humidity and surface temperature parameterization; roughness length and albedo per LPJmL; sublimation parameterization to reduce high-latitude temperature bias. SPITFIRE improvements included optimization and a new fire danger index for more realistic South American fire dynamics. Outputs analyzed included aboveground biomass, burnt area fraction, temperature, precipitation, and spatial recovery patterns. Bi-stability was assessed by comparing end states of recovery runs versus controls under identical forcings.

• Deforestation-induced climate shifts: At 450 ppm, replacing forest with grass reduced aboveground biomass from ~20–30 to ~1–3 kg C/m2, increased central Amazon surface air temperature by ~4°C, and decreased basin-wide precipitation by ~1–2 mm/day relative to controls. Latent heat loss decline from reduced evapotranspiration dominated over albedo cooling, causing net warming and drying.
• Fire regime intensification: Warmer, drier conditions increased average annual burnt area by a factor of ~2–3 across much of the basin (especially southern Amazon), though extremely hot/dry northern areas experienced fuel limitation and reduced burnt area.
• Recovery without fire: After 250 years of recovery, biomass regained ~98% (284 ppm), 97% (450 ppm), 95% (750 ppm), and 80% (1200 ppm) of control biomass; locked grassland area was minimal (0–71 million ha; 0–10.5%).
• Recovery with fire: Fire drastically limited biomass recovery to ~40% (284 ppm), 50% (450 ppm), 50% (750 ppm), and 20% (1200 ppm) of controls; 56–82% of the Amazon area (353–515 million ha) remained locked as grassland after 250 years, depending on CO2. Mean burnt area during recovery stayed high (~18–20%) versus controls (~6–8%).
• Spatial dynamics: With fire, forests re-established first in the moist western Amazon near the Andes where Atlantic-sourced moisture precipitates; eastern Amazon recovery depended on slow reactivation of moisture recycling, with canopy closure reducing wind and fire risk and enhancing downwind rainfall. Large persistent grasslands experienced up to ~50% of grid-cell area burned in a year.
• CO2 effects: From 284 to 750 ppm, increased CO2 fertilization and water-use efficiency partly counteracted warming, aiding growth; at 1200 ppm, extreme heat and aridity suppressed tree growth, yielding widespread grassland/bare soil. Fire activity at 1200 ppm was not higher than at 750 ppm due to fuel limitation (more bare soil), indicating a shift from fire-dominated to temperature-controlled dynamics.
• Bi-stability and path dependence: Under identical climate and ignition forcings, controls (intact forest initial state) largely remained forested, whereas recovery from grassland frequently remained treeless, demonstrating history-dependent bi-stability between forest and grassland over large Amazon regions (especially eastern Amazon).

The study directly addresses whether fire inhibits Amazon forest recovery after large-scale deforestation by integrating coupled fire–vegetation–climate feedbacks over centuries. Results show that deforestation-induced warming and drying elevate fire activity, which suppresses tree establishment and favors fire-adapted grasses, thereby reinforcing aridity and fire through reduced evapotranspiration and moisture recycling. This positive feedback loop locks much of the basin into a grassland state despite the same boundary conditions that sustain forests when starting from an intact canopy, evidencing hysteresis and bi-stability. Spatial patterns align with known moisture dynamics: western Amazon, with strong Atlantic moisture influx and orographic precipitation, recovers earlier; eastern Amazon depends on re-establishing forest-mediated rainfall recycling. The interplay between CO2 fertilization and heat stress explains partial recovery at moderate CO2 and widespread failure at 1200 ppm. Persistent high burnt area fractions and incomplete biomass recovery even after 250 years underscore the long-term, path-dependent nature of the system. The findings highlight that initial conditions and fire disturbances are pivotal in determining whether reforestation and natural regeneration can succeed, implying that restoration without fire control may be ineffective in many regions under warming climates.

Using a fire-enabled Earth system model, the study shows that after large-scale deforestation, fire can prevent recovery across 56–82% (353–515 million ha) of the Amazon, leading to a history-dependent lock-in of grassland and establishing bi-stability with forest under identical external forcings. Without fire, forests largely recover over centuries, emphasizing fire’s critical role in irreversibility. Positive feedbacks—reduced evapotranspiration and precipitation, higher temperatures, and increased burning—stabilize treeless states, especially under higher CO2 levels. These results argue for stringent protection of existing forests, cessation of deforestation, reduction of CO2 emissions, and robust fire management; they question the feasibility of large-scale reforestation without controlling fire. Future research should: (1) explore a spectrum of partial deforestation patterns and sizes to map regeneration likelihood; (2) test alternative ignition scenarios and socio-economic pathways with explicit human–environment feedbacks; (3) improve model processes (nutrient limitation, leaf thermal constraints, plant trait diversity, higher spatial resolution) to refine recovery projections; and (4) assess targeted fire suppression and moisture recycling restoration strategies to overcome bi-stability.

• Model and projection uncertainties: Earth system models are tuned to historical conditions; future projections suffer from uncertainties in climate, carbon cycle feedbacks, and biases. Differences across CMIP-class models in warming and carbon fluxes can alter Amazon outcomes.
• CO2 fertilization likely overestimated: Current model version lacks nutrient limitations (e.g., nitrogen) and leaf thermal constraints, potentially inflating biomass recovery under elevated CO2.
• Vegetation trait simplifications: Limited representation of plant trait diversity (e.g., rooting strategies) may bias drought responses and recovery potential.
• Fire model constraints: Calibration relies on limited satellite-era data; coarse 0.5° resolution aggregates heterogeneous fire behavior, representing burnt area as cell fractions and not capturing fine-scale fire frequency variability.
• Ignition forcing assumptions: Human and lightning ignition patterns were idealized (Cerrado-like human ignitions, Caatinga lightning) and applied uniformly, including to controls where they are unrealistic for intact forests; used to ensure comparability but still an approximation.
• Scope: Not intended to predict exact recovery rates/patterns or prescribe specific policy measures; computational constraints limit exploration of full scenario space.