Tropical forest restoration under future climate change

Tropical forest restoration area

To determine the geographic distribution of land available for tropical forest restoration, we used a widely applied global forest restoration map². This dataset limits potential restoration area to regions that are biogeophysically suitable for forest, and excludes croplands. To define the tropics, we masked the potential restoration map with the following three ecoregions from the Ecoregions2017 vegetation map³⁴: ‘Tropical and Subtropical Moist Broadleaf Forests’, ‘Tropical and Subtropical Dry Broadleaf Forests’, and ‘Tropical and Subtropical Coniferous Forests’. The resulting restoration mask includes all tropical and subtropical forest ecoregions with some that are outside the tropical latitudes, but excludes wetlands and high mountain areas (Extended Data Fig. 4). The restoration mask was converted from a presence–absence raster at its native ~350 m resolution to a 0.5° geographical grid by aggregating to the fraction of each 0.5° grid cell available for restoration. Any uncertainties in the allocation of restorable area, distinguishing crop and pasture, and forest to non-forest classification from the original forest restoration map were also implicitly included in our restoration extent. While the resulting restoration area is relatively small, its spatial distribution is representative for most of the humid tropics.

To prioritize for carbon uptake capacity, we selected all grid cells with restoration area greater than 1 ha and ranked these by carbon storage density (above ground and below ground; g m⁻²) at 2100 under the default scenario. We then selected the top n grid cells with greatest carbon density until cumulatively 64 Mha of restored area was reached. Similarly, for cost we calculated the restoration cost for each grid cell following ref. ²⁷ and sorted the grid cells by their cost, beginning with the lowest value, until 64 Mha were reached. To consider the combined impact of carbon uptake and restoration costs, we divided our restoration cost layer by the total carbon uptake per grid cell from restoration and ranked the cost per carbon uptake from cheapest to most expensive, selecting the n grid cells with the lowest values until 64 Mha were reached. We then used the selected grid cells to mask carbon uptake under the various climate change and CO₂ fertilization scenarios. To factor in climate change in the prioritization process, we used the same restoration cost layer but used the carbon density and total carbon uptake layers with climate change impacts in CO2₂₀₁₄ for the year 2100.

Vegetation model

We used the LPJ-LMfire DGVM¹⁹, a version of the Lund-Potsdam-Jena DGVM (LPJ)³⁵. LPJ-LMfire is driven by gridded fields of climate, soil texture and topography at 0.5° resolution, and with a time series of atmospheric CO₂ concentrations (see Supplementary Information). To simulate land use, LPJ-LMfire separates grid cells into fractional tiles of ‘unmanaged’ land that has never been under land use, ‘managed’ land, and areas ‘recovering’ from land use³⁶. Restoration removes land from the ‘managed’ tile and transfers it to the ‘recovering’ tile; land is never reallocated to the ‘unmanaged’ tile. The tiles are treated differently with respect to wildfire: on the ‘unmanaged’ and ‘recovering’ tiles, lightning-ignited wildfires are not suppressed, while fire is excluded from ‘managed’ tiles. For our analysis of total carbon (above and below ground), we only used the ‘recovering’ tile.

Climate data

Climate forcing used to drive LPJ-LMfire comes from the output of 13 GCMs in simulations produced for the CMIP6 Supplementary Table 2 (refs. ^37,38). For each GCM, we obtained simulations for the historical period (1850–2014) and four future SSPs (SSP1-26, SSP2-45, SSP3-70 and SSP5-85 covering 2015–2100). We used only GCMs that archived all seven climate variables needed to run LPJ-LMfire: 2 m temperature (tas, K), precipitation (pr, kg m⁻² s⁻¹), convective precipitation (prc, kg m⁻² s⁻¹), cloud cover (clt, %), minimum and maximum daily temperature (t_min, t_max, K), and 10 m surface wind speed (sfcWind, m s⁻¹) (Supplementary Fig. 2). For each model, we concatenated the historical simulation with a future scenario, calculated anomalies with respect to 1971–1990 and added those to observed 30 year climatologies to create bias-corrected monthly climate time series covering 1850–2100 (see Supplementary Information). Where multiple ensemble members were available from a GCM, we chose the first simulation.

Simulation protocol

We drove LPJ-LMfire with the GCM simulations described in the previous section, and the same atmospheric CO₂ concentrations and land use boundary conditions as those used in the CMIP6 simulations. All forcings cover the historical period (1850–2014) and the individual future SSPs (2015–2100). Each LPJ-LMfire simulation was initialized for 1,020 years with 1850 atmospheric CO₂ and land use, and the 1850s climatology of each CMIP6 GCM. This was followed by simulations with transient climate from 1850 to 2100 for each CMIP6 GCM under each of the four SSPs. For each the 13 CMIP6 GCMs running each of the SSP scenarios, we conducted two CO₂ experiments (CO2₂₀₁₄ and CO2_free) and two fire experiments. In total, we ran 221 vegetation model simulations covering the range of future climate, CO₂ and fire scenarios.

Atmospheric CO₂ in these simulations either followed the CMIP6 historical and SSP trajectory for the entire 1850–2100 run (CO2_free), or followed the historical CMIP6 trajectory until 2014, and was then fixed at 2014 concentrations for the remainder of the simulation (CO2₂₀₁₄). This allowed us to test the vegetation response to future climate change in the absence of additional CO₂ fertilization of photosynthesis. Our simulations ended with the standard SSP projections in 2100, 80 years after restoration begins. We therefore could not assess the fate of restored carbon beyond that point. On the basis of the trends in the multi-model mean carbon uptake rates, we estimated that only under severe climate change will carbon storage be reduced shortly after 2100 in CO2₂₀₁₄.

In control simulations, land use followed the historical CMIP6 trajectory until 2014, after which it was fixed under 2014 conditions until 2100. Land use after 2014 was fixed at 2014 levels because it is the last year with common land use between all scenarios, which allowed us to identify future climate change impacts on restoration permanence and avoid influences from land abandonment and expansion prescribed in the different SSP scenarios.

In the restoration experiments, land use also followed the historical CMIP6 trajectory until 2014, but then diverged: cropland extent remained at 2014 levels until 2100, while pasture (or non-cropland land use) remained constant from 2014 to 2020 and was then linearly reduced by the restoration area from 2020 to 2030. From 2030, land use remained constant at that lower level until 2100. The amount of restoration in a grid cell was limited by the pasture area, that is, once all of the available pasture area had been restored, no additional restoration took place. Because it is highly unlikely to be practical to restore the entire target area of tropical forest at once, we linearly increased the restoration area from 2020 to 2030, which caused an expansion-driven increase in carbon uptake over the 11 year period (Extended Data Fig. 1). This means that two factors controlled carbon uptake over time in our experimental design: first the expansion of the restoration area, accounting for approximately 19.7 Pg C, and second the long-term effect of carbon accumulation (Extended Data Fig. 5).

Primary climate change impacts, such as drought and heat stress that reduce carbon uptake, were implicitly included in the climate forcing data, while secondary climate change impacts from wildfire were simulated by LPJ-LMfire on the basis of climate. To quantify the contribution of wildfire on the carbon storage from restoration, we repeated the simulations described above with fires turned off in LPJ-LMfire.

Restoration opportunity index

We created a restoration opportunity index to evaluate the suitability of locations for restoration on the basis of the ability for restoration to result in net carbon uptake over 2020–2100 and to store this carbon without episodes of major loss. For each of the 13 realizations of the four SSPs in the CO2₂₀₁₄ experiment, we identified all restoration grid cells (1) that had a net carbon uptake by 2100 relative to 2030, and (2) where temporal reductions in total carbon storage over 2030–2100 were <10% of the 2030–2100 mean. For each simulation, grid cells meeting these criteria were then used to mask the rank sum of carbon uptake (grid cells ordered high to low) and opportunity cost (grid cells ordered low to high). We then summed this quantity across all 13 future simulations from the different GCMs, and standardized the resulting layer from 0–1. A score of 1 means that the restoration opportunity index is 1 under all future climate model simulations.

Source: Ecology - nature.com