Substantial carbon removal capacity of Taiga reforestation and afforestation at Canada’s boreal edge

Abstract

Large-scale tree planting at Canada’s northern boreal forest edge offers potential for climate change mitigation, but this capacity is uncertain due to a lack of spatially explicit models. This study quantifies the carbon removal capacity of tree planting at the northwestern boreal edge using a carbon budget model and Monte Carlo estimates. Combining satellite inventory data with probabilistic fire regimes, we simulated total ecosystem carbon under scenarios considering fire return intervals, land classes, planting mortality, and climate variables. Our results indicate that planting ~6.4-32 million hectares could sequester ~3.88-19.4 Gigatonnes of carbon dioxide equivalent over 75 years, with the Taiga Shield West ecozone showing the most potential. Even the conservative estimate is over five times Canada’s annual greenhouse gas emissions, a substantial contribution to its 2050 net-zero goal. Further research is needed to refine these estimates, assess economic viability, and investigate impacts on permafrost and albedo.

Afforestation at Canada’s northern boreal forest edge has the potential to remove substantial amounts of atmospheric carbon dioxide and store it for the next 75 years, suggest Monte Carlo simulations with an open-source carbon budget model.

Introduction

Climate change is occurring at an accelerated pace in higher latitudes, with the Arctic warming about four times faster than the global average^1,2. This rapid warming poses challenges for northern ecosystems, but also creates opportunities for nature-based climate solutions (NbCS). Canada’s federal government aims to reduce national greenhouse gas (GHG) emissions by 40% below 2005 levels by 2030 and to achieve net zero emissions by 2050³. Realizing these targets will require both industrial decarbonization and substantial contributions from NbCS to remove and store carbon. Canada’s boreal forests harbor vast carbon stocks, with managed boreal forests by themselves storing nearly 28 gigatonnes (Gt) of carbon, while unmanaged lands hold even larger quantities^4,5.

However, the boreal forests are shifting northward, albeit slowly, driven by rising temperatures and retreating sea ice^6,7. Exactly how this shift will alter net carbon stocks remains largely uncertain because of the interplay between multiple ecosystem factors such as vegetation and albedo feedbacks, permafrost, and changing temperatures^6,8. Such a northward shift of the boreal has been observed by analysing inter-annual trends in annual maximum vegetation greenness using satellite observations, providing early indicators⁷; however, further analysis shows that this expansion of tree cover into the southern arctic is slow and is not enough to compensate for the decline of tree cover in the southern boreal boundary due to wildfires and timber logging, indicating signs of biome contraction⁹. This duality, in which the boreal forest is both expanding northward and contracting in parts of its southern range, raises the question of whether targeted reforestation/afforestation efforts at the boreal–taiga interface could help stabilize and potentially expand this critical carbon sink.

Despite growing interest in boreal afforestation^10–12, major uncertainties persist. Boreal tree planting at northern latitudes ranks as a high impact, high uncertainty NbCS due to variables such as permafrost thaw, albedo feedbacks, site limitations, high wildfire frequency, and unpredictability in long-term carbon durability^10,13. It is crucial to address these uncertainties to better understand the role of northern boreal tree planting in meeting Canada’s climate change mitigation targets (see section Discussion), however, an even more basic uncertainty from the perspective of climate change mitigation is the actual carbon removal capacity of reforestation/afforestation projects considering existing landscape details, changing climatic conditions, variations in possible scenarios, and spatially explicit modeling at appropriate ecozone scales, making it difficult to pinpoint which subregions and scenarios will provide the greatest removal capacity.

Historically, assessing this potential has been limited by methodological constraints. Many early carbon budget models were aspatial, relying on aggregated data that could not fully capture the fine-scale heterogeneity of ecosystems or the precise impact of localized disturbances^14–19. Recent progress in the field has shifted towards spatially explicit frameworks capable of integrating high-resolution remote sensing data to model complex carbon dynamics across landscapes. The Generic Carbon Budget Model (GCBM)¹¹ represents a key advancement in this area. As an open-source, spatially explicit framework, the GCBM simulates all five IPCC-required carbon pools and is designed to process the large, geographic datasets essential for modern land-sector monitoring¹¹. Its scientific foundation is rooted in the well-established CBM-CFS3 model¹¹, which is used for Canada’s national forest carbon reporting (see section Methods for more details).

The GCBM’s ability to integrate satellite-based forest inventories and model disturbances at a granular level makes it uniquely suited to reduce the uncertainties in northern carbon removal estimates. For example, the model’s granularity allows for finer properties, such as specific land histories and probabilistic fire regimes, to be assigned to each grid cell. This detailed, cell-level approach provides the foundation for Monte Carlo simulations that generate more robust statistical estimates of carbon removal under various scenarios. To our knowledge, a spatially explicit model with this capability has not been previously applied to assess the carbon budget impacts of taiga reforestation and afforestation at the western boreal edge. Therefore, this study presents a novel application of the GCBM to provide a more robust, spatially detailed assessment of carbon dynamics in this critical, understudied region. Although national-scale analyses^14,15 and few province-scale analyses^16–19 have evaluated mitigation potentials of boreal tree planting, few focus specifically on the boreal boundary, include factors such as region-specific fire regimes, conduct meaningful ablations, or consider spatially explicit inventory. We address this by considering the north-western boreal boundary belonging to the Taiga Plains (TP) and the Taiga Shield West (TSW) ecozones as our region of interest (see Fig. 1a), and use the National Terrestrial Ecosystem Monitoring System (NTEMS) Satellite-Based Forest Inventory (SBFI)²⁰. We use GCBM to conduct spatially explicit carbon budget modeling in a range of scenarios and provide monte carlo estimates of carbon removal capacity in these scenarios. For each scenario, which considers different land histories, planting mortality rates, and climatic variables, we run a suite of simulations to arrive at these estimates. We project forest growth using baseline and regional yield curves and incorporate probabilistic fire regimes to simulate natural disturbances. Our results provide insight into the role played by different relevant parameters in the eventual removal capacity, the outcomes resulting from the spread of scenarios, and how these can potentially be used to guide policy decisions²¹.

Fig. 1. Region of interest and TEC at 2100 as a function of mean fire return interval (FRI).

a Region of interest chosen from the NTEMS-SBFI. The strip was chosen from the NTEMS-SBFI²⁰ by considering the boreal forest edge in the north-west, spanning the TP and TSW ecozones, and the provinces of Northwest Territories (NT), Manitoba (MB), and Saskatchewan (SK). Yukon (YT), Nunavut (NU), and Taiga Taiga Cordillera (TC) were chosen, but were omitted for analysis because of issues with GCBM (see Methods section). b TEC for afforestation on FL and NFL, and baseline and afforested configurations. c TEC for afforestation on FL with and without generic mortality. 90% Mort refers to 90% mortality of afforested trees after 5 years. d TEC for afforestation on NFL with and without generic mortality. FL forested land, NFL non-forested land. The baseline refers to existing forests without additional afforestation in free areas. The plotted values are averaged across a window of 100. The lines denote the mean and the spread shows the standard deviation computed across the 1000 independent simulation runs per scenario.

Results

In this study, we consider reforestation/afforestation in the free areas (see Methods) of the region of interest at the edge of the north-western boreal in 2025 and track changes in total ecosystem carbon (TEC) until 2100 (75 years). Hereafter, we refer collectively to reforestation and afforestation as afforestation, noting that historically forested lands undergoing afforestation represent reforestation. Subsequent sections examine carbon removal potential and investigate various experimental configurations to reveal interactions within the afforestation system. Unless stated otherwise, simulation parameters follow those detailed in the Methods section.

The fire return interval exerts strong control on TEC

Fire is a critical disturbance factor in boreal forests, shaping forest age, composition, and diversity^22,23. The mean fire return interval (FRI), defined as the average years between consecutive fires at a given location, is the primary indicator used to characterize fire regimes^24–26. Typically, mean FRI is derived by fitting a negative exponential distribution to the proportion of surviving forest^24,27–29. In contrast, this study samples FRI values directly from ecozone-specific Weibull distributions (Eq. (1) in Methods), parameterized using historical data^30,31, which effectively capture observed FRIs across boreal ecozones^32–35. For example, Landsat-based estimates suggest FRIs ranging from 150 to 500 years at the TP edge and 500 to 1500 years at the TSW edge of the northwestern boreal range³². We set our Weibull scale and shape parameters accordingly (see Methods and Supplementary Fig. 1).

Total ecosystem carbon (TEC), the sum of carbon stored in aboveground biomass, belowground biomass, and dead organic matter, is reported as tonnes per hectare (see Supplementary Fig. 2 for area distribution), considering only carbon accumulated during the experiment relative to initial conditions. Our results show that TEC at 2100 increases with longer FRIs but saturates near an FRI of 500 years (Fig. 1b–d), consistent with our fire regime simulations (see Methods). Additionally, TEC at 2100 is considerably higher in historically forested land (FL) compared to historically non-forested land (NFL) following afforestation (Fig. 1b). One explanation could be that previously forested areas inherently support forest growth better, whereas non-forested lands may require substantial interventions (e.g., soil amendments, nutrients) to achieve comparable carbon storage capacity (see Supplementary section Ablations with Land Classes, Soil Types, and Fire Regimes). Afforestation in available areas within already forested land also substantially boosts TEC compared to a baseline scenario of allowing existing forests to grow without additional planting (Fig. 1b). These patterns remain consistent across the explored FRIs (Supplementary Figs. 3 and 4).

We further investigated the effects of poor site quality, limited afforestation success, and failed seeding strategies by imposing a generic 90% tree mortality rate. This simulates real-world scenarios, as >95% of Canadian afforestation occurs via planting rather than seeding due to high seeding failure rates³⁶. Under these extreme mortality conditions, TEC at 2100 is only marginally better than baseline in historically forested land (Fig. 1c). Conversely, in historically non-forested areas, introducing 90% mortality has minimal impact on TEC because survival rates are already low, making additional mortality less important (Fig. 1d). This relationship holds across the FRI ranges explored. To enhance robustness, future analyses could incorporate explicit fire-spread modeling (e.g., Burn-P3³⁷), accounting for ignition sources and wind or terrain effects. Additionally, integrating climate projections to examine how shifts in temperature and precipitation might alter FRI distributions could provide deeper insights into future TEC trajectories³⁸.

The effect of generic mortality on ecosystem carbon growth

Examining TEC growth trajectories over time helps us understand how disturbances and historical land classes influence forest carbon accumulation up to 2100^39,40. Initially, afforestation on NFL leads to reduced TEC, improving only later in the century (Fig. 2a). This early decrease likely arises from poor growing conditions limiting biomass accumulation relative to disturbance-related losses⁴¹. Interestingly, applying high generic mortality (90%) results in similar TEC states by 2100 regardless of whether afforestation occurs on FL or NFL (Fig. 2a), suggesting that at extreme mortality rates, the original land class becomes less influential⁴². Moreover, timing of mortality events (2030 vs. 2060) has little effect on TEC outcomes at equivalent mortality rates (Fig. 2a). In contrast, the FL baseline scenario exhibits a plateau, whereas additional afforestation notably enhances TEC accumulation by 2100 (Fig. 2a).

Fig. 2. TEC over the years up to 2100.

a FL forested land, NFL non-forested land, X% Mort - X% mortality of afforested trees after 5 years. The baseline refers to existing forests without additional afforestation in free areas. The plotted values are averaged across all other variables. The lines denote the means. The standard deviations can be found in Supplementary Fig. 5. b TEC with different percentages of mortality. Mortality acts as a control on the slope of TEC growth over the years. Standard deviation spread not shown to retain clarity. c TEC with different mean fire return intervals (FRI) for afforestation on forested land (FL). The FRI affects the slope of TEC growth over the years to a lesser extent compared to mortality. Standard deviation spread not shown to retain clarity.

To better understand mortality’s role in TEC growth, we varied generic mortality rates from 10% to 90%, observing a mostly linear effect on TEC growth, with minor deviations between 40% and 50% mortality possibly due to GCBM modeling artifacts (Fig. 2b). Thus, mortality regulates TEC growth slopes and can serve as an adjustable modeling parameter. Fire regimes, particularly fire return interval (FRI), also control TEC growth trajectories but have a smaller impact (Fig. 2c; see section The Fire Return Interval Exerts Strong Control on TEC). Integrating species-specific mortality rates, distribution of mortality events spread out over the first few decades, and tying mortality to soil productivity or microclimate data, could be interesting extensions of this mortality analysis.

Administrative-ecozone combinations and climate sensitivity

Our region of interest includes four key administrative-ecozone combinations: Northwest Territories (NT)-Taiga Shield West (TSW), NT-Taiga Plains (TP), Manitoba (MB)-TSW, and Saskatchewan (SK)-TSW. Separating the sequestration potential according to these combinations gives us an idea of how conducive each of these are for afforestation^4,43. We observe that NT-TSW has the highest sequestration potential as given by TEC at 2100 both for FL and NFL historical land classes (Fig. 3a). While SK-TSW has the second highest TEC in FL, its potential in NFL is marginally better than MB-TSW, which is the lowest in NFL (Fig. 3a). Climate change, altering mean annual temperature (MAT) and total annual precipitation (PCP), impacts carbon sequestration. Direct MAT input to GCBM minimally changes TEC (see Supplementary Figs. 6–8), indicating yield curves primarily account for climate effects in GCBM. Adjusting MAT and PCP in yield equations reveals increased sequestration only with rising temperatures (Fig. 3b). An exception occurs in NT-TSW with reduced MAT and PCP (−20%) slightly outperforming averages (Fig. 3b). Generally, lowest MAT (−40%) and highest PCP (+40%) scenarios reduce TEC the most (Fig. 3b). Province-specific yield curves^44–47 and climate-driven disturbance regimes³⁸ could further refine these estimates.

Fig. 3. TEC at 2100 in different admin-eco combinations and with changes in MAT and PCP.

a FL forested land, NFL non-forested land. The baseline refers to existing forests without additional afforestation in free areas. The plotted bars are averaged across all other variables. The bars denote the means and spread denotes the standard deviations computed across the 1000 independent simulation runs per scenario. b TEC with different percentage changes in mean annual temperature (MAT) and total annual precipitation (PCP). The TEC accounts for existing plus afforested trees in FL. c Yield volume as a function of age up to 100 years for species group 3 in the Taiga Plains eco-zone. d TEC over the years up to 2100. The TEC accounts for existing plus afforested trees on forested land (FL). e TEC at 2100 as a function of mean fire return interval (FRI). The plotted values are averaged across a window of 100. f TEC at 2100 as a function of the afforestation area. A combination of existing simulations is sampled and combined to get higher afforestation area and the resulting TEC is added. Legend shows varying combination of increase or decrease in mean annual temperature (MAT) and total annual precipitation (PCP). Standard deviation spread not shown to retain clarity. The TEC accounts for existing plus afforested trees on forested land (FL).

Although the yield curves we use are parameterized by environmental variables, unbalanced sampling renders them a non-ideal candidate for faithfully modeling the impacts of climate change⁴⁸. This is mainly due to the limited sampling in climate extremes and areas where climate is a limitation for tree growth, as well as the inability to handle non-linear effects⁴⁸. However, climate sensitivity is a major limitation in most of existing Canadian growth and yield models, and obtaining reliable yield curves as a function of climate parameters is an ongoing effort^49–51. Therefore, in this section, keeping in mind the limitations of our yield curves, we use them to stress test how the yield might behave with changing environmental conditions. Though this is not perfectly reliable, it allows us to speculate how the climate parameters might regulate growth in the future.

Altered yield curves (Eq. (2) in section Methods) for species group 3 in the TP eco-zone (for other species groups in the TP eco-zone see Supplementary Fig. 9), illustrate that MAT strongly affects yield (higher MAT increases yield, lower MAT decreases yield), while PCP impacts yield modestly (Fig. 3c). TEC trends closely follow these yield patterns, except for some specific MAT-PCP combinations (Fig. 3d). Similar relationships appear between TEC, fire regimes, and afforestation area, emphasizing MAT’s dominant role (Fig. 3e, f). Future work should focus on refining climate-sensitive yield curves and modeling climate-influenced disturbances. Both the fire regime and the afforestation area scaling act as separate levers, acting in conjunction with the climate state (Fig. 3e, f), however, the changing climate will certainly alter the fire regime and therefore restrict the space of possible scenarios. Altering the disturbance regimes with the changing climate and using proper climate-sensitive yield curves are natural extensions for future work.

How does carbon removal capacity scale with afforestation area

We run simulations on patches within gridded cells of ~7–11 hectares (Supplementary Fig. 2) to maintain computational tractability. To explore how total ecosystem carbon (TEC) changes with larger afforestation areas, we combine afforestation areas and TEC from multiple independent experiments. Historical land class and mortality influence TEC growth slopes more relative to afforestation area (Fig. 4a). Both FL and NFL with 90% mortality exhibit growth patterns similar to the FL-baseline scenario, whereas NFL without mortality shows marginally higher growth (Fig. 4a). FL-afforested areas (part of FL-baseline+afforested) achieve the highest TEC growth rates (Fig. 4a). In contrast, FRI has a lesser impact on TEC growth slopes on FL, except for the smallest FRI (FRI = 30), which performs notably worse (Fig. 4b). Other FRIs cluster closely, indicating similar growth patterns (Fig. 4b). However, FRI dominantly controls TEC growth on NFL (Supplementary Fig. 10).

Fig. 4. TEC at 2100 as a function of the afforestation area and over the years comparing yield curves by Timberworks Inc. and Ung et al.

a FL forested land, NFL - non-forested land, 90% Mort − 90% mortality of afforested trees after 5 years. The baseline refers to existing forests without additional afforestation in free areas. The lines denote the mean and the spread shows the standard deviation. b TEC with different mean fire return intervals (FRI) for afforestation on forested land (FL). The TEC accounts for existing plus afforested trees in FL. A combination of existing simulations is sampled and combined to get higher afforestation area and the resulting TEC is added. Standard deviation spread not shown to retain clarity. c TEC with high existing forest density. d TEC with low existing forest density. The TEC accounts for existing plus afforested trees on FL. The yield curves from Ung et al. produce a lower bound for the TEC.

TEC increases roughly linearly with afforestation area when locations are sufficiently independent, suggesting greater TEC benefits from afforesting multiple smaller, independent patches rather than fewer correlated ones. Although our experiments combine patches randomly, deliberate selection minimizing parameter correlations may further enhance TEC growth slopes. Nonetheless, summing TEC from independently treated patches overlooks correlated risks, such as widespread wildfires²³, which could be addressed by explicitly modeling adjacency and connectivity effects^52,53. Additionally, our experiment does not capture non-linear, large-scale afforestation impacts on processes like competition, nutrient cycling, and hydrology⁵⁴, limitations stemming from GCBM constraints and patch combination methods. Resolving this is an interesting direction for future research.

Practical yield curves and a potential lower bound

In the previous sections, we used yield curves from Ung et al.⁴⁸ (Eq. (2) in Methods) to simulate forest growth. While these yield curves offer a useful starting point, being derived from sample plots across Canada, they may not fully represent our region due to local differences in site quality and climate. Additionally, these curves average across varying densities, limiting explicit analysis of forest density effects^48,55–61. To capture a realistic range of carbon sequestration estimates, we therefore consider alternative yield curves and their variability. We obtain yield curves from Timberworks Inc. in NT⁶² and compare them against Ung et al.⁴⁸. Timberworks Inc. used Alberta’s Growth and Yield Projection System (GYPSY) to derive curves categorized by site productivity (good, medium, poor) and canopy density (dense or open)⁶². Supplementary Fig. 12 illustrates yield variations for NT across combinations of species, density, and site quality (Timberworks). Supplementary Figs. 9 and 13 provide corresponding data for TP and TSW using Ung et al.’s yield curves.

To explicitly analyze density effects, we simulated scenarios with high (Fig. 4c) and low (Fig. 4d) existing forest densities. We plot the TEC from the Ung et al. yield curve on both these sets of scenarios, and observe that it is a lower bound of the estimates (Fig. 4c, d). Results indicate that site quality had a stronger influence on TEC than afforestation density, though higher initial density increased overall TEC, likely due to faster carbon accumulation in mature, denser forests (Fig. 4c, d). See Supplementary section Modulating Forest Density Using Mortality as a Surrogate for additional analyses on forest density.

Carbon removal capacity in the Taiga

We converted TEC into CO₂e removed (Tonnes), summarizing mean and standard deviation (computed across the 1000 independent simulation runs per scenario) for 2050, 2075, and 2100 (Table 1), assuming afforestation begins in 2025. Our analysis included afforestation on historically forested (FL) and non-forested lands (NFL), totaling approximately 6.4 M hectares, alongside existing forests covering ~8.6 M hectares. We see that the Ung et al. lower bound is much less compared to the average and best case carbon removal capacity obtained from yield curves from Timberworks (Table 1). By 2100, in the average case, afforesting 6.4 M hectares remove ~3.88 ± 0.98 Gt CO₂e, while existing forests remove 3.15 ± 0.92 Gt CO₂e. (Table 1). These estimates incorporate spatially explicit inventory data and fire regimes (see Methods), excluding additional mortality and climate change factors. Incorporating these factors requires using appropriate multiplicative adjustments. For instance, a 30% mortality within five years post-afforestation reduces carbon removal by approximately 20% (Fig. 2b), whereas a warmer and drier climate scenario (+40% MAT, −40% PCP) increases carbon removal by around 35% (Table 2). The spatial distribution of this potential is illustrated in Fig. 5 for Northwest Territories, the biggest region by area in our study.

Table 1.

Carbon removed via additional afforestation and existing forests

Gt CO2e removed		Ung et al. Lower bound	Average	Best-case
Mean - afforestation	2050	0.14	0.47	0.76
	2075	0.71	2.32	3.72
	2100	1.19	3.88	6.2
Standard deviation - afforestation	2050	0.45	0.48	0.46
	2075	0.75	0.74	0.76
	2100	0.99	0.98	1.01
Mean - existing forests	2050	0.70	2.28	3.66
	2075	0.89	2.9	4.65
	2100	0.97	3.15	0.505
Standard deviation - existing forests	2050	0.65	0.82	0.91
	2075	0.71	0.86	0.95
	2100	0.75	0.92	0.99

Carbon removed in 2050, 2075, and 2100 with the total area afforested being ~6.4 M hectares. Mean and standard deviations shown for the Ung et al. lower bound, the average, and the best case scenarios. The average and best carbon removal capacity is calculated by using the different yield curves in NT for varying density and site quality by Timberworks Inc.⁶².

Table 2.

Carbon removed in 2100 under ten illustrative scenarios

Scenario	Key parameters	Approx. multiplier	Resulting Gt CO₂e (~15% of available area afforested)	Resulting Gt CO₂e (~75% of available area afforested)	Rationale/multiplier math
Baseline	- 6.4 M ha (3.2 M ha FL + 3.2 M ha NFL) - Weibull-based FRIs (TP & TSW) - No Mortality - Current Climate	1.00	3.88	19.4	Starting point, combining both FL and NFL, modest disturbance regimes, and minimal seedling loss.
Moderate mortality	- Same as Baseline except 50% Mortality (seed or establishment failures)	0.60	2.33	11.65	A ~40% reduction from Baseline: 3.88 × 0.60 ≈ 2.33 Gt.
Extreme mortality	- Same as Baseline except 90% Mortality (severe early failures)	0.25	0.97	4.85	Only 25% of Baseline remains: 3.88 × 0.25 ≈ 0.97 Gt.
FL only	- Afforest only 3.2 M ha of historically forested land	0.64	2.47	12.35	FL yields ~1.8x what NFL yields. NFL area = 3.2 M ha → total NFL carbon = 3.2x. FL area = 3.2 M ha → total FL carbon = 3.2 × 1.8x = 5.76x. Combined carbon = 3.2x + 5.76x = 8.96x = 3.88 Gt → x ≈ 0.43 Gt. So, NFL portion ≈ 3.2x = 1.37 Gt, FL portion ≈ 5.76x = 2.47 Gt. 2.47/3.88 ≈ 0.64.
NFL only	- Afforest only 3.2 M ha of historically non-forested land	0.36	1.37	6.85	NFL portion alone is 1.37 Gt; 1.37/3.88 ≈ 0.36.
Short FRI	- Same as Baseline except use a short FRI (≈30 years)	0.60	2.33	11.65	A ~ 40% reduction from Baseline: 3.88 × 0.60 ≈ 2.33 Gt.
Long FRI	- Same as Baseline except use a long FRI (≈500 years)	1.20	4.66	23.3	Reduced disturbance extends TEC accumulation by ~20%: 3.88 × 1.20 ≈ 4.66 Gt.
Warmer climate	- Same as Baseline except ~+40% MAT (simulated in yield curves)	1.35	5.24	26.2	~35% greater growth: 3.88 × 1.35 ≈ 5.24 Gt.
Cooler climate	- Same as Baseline except ~−20% MAT (in yield curves)	0.80	3.10	15.5	~20% decrease in growth: 3.88 × 0.80 ≈ 3.10 Gt.
Long FRI + warmer	- FRI ~500 years (×1.20) - +40% MAT (×1.35) - No Mortality - Full 6.4 M ha	1.62	6.30	31.5	Multiply both boosts: 3.88 × 1.20 × 1.35 ≈ 6.30 Gt.

Projected CO₂ removal by 2100 under ten scenarios varying fire regime (FRI), mortality, climate (MAT), and land class (FL vs. NFL). Each row lists an approximate multiplier vs. Baseline and the resulting removals for ~15% and ~75% of available area afforested. Baseline assumes 6.4 Mha (3.2 FL + 3.2 NFL), Weibull FRIs, no mortality, and current climate; notes show the rationale/multiplier math.

Fig. 5. Spatial map of average carbon removal capacity at 2100 in Tonnes of CO₂e with additional afforestation.

Shown for Northwest Territories, the biggest region by area in our study. The afforested regions include both FL and NFL. The average carbon removal capacity is calculated by using the different yield curves in NT for varying density and site quality by Timberworks Inc.⁶².

Our estimate of 6.4 M hectares available for afforestation in northern regions is quite conservative. We assume that only ~15% of the total calculated free area in the study region (41.12 M hectares, see Table 3 for a breakdown according to regions) is available, due to factors like logistical problems, biodiversity considerations, land ownership issues, and site amenability for afforestation. Moreover, we exclude territories like Yukon and Nunavut, the Taiga Cordillera ecozone, and the region immediately to the north of the Taiga. Given considerable potential carbon removal in forest gaps and NFL regions of TP and TSW, evaluating the feasibility of taiga afforestation involves four key considerations: 1) future scenarios involving fire, mortality, disturbances, and climate changes that impact carbon sequestration⁶³; 2) economic viability relative to alternative carbon removal strategies^64–66; 3) ecological impacts including permafrost thaw, albedo feedbacks, energy fluxes, and ecosystem resilience¹³; and 4) Role of adaptive forest management strategies, including partial cutting, salvage logging, resilient species selection, and strategic fire management in maximizing long-term carbon storage. We plan to address these questions in detail in future research, however, Table 2 shows an example of what the answer to the first question might look like.

Table 3.

Study region area breakdown

Admin-zones	Eco-zones	Existing forest area (M hectares)	Afforestation area: FL (M hectares)	Afforestation area: NFL (M hectares)
Manitoba	Taiga Shield West	0.71	1.65	4.9
Northwest Territories	Taiga Plains	2.9	7.49	6.38
Northwest Territories	Taiga Shield West	4.86	11.36	8.96
Saskatchewan	Taiga Shield West	0.16	0.29	0.09
	Total Area	8.63	20.79	20.33

A breakdown of total areas considered in the study region according to different admin-eco combinations. We get a total of 41.12 M hectares of free area available for afforestation and 8.63 M hectares of existing forests.

We tabulate an illustrative set of ten scenarios (Table 2) to show how different parameters, fire regime (FRI), mortality rates, climate, and historical land class (forested [FL] vs. non-forested [NFL]), might affect total carbon (CO₂e) by 2100 across the 6.4 M ha considered (~15% of calculated free area, 3.2 M - FL, 3.2 M - NFL). We also compute best case estimates for when ~75% of the calculated free area may be available for afforestation. The baseline (Scenario 1) corresponds to the ~3.88 Gt CO₂e figure derived from simulations using the Weibull-based FRI distributions for the Taiga Plains (TP) and Taiga Shield West (TSW), no early mortality, and no major climate departure. All other scenarios adjust key drivers and use approximate multipliers grounded in the relationships discussed in the Results section. Some key takeaways are: a) FL sequesters 1.8× more carbon per hectare than NFL in the baseline, so planting only FL or NFL yields about 1.38 Gt or 2.50 Gt, respectively, b) Moderate mortality (50%) reduces carbon by ~40%; extreme mortality (90%) cuts it to ~25% of baseline, c) Going from moderate, Weibull-based intervals to very short (~30 years) or very long (~500 years) can halve or boost carbon by 20%, d) + 40% warming can increase overall yields by ~35%, whereas cooler scenarios (−20% MAT) cut carbon ~20%. However, the interaction between temperature and fire regimes and other disturbances are not taken into account here, as this is purely a yield curve based estimate, e) Multiplying individual factors can push net CO₂e anywhere from <1 Gt in worst cases (frequent fires, extreme mortality) to over 6 Gt under ideal, low-disturbance, high-temperature scenarios.

We validated our GCBM projections against plot-level data from Canada’s National Forest Inventory (NFI)⁶⁷. The validation revealed a strong model performance in estimating carbon stocks and acceptable performance for predicting annual growth increments. As detailed in the Validation Methodology section in the Methods section, performance was assessed by comparing the mean NFI-derived values for each Age-Density-Species stratum against the corresponding mean of the outputs from GCBM. The overall R² and relative RMSE values across all strata are summarized in Table 4. The results align with expectations for forest carbon modelling. Our model demonstrated a good fit for Live Carbon Stocks (R² = 0.65), explaining 65% of the variance observed in the NFI data with a relatively low error (24%). Performance for TEC was lower (R² = 0.57), which is expected given the inclusion of the highly variable soil and deadwood pools. We consider the model’s ability to predict the Live Carbon Increment (R² = 0.43) acceptable, as capturing the noisy, net result of annual growth and mortality is a major challenge. Moreover, an R² value of 0.43 suggests that our modeling framework with GCBM correctly captures a good portion of the signal in this noisy process, successfully simulating the general trend and magnitude of forest growth dynamics in the different strata. Overall, these values fall within the ranges reported by other boreal forest carbon studies^68–70.

Table 4.

Overall model performance metrics for carbon stocks and increments

Carbon pool	R² (coefficient of determination)	RMSE
Live Carbon Stocks (AGB + BGB)	0.65	24%
Total Ecosystem Carbon (TEC) Stocks	0.57	33%
Live Carbon Increments	0.43	39%

Performance of GCBM against NFI ground-plot means for carbon stocks and increments across all strata. Metrics reported are R² and RMSE (%). Accuracy is moderate for live stocks (R² = 0.65 R, RMSE = 24%), and weaker for TEC (0.57, 33%) and increments (0.43, 39%).

Discussion

Our simulations demonstrate that fire return interval (FRI), historical land cover, climate parameters, and planting mortality collectively determine carbon removal from afforestation at the boreal–taiga interface. Longer FRIs greatly enhance total ecosystem carbon (TEC), as biomass accumulation has more time before fire resets succession. Conversely, short FRIs (<50 years), especially on historically non-forested land, limit carbon storage potential, underscoring the importance of effective fire management. Reforestation consistently surpasses afforestation on historically non-forested land due to better soil and microclimate conditions, although targeted soil amendments and species selection⁷¹ may improve outcomes. High seedling mortality (around 90%) significantly undermines carbon gains, necessitating detailed site assessments, robust silviculture, and ongoing monitoring.

Yield curves showed sensitivity primarily to mean annual temperature (MAT), with precipitation (PCP) playing a secondary role. Moderate warming generally increases carbon removal, but extreme temperatures risk intensified fires or species stress, emphasizing the need for site-specific climate projections and calibrated yield models. Carbon accumulation scales linearly with afforestation area only if stands are relatively independent, suggesting a distributed approach to afforestation to mitigate risks from large disturbances. Incorporating density-specific yield curves, like those from GYPSY, TIPSY, and VDYP, might improve accuracy, addressing potential underestimation by default models⁷². For a deeper exploration of how historical land classes, soil types, and alternative fire-regime assignments influence total ecosystem carbon, see Supplementary section Ablations with Land Classes, Soil Types, and Fire Regimes.

We don’t see organic soils play a dominant role in the big picture of TEC in our analyses. While GCBM includes soil organic carbon as a component of TEC, we don’t see noticeable differences in TEC between different soil types. A potential reason could be that the yield curves used to model tree growth do not directly account for interactions with soil type. Better integrating such complex interactions between growth models, soil characteristics, and climate is an active area of research for improving forest carbon modeling with GCBM. Another reason could be that the powerful and frequent effect of disturbances like wildfire, which resets ecosystem succession and releases large amounts of carbon, could overwhelm or mask the subtler, long-term influence of soil type on carbon accumulation. Finally, the simulations track carbon changes over a 75-year period, from 2025 to 2100. In slow-growing boreal ecosystems, the full impact of soil quality on biomass accumulation, and the subsequent transfer of carbon to dead organic matter and soil pools, may take longer than this timeframe to become statistically significant, especially when compared to the more immediate impacts of fire or planting mortality.

Although this study’s quantitative findings are specific to the boreal taiga, its primary conclusions can be mapped onto the unique conditions of high-altitude regions^73,74. The critical influence of disturbance regimes like fire return intervals on carbon storage suggests that high-altitude disturbances, such as avalanches or pest outbreaks, will similarly govern sequestration success. Our finding that reforestation on historically forested land is far more effective than on non-forested sites implies that prioritizing high-altitude areas that recently lost tree cover would be more viable than attempting to plant in stable alpine meadows. The severe impact of high seedling mortality on carbon gains, the complex sensitivity of growth to climate variables, the risk-mitigation benefits of spatially diversifying plantations, and the necessity of weighing ecological trade-offs like altered albedo and hydrology are all directly transferable concepts that can inform more resilient and effective carbon sequestration strategies in high-altitude environments^73,74.

The importance of this research extends beyond quantifying carbon sinks, positioning boreal afforestation within the broader portfolio of global climate solutions where it must be carefully weighed. As a Nature-based Climate Solution (NbCS), afforestation is a vital complement to, but not a substitute for, the primary goal of industrial decarbonization and reducing greenhouse gas emissions. When considered alongside technological options like direct air capture or other NbCS such as protecting existing forests, its implementation requires a clear-eyed assessment of its inherent challenges. These include higher financial and logistical costs in remote regions, the high uncertainty and risk of carbon reversal from disturbances like wildfire, and complex ecological trade-offs involving albedo feedbacks and regional hydrology. Therefore, studies like this one are crucial not to present afforestation as a standalone panacea, but to reduce its uncertainties, enabling policymakers to strategically deploy it where it is most cost-effective, resilient, and ecologically responsible as part of an integrated climate mitigation strategy.

Our findings provide policymakers with a scenario-based view of the potential carbon gains, and inherent uncertainties, associated with boreal afforestation. In particular, the results show the importance of aligning afforestation projects with strategic fire management (e.g., ensuring longer fire return intervals), selecting historically forested sites first (where growth is faster), and tailoring interventions to local climate projections. Quantifying how temperature and precipitation shifts, fire regimes, and early planting mortality impact overall carbon uptake enables governments and stakeholders to weigh costs and benefits more effectively against alternative climate mitigation strategies. Moreover, modeling different scenarios illuminates the range of possible outcomes, helping decision-makers identify no-regrets policy pathways, such as diversifying planting locations and optimizing species selection, that reduce the risk of large-scale carbon loss from fires, pests, or climatic extremes. Ultimately, this research offers a framework that can be integrated into broader land-use and climate policy discussions, guiding investments in northern afforestation where it can most sustainably and cost-effectively contribute to national and international carbon targets.

While afforestation supports natural boreal forest expansion, ecological and logistical challenges include limited infrastructure for seedling transport, sparse availability of nursery stock adapted to harsh boreal conditions, and higher costs of planting and maintenance in remote northern regions. Projected intensified wildfires could severely limit carbon retention, compromising large-scale planting efforts. By 2100, afforestation (~6.4 M hectares, conservative) could sequester around 3.88 Gt CO₂e, contributing substantially to Canada’s climate goals, yet associated costs, wildfire uncertainty, and ecological trade-offs (albedo, permafrost) must be weighed. Forest canopies insulate permafrost¹³, but lower albedo could enhance warming, emphasizing the need for regionally specific analyses^75–78 to identify optimal sites⁷⁹. The management regime we use in our study is derived from the historical effects of management that are implicitly embedded within the empirical yield curves we use for projecting forest growth. However, adaptive forest management (e.g., partial cutting, salvage logging, resilient species selection) is crucial for mitigating leakage, baseline, and reversal risks, maximizing long-term carbon storage^80–85. Future work could integrate such explicit management scenarios into our simulation framework. Future studies could also integrate province-specific yield curves, insect disturbances, and diverse management strategies (partial cutting, strategic harvesting). Explicit soil data from NSDB⁸⁶ and better modeling of cryosolic soils at the Arctic edge, alongside refined fire regime models (e.g., Burn-P3³⁷), will improve predictive accuracy for carbon sequestration (see Supplementary section Future Research).

While this study identifies substantial afforestation potential by targeting free areas, the methodology classifies ecologically active systems such as wetlands, shrublands, and grasslands as land available for planting. These non-forested ecosystems possess their own intrinsic biodiversity and provide critical ecological functions that would be lost upon conversion to forest⁸⁷. For instance, the sedge meadows and floodplains identified as potential free areas for afforestation are crucial year-round grazing habitats for wild herds of Wood Bison. These same open wetlands provide essential breeding pools for amphibians like the Wood Frog and are nesting grounds for shorebirds such as the Lesser Yellowlegs. The open, lichen-rich peatlands and shrub barrens targeted for planting are critical summer foraging and calving grounds for Boreal Woodland Caribou. The unique mix of stunted trees and open tundra in this region is also the exclusive breeding habitat for endemic species like the Harris’s Sparrow. Afforestation would directly compromise the survival of these species specifically adapted to open landscapes. Therefore, there is a conflict between maximizing carbon sequestration and conserving existing habitat, a trade-off that would practically limit the total area suitable for afforestation.

Our spatially explicit modeling of boreal–taiga afforestation at the north-western edge of Canada’s boreal forest indicates that strategic planting could substantially increase total ecosystem carbon (TEC) by 2100. Longer fire return intervals, historically forested land cover, and low early seedling mortality all drive higher carbon gains. When scaled up, these findings emphasize that large swaths of Canada’s northern boreal, especially areas with prior forest history, could play a more prominent role in climate mitigation portfolios. However, afforestation in these regions also faces risks from the very dynamics that define boreal ecosystems, notably wildfire and harsh winter conditions. For boreal planting initiatives to succeed long-term, careful site selection, adapted silvicultural practices, monitoring, and adaptive management will be paramount.

Methods

The following sections detail the simulation softwares, data, pre-processing methods, and assumptions used for our spatially explicit carbon budget modeling.

Generic Carbon Budget Model (GCBM)

The GCBM is a flexible, open-source framework for modeling forest carbon dynamics at the stand and landscape levels¹¹. It generates a time-series output of spatially explicit and tabular indicators of forest carbon stocks and fluxes. Moreover, it adheres to the carbon estimation guidelines set by the Intergovernmental Panel on Climate Change (IPCC), simulating the dynamics of forest carbon stocks, including, aboveground biomass, belowground biomass, litter, dead wood, and soil organic carbon. The GCBM is meant to succeed the Canadian Forest Service’s (CFS) CBM-CFS3 model in Canada’s National Forest Carbon Monitoring, Accounting and Reporting System (NFCMARS)⁸⁸, which is responsible for tracking Canada’s managed forest carbon balance for international reporting purposes. The GCBM is rooted in the same scientific foundation as the CBM-CFS3, but with the added benefit of spatial capabilities and enhanced data from national ecological parameter databases. While the CBM-CFS3 is an aspatial model, the GCBM requires spatial inputs, including forest inventory data and disturbance area information, to operate effectively. Notably, the GCBM is designed to integrate seamlessly with the Full Lands Integration Tool⁸⁹ software platform, developed by moja global. Recent studies have conducted accurate parameterization, robust uncertainty assessments, and sensitivity analyses, for CBM-CFS3, and by extension for GCBM^90–92. Therefore, we don’t study the inherent parameter uncertainty in GCBM in detail here.

Region of interest

The north-western edge of the boreal forests is chosen as the region of interest from the National Terrestrial Ecosystem Monitoring System (NTEMS) Satellite-Based Forest Inventory (SBFI)²⁰ (Fig. 1a). The NTEMS-SBFI data is available as rectangular layers, each consisting of multiple polygons. The grid layers along the edge of the forests are selected to be included in the inventory. The region of interest includes the eco-zones Taiga Plains (TP), Taiga Shield West (TSW), and Taiga Cordillera (TC), and the admin-zones Northwest Territories (NT), Yukon (YT), Manitoba (MB), Saskatchewan (SK), and Nunavut (NU).

Compiling the NTEMS-SBFI data in the region of interest

The region of interest is gridded to 0.06 ×0.06 degrees (latitude, longitude) resolution and the polygons in the NTEMS-SBFI are assigned to these grid cells if they lie inside the cell or overlap with the cell. The data belonging to the polygons in each grid cell is converted to aggregate data for the grid cell. For instance, for a given cell, the tree species percentages are averaged across all the polygons in the cell and assigned to the cell. Similarly, forest age, fire fractions, percentage free area, and percentage forested area are averaged across all the polygons belonging to the cell. The historical fire year, admin-zone and eco-zone are assigned according to the most common occurrence among all the polygons in the cell. The mean average temperature (MAT) and total annual precipitation (PCP) are downloaded from ClimateDataCA⁹³, gridded to the same resolution as the grid cells (0.06 ×0.06 degrees), and assigned to them. Next, the gridded cells are retained if they belong to TP or TSW eco-zones. Currently, the TC eco-zone is not considered in the inventory because GCBM displays errors when running simulations in this eco-zone. The TC eco-zone will be considered in future experiments once these errors are resolved.

The resulting grid cells are further filtered depending on whether they have >35% free area. The free area is calculated by summing the area comprising exposed regions, bryoids, shrubs, herbs, and wetland, and excludes areas containing water, snow, rock, and forest. This is done to identify gaps along the boreal edge where afforestation can potentially be carried out. Therefore, the resulting estimates could be considered a lower bound because grid cells with <35% area are ignored and some snow covered regions which could be afforested are ignored. On the other hand, many exposed regions and regions with bryoids, shrubs, herbs, and wetland may not be amenable to afforestation. The conservative estimates balance these situations. Each grid cell in the retained inventory consists of the following relevant compiled data: a) percentages of different species on forested land, b) fraction of the region affected by fire, c) the year in which the historical fire disturbance occurred, d) admin-zone and eco-zone, e) mean forest age for forested region, f) percentage free and forested area, and d) MAT and PCP. For more details on the compiled forest inventory data in the region of interest, see section Forest Inventory of the Region of Interest in the Methods section.

Creating scenarios

For an experiment with a chosen configuration, 1000 independent simulations per scenario configuration are created. We chose n = 1000 as we see that our variable ranges are sufficiently spanned at this n and it keeps simulations computationally tractable. If the experiment is decided to be run on historically forested land (FL), then the grid cells are filtered using the condition that the mean forest age in the cell is not zero and the percentage free area is <70%. If the experiment is decided to be run on historically non-forested land (NFL), cropland (CL), or grassland (GL), then the grid cells are filtered using the condition that the mean forest age in the cell is zero and the percentage free area is ≥70%. Next, each scenario is created by doing the following:

1) A random cell is sampled from the grid, dividing the cell further into a 0.002 ×0.002 degrees grid, and sampling a sub-cell randomly. The sub-cell region is assigned as the inventory region and the historical land class for the inventory region is assigned according to the experiment type (either FL or NFL). 2) If the historical land class is FL, forest age is assigned to the inventory region by noting the mean forest age for the cell and sampling from a normal distribution with mean = mean forest age and standard deviation = 0.1. If the historical land class is NFL, forest age is assigned as 0. 3) Eco-zone of the cell and the dominant soil type mapping for the eco-zone (Table 5) are noted. A dominant soil probability (dom_soil_prob) (see Table 6) is sampled randomly and the dominant soil type is assigned according to dom_soil_prob and Average soil type is assigned according to 1-dom_soil_prob. Here it would be ideal to include the full spatially explicit soil types from Soil Landscapes of Canada (SLC v3.2)⁸⁶, but this data is not used currently. 4) Historic species are assigned to the inventory region by a weighted sampling according to species percentages in the cell. If the cell has no trees, a historical species of Nope is assigned. 5) An afforestation region is chosen based on the percentage of free area in the cell. This region is assigned starting from the top portion of the inventory region, with lat range=percentage of free area*lat range of inventory region. The exact area with trees in the cell, combining both existing and afforested trees, is decided by GCBM and spans a distribution (see Supplementary Fig. 2). 6) The species to be afforested is chosen based on historical land class. If FL, the species is chosen via weighted sampling according to species percentages in the cell. If NFL, it is chosen via weighted sampling according to species percentages representative of the eco-zone of the cell (see Table 7). 7) The admin-zone, eco-zone, and MAT of the cell are assigned to the inventory and afforestation regions. 8) The simulation start and end years are set to 2015 and 2100 respectively. The afforestation year is set to 2025. If the experiment configuration involves generic mortality, generic mortality percentage and year of disturbance are set. Generic mortality is considered a proxy for planting strategy (perhaps also site quality), where seeding risks high mortality and transplanting results in low mortality. A similar relationship can be drawn for site quality. 9) Fire events are assigned by sampling fire return intervals (FRIs) from a weibull distribution (Eq. (1)). The weibull distribution parameter scale is sampled according to eco-zone specific mapping (see Table 8) based on the eco-zone of the cell. The weibull shape parameter is randomly sampled from a range (see Table 6). The scale parameter is a measure of the time scale of a fire regime, whereas the shape parameter connects the hazard of burning with stand age, with shape > 1 denoting an increase in hazard of burning with stand age. No fire events are assigned for the first 3 years after afforestation, after which fires are assigned every year according to a binomial probability of a fire event happening in the current year given the previous fire event and difference between the years. The base probability of fire happening on any given year is set to 1/FRI.

$$f\left(x;\lambda ,k\right)=\left\{\begin{array}{cc}\hfill \frac{k}{\lambda }{\left(\frac{x}{\lambda }\right)}^{\mathrm{k}-1}{e}^{{-\left(x/\lambda \right)}^k},\hfill & \hfill x\ge 0,\hfill \\ \hfill 0,\hfill & \hfill x<0.\hfill \end{array}\right)$$ 1

Table 5.

Dominant soil type mapping according to eco-zone

Eco-zone	Soil type
Taiga Plains	Luvisolic (W. Canada), Cryosolic
Taiga Shield West	Brunisolic, Cryosolic

Mapping derived from literature and the Soil Landscapes of Canada (SLC v3.2). Taiga Plains: Luvisolic (W. Canada), Cryosolic; Taiga Shield West: Brunisolic, Cryosolic.

Table 6.

Simulation-related parameters and their assumed ranges

Parameter	Range/List
Dominant soil probability (dom_soil_prob)	[0.6, 1, 0.05]
Weibull shape	[1.05, 1.61, 0.05]32

Bracketed notation [A,B,C] denotes an inclusive list from A to B in steps of CC. Parameters shown: dominant soil probability (dom_soil_prob) and Weibull shape (range per³²).

Table 7.

Species and species percentages according to eco-zone based on NTEMS-SBFI

Eco-zone	Species	Percentage
Taiga Plains	Black spruce	81
	Trembling aspen	13
	Lodgepole pine	6
Taiga Shield West	Black spruce	73
	Lodgepole pine	27

Values are rounded; Reported species: Black spruce, Trembling aspen, and Lodgepole pine (Taiga Plains); Black spruce and Lodgepole pine (Taiga Shield West).

Table 8.

Weibull scale parameter ranges according to eco-zones

Eco-zone	Weibull scale parameter
Taiga Plains	[5,100,200]
Taiga Shield West	[5,500,700]

Bracketed notation [A,B,C] denotes an inclusive range from A to B in steps of C. Values for Taiga Plains and Taiga Shield West are derived from ref. ³².

Equation 1: Weibull distribution with shape parameter k and scale parameter λ.

10) The fire fraction is set to the historical fire fraction of the cell (see Supplementary Fig. 14). The fire regions are assigned for successive years (if more than 1 event) by starting from the bottom of the inventory region and assigning non-overlapping regions with lat range = fire fraction*lat range of inventory region. 11) The rest of the parameters in the GCBM are set to default including site quality and planting density.

Generating yield curves

$$\mathrm{Yield}=\exp \left(a_0+a_1\cdot \mathrm{MAT}+a_2\cdot \mathrm{PCP}+\frac{a_3+a_4\cdot \mathrm{MAT}+a_5\cdot \mathrm{PCP}}{T_s}\right)\cdot a_6$$ 2

Equation 2: Yield equation derived from Ung et al.⁴⁸.

We use two sets of yield curves. The first set of yield curves are generated according to the equations given in McKenney et al.⁹⁴ for different eco-zones and different species groups (groups 0–7). These equations are derived from Ung et al.⁴⁸ and are obtained by fitting growth and yield models to temporary sample plots across Canada. As these curves are parametrized using climate variables and were shown to be applicable across provinces, we employ them in our study to study overall trends and consider them as a lower bound. Our region of interest spans multiple provinces and ecoregions, and as these yield curves have been fitted to diverse climates across Canada, they are appropriate for an exploratory study such as this. The general form of the equation is as given in Eq. (2). The species are assigned to species groups as in Table 9. The coefficients for the yield equations for different species groups are as given in Table 10. The MAT and PCP values used are averages for the eco-zones. The second set of yield curves were obtained by contacting the forestry department of the government of northwest territories (GNWT). The yield curves shared by GNWT are the ones used in the forest management agreement and 25 year strategic plan document prepared by Timberworks Inc in support of their land use permit application⁶². These yield curves are meant to be used in the province of Northwest Territories, and comprise of 5 species types (Deciduous, Black Spruce, White Spruce, Mixed Wood, Pine), 2 density configurations (High, Low), and 3 site qualities (Good, Medium, Poor). In subsequent phases, province specific yield curves will be explored^44–47. Both these yield curves, other data, and project code can be found under the data folder in the GitHub Repository⁹⁵.

Table 9.

Species assignment to species groups

Species group	Species
0	Not stocked
1	ABIE.AMA, PSEU.MEN
2	CHAM.NOO, THUJ.PLI, TSUG.CAN, TSUG.HET, TSUG.MER
3	PINU.ALB, PINU.BAN, PINU.CON, PINU.PON, PINU.RES, PINU.STR
4	PICE.ABI, PICE.ENG, PICE.GLA, PICE.MAR, PICE.RUB, PICE.SIT
5	ABIE.BAL, ABIE.LAS
6	BETU.PAP, POPU.BAL, POPU.GRA, POPU.TRE
7	ACER.MAC, ACER.RUB, ACER, SAH, ALNU.INC, ALNU.RUB, BETU.ALL, FRAX.NIG, LARI.LAR, LARI.OCC, QUER.RUB, ULMU.AME

Mapping of species to species groups (0–7) used in the yield and growth analyses. Abbreviations follow the taxonomic codes in ref. ⁹⁴ (e.g., ABIE.AMA = Abies amabilis, PSEU.MEN = Pseudotsuga menziesii). Group 0 denotes unstocked stands; all other groups list representative conifer and broadleaf species.

Table 10.

Coefficients of yield equation

Species group	a0	a1	a2	a3	a4	a5	a6
0	0	0	0	0	0	0	0
1	5.7	0.0636	−0.0001	−173.859	14.5291	0	1.0423
2	5.7	0.0636	−0.0001	−173.859	14.5291	0	1.0423
3	6.4755	0.1271	−0.0008	−67.4993	2.6486	0.0119	1.0777
4	6.443	0.0981	0.0013	−37.4046	7.5551	−0.0362	1.2127
5	4.8421	0	0.0007	−74.8932	4.3921	0	1.2453
6	6.6358	0	−0.0004	−55.6634	0.8537	0	1.0289
7	6.605	0	−0.0009	−47.1154	1.6253	0	1.0819

Coefficients a₀–⁣a₆ for the species-group yield equation (Eq. (2)). Species groups 0–7 follow the classification in ref. ⁹⁴; coefficient values are from ref. ⁴⁸. Zero entries indicate omitted terms for that group; units and conventions follow the original sources.

It is important to note that GCBM does not dynamically model the full hydrology and energy fluxes of the forest ecosystem. Therefore, using MAT and PCP as inputs for yield curves serves as a surrogate for the complex water-energy nexus that truly governs tree growth and mortality. To get a more accurate estimation of these interactions, combined or coupled modeling approaches are necessary. Future research could involve integrating GCBM’s detailed carbon budget accounting with other specialized models, such as a dedicated hydrological model to better simulate soil moisture and water stress, or a land surface model to explicitly calculate energy fluxes and albedo feedbacks. This integrated approach would provide a more holistic understanding of how afforestation impacts, and is impacted by, the complete water and energy balance of the region.

Research framework and simulation workflow

To provide a clear and logical overview of our study, this section outlines the research framework used to estimate the carbon removal capacity of taiga afforestation. The workflow is divided into four main phases: (1) Data Acquisition and Pre-processing, (2) Monte Carlo Simulation Setup, (3) GCBM Carbon Budget Modeling, and (4) Analysis and Scenario Evaluation. The process is designed to integrate spatially explicit inventory data, probabilistic disturbance regimes, and climatic scenarios to generate robust estimates of carbon dynamics. The logical flow of this framework is illustrated in the flowchart given in Fig. 6.

Fig. 6. Research framework and simulation workflow.

Outlines the research framework used to estimate the carbon removal capacity of Taiga afforestation. The workflow is divided into four main phases: Data Acquisition and Pre-processing, Monte Carlo Simulation Setup, GCBM Carbon Budget Modeling, and Analysis and Scenario Evaluation.

Phase 1: Data acquisition & pre-processing

The initial phase focused on gathering and preparing the necessary spatial and tabular data. We selected the north-western edge of Canada’s boreal forest as our region of interest from the NTEMS-SBFI²⁰. This region spans the Taiga Plains (TP) and Taiga Shield West (TSW) ecozones. The raw polygon data from NTEMS-SBFI was gridded to a 0.06 ×0.06 degree resolution. For each grid cell, key attributes such as tree species percentages, forest age, and historical fire data were aggregated and averaged. Climate data, specifically mean annual temperature (MAT) and total annual precipitation (PCP), were downloaded from ClimateDataCA⁹³ and assigned to the corresponding grid cells. Finally, the grid was filtered to retain cells with greater than 35% free area available for potential afforestation, creating a baseline inventory for our simulations. See subsections Region of Interest and Compiling the NTEMS-SBFI data in the region of interest in the Methods section for more details.

Phase 2: Monte Carlo simulation setup

To account for variability and uncertainty, we employed a Monte Carlo approach, creating 1000 independent simulations for each experimental scenario configuration. Each scenario was built by randomly sampling a grid cell from the pre-processed inventory. Within each scenario, key parameters were assigned based on the specific experiment being run. This included defining the historical land class (historically forested FL or non-forested NFL), assigning forest age, and selecting species for afforestation based on ecozone-specific data. A critical component of this phase was the assignment of fire events. Fire return intervals (FRIs) were sampled from ecozone-specific Weibull distributions, which effectively capture the observed fire regimes in the boreal region. Other parameters, such as generic mortality rates and the timing of afforestation (set to 2025), were also defined at this stage. See subsection Creating scenarios in the Methods section for more details.

Phase 3: GCBM carbon budget modeling

With the scenarios fully parameterized, we used the GCBM to simulate forest carbon dynamics. The GCBM simulates changes in all five IPCC carbon pools: aboveground biomass, belowground biomass, litter, dead wood, and soil organic carbon. For each of the 1000 scenarios, we ran a simulation from the start year of 2015–2100. Growth and yield were modeled using two sets of yield curves: a conservative set derived from Ung et al. and a more optimistic set from Timberworks Inc. based on Alberta’s GYPSY model. The model’s output consisted of time-series data on the total ecosystem carbon (TEC) for each simulated patch. See subsections Generic Carbon Budget Model (GCBM) and Generating yield curves in the Methods section for more details.

Phase 4: Analysis & scenario evaluation

In the final phase, the TEC outputs from all simulations were aggregated and analyzed. We calculated the mean and standard deviation of carbon removal across various scenarios to understand the influence of key variables like FRI, mortality, historical land class, and climate parameters (MAT and PCP). The results were evaluated for different administrative-ecozone combinations (e.g., NT-TSW) to identify regions with the highest sequestration potential. Finally, we converted the TEC (tonnes of Carbon) into tonnes of CO₂ equivalent (CO₂e) and scaled the results to the total available afforestation area (~6.4 M hectares) to provide comprehensive estimates of carbon removal potential by 2050, 2075, and 2100 under different illustrative scenarios.

Forest inventory of the region of interest

The region of interest at the northwestern boreal edge exhibits distinct spatial patterns in its forest characteristics (Fig. 7). For clarity and to better illustrate the spatial nuances of the inventory data, we focus specifically on the Northwest Territories (NT). This region was chosen because it is the largest administrative zone by area in our study and our results indicate it has the highest carbon sequestration potential. Focusing on the NT provides a detailed and representative view of the key environmental gradients and forest characteristics without the visual clutter of showing all administrative regions at once. The compiled, gridded inventory data has been made available in the project’s GitHub repository to allow for replication and to serve as a resource for other researchers.

Fig. 7. Forest inventory characteristics for the Northwest Territories portion of the study area.

The data is compiled from the NTEMS-SBFI and gridded. Plotted variables include: a forest age in years, b percentage of free area available for afforestation, c percentage of existing forest cover, d mean annual temperature in degrees Celsius, and e dominant tree species. Species abbreviations are as follows: BETU.PAP (Betula papyrifera), LARI.LAR (Larix laricina), PICE.GLA (Picea glauca), PICE.MAR (Picea mariana), PINU.BAN (Pinus banksiana), and POPU.TRE (Populus tremuloides). Note: Each cell can contain a list of species according to their observed percentages. Only the most dominant species is visualized here.

The forest age distribution reveals that older stands, with ages exceeding 75 years, are predominantly located in the southern and central parts of the study area, while younger forests are scattered throughout (Fig. 7a). The percentage of free area, available for potential afforestation, is highest in the northern parts of the region, often exceeding 70%, which corresponds to the transition into the taiga and tundra ecosystems (Fig. 7b). Conversely, the percentage of forested area is densest in the southern portion, with coverage frequently greater than 30%, and gradually decreases northward (Fig. 7c). A clear latitudinal gradient is observed for mean annual temperature, with warmer temperatures (around −2.1 °C) in the south and progressively colder conditions toward the north, reaching as low as −9.9 °C (Fig. 7d). The dominant tree species across the region is Black Spruce (Picea mariana), which covers the majority of the landscape. Pockets of Jack Pine (Pinus banksiana), Tamarack Larch (Larix laricina), and White Spruce (Picea glauca) are also present, with minor occurrences of other species like Trembling Aspen (Populus tremuloides) and Paper Birch (Betula papyrifera) (Fig. 7e).

Validation methodology

To assess the credibility of the GCBM projections, we conducted validation against plot-level data from Canada’s National Forest Inventory (NFI)⁶⁷. Data processing and variable definitions followed the official NFI compilation procedures⁹⁶ and data dictionaries⁹⁷. The NFI ground plots were selected as the primary validation dataset due to their systematic national-scale design and comprehensive carbon pool measurements. This validation was designed to test the model’s performance across a range of stand structures and developmental stages found within the study area.

Primary validation dataset and processing

Our validation relies on data from NFI ground plots located within the Northwest Territories portion of the Taiga Plains and Taiga Shield West ecozones. All complete plot visits from the available measurement cycles were processed to derive carbon stocks for five key ecosystem pools.

Live biomass carbon (AGB + BGB)

The live tree carbon pool was calculated from individual tree measurements within the NFI database. Plot-level Above-Ground Biomass (AGB) (Tonnes ha⁻¹) was calculated by summing the per-stem biomass values provided in the NFI tree tables (biomass_total for large trees, smtree_biomass for small trees) and scaling the total by the plot area. Below-Ground Biomass (BGB) was not directly measured but was estimated by applying standard root-to-shoot ratios to the AGB of each plot. Based on national defaults, a ratio of 0.20 was used for conifer-dominated plots and 0.26 for broadleaf-dominated plots. Live Carbon was calculated by summing AGB and BGB and applying a carbon fraction of 0.5.

Woody debris carbon

The carbon stock in the dead wood pool was calculated directly from the nt_gp_wd_summary.csv file. The values from the four component columns (plotbio_swd, plotbio_wd, plotbio_roundwd, plotbio_oddwd), which represent biomass in tonnes per hectare, were summed and multiplied by a carbon fraction of 0.5 to yield total woody debris carbon in Tonnes Cha⁻¹.

Soil organic carbon (SOC)

The soil organic carbon pool was calculated to the full depth of measured samples by combining data from the three NFI soil sample files: nt_gp_for_flr_org_sample.csv, nt_gp_soil_org_sample.csv, and nt_gp_soil_mineral_sample.csv. The calculation did not require the horizon description file and instead relied on the depth measurements recorded with each individual sample. Layer Thickness, the thickness of each soil layer was determined directly from the sample_upper and sample_bottom columns in the sample files. To handle inconsistencies across files, a unified data-selection logic was implemented. For organic layers, carbon content was derived from the tc_8mm column. For mineral layers, the tc column was used. These were converted from gkg⁻¹ to percent by dividing by 10. For bulk density, bulk_density_total was prioritized, with bulk_density_2mm used as a fallback. To ensure physical realism, a data quality filter was applied, removing any soil sample records with a calculated bulk density outside the range of 0.1–1.5 gcm⁻³ or a carbon content outside the range of 0–60%. The carbon stock for each layer (in Tonnes Cha⁻¹) was calculated by multiplying its thickness (m), bulk density (kgm⁻³), and carbon fraction, adjusted by the appropriate unit conversion factor. These values were then summed to the plot level.

Total ecosystem carbon (TEC)

The total ecosystem carbon stock for each plot-visit was calculated as the sum of the live carbon (AGB + BGB), woody debris carbon, and soil organic carbon pools.

Stratification for model comparison

To facilitate a direct comparison with GCBM outputs, the processed NFI plot data was stratified based on three key stand attributes:

1. Stand age: Plots were grouped into 20-year age classes (e.g., 0–20, 20–40 years).

2. Density class: Following our yield curve parameterization, plots were classified as either Sparse (<900 live stems ≥9 cm DBH per hectare) or Dense (≥900 stems ha⁻¹).

3. Leading species: The dominant species for each plot was identified as the one with the highest percentage contribution in the NFI’s species composition tables.

Validation approach and performance metrics

The validation was conducted in two distinct parts to robustly assess different aspects of model performance:

1. Stock validation: The primary validation was performed on carbon stocks (Tonnes Cha⁻¹). The mean (from n NFI plots) of calculated carbon pools (Live C, TEC) was compared against the corresponding mean stock generated by the n GCBM outputs for each stratum. This approach provides an assessment of the model’s ability to represent the state of the ecosystem at different developmental stages.

2. Increment validation: The analysis of annual carbon increments (Tonnes Cha⁻¹yr⁻¹) was confined to the Live Carbon pool only. While soil depth was harmonized between visits to ensure a valid comparison, the resulting high variance in the soil and deadwood increment data confirmed that the measurement uncertainty over short NFI re-measurement intervals is too large to provide a reliable signal for validating these slow-turnover pools. The live carbon increment, however, offers a clear and direct measure of forest growth dynamics. To facilitate a direct comparison, GCBM outputs for each stratum were simulated forward for a duration matching the average re-measurement interval of the corresponding NFI plots, from which an annual increment was calculated. The mean annual increment predicted from the n GCBM runs was then compared with the mean annual increment from the n NFI measurement intervals.

Model performance for both stocks and increments was evaluated using two standard metrics. For each carbon pool, the performance metrics, Coefficient of Determination (R²) and Relative Root Mean Square Error (RMSE) were calculated across the full set of strata, comparing the list of NFI-derived means against the list of GCBM-predicted means. This provides a robust assessment of model performance over the entire range of simulated conditions.

Rationale for using Total Ecosystem Carbon (TEC) as the main indicator

In this study, Total Ecosystem Carbon (TEC) was selected as the primary indicator for quantifying the climate mitigation potential of afforestation. This choice is deliberate and serves several key purposes:

1. Comprehensive scope: TEC provides a holistic measure of carbon dynamics by accounting for all five ecosystem carbon pools mandated by the Intergovernmental Panel on Climate Change (IPCC): aboveground biomass, belowground biomass, litter, dead wood, and soil organic carbon. This approach avoids a narrow focus on only live biomass, which would overlook carbon stocks and fluxes within the dead organic matter and soil pools, pools that are particularly large and vulnerable in boreal ecosystems.

2. Policy relevance: As the scientific foundation of the GCBM is the CBM-CFS3 model used for Canada’s national forest carbon reporting, using TEC aligns our study with established national and international reporting standards. This ensures that our findings are directly comparable and can be integrated into broader assessments of how land-use activities contribute to Canada’s climate targets.

3. Disturbance impact assessment: Boreal forests are disturbance-driven ecosystems, with factors like fire playing a powerful role in carbon cycling. TEC is particularly well-suited to capture the net effect of these disturbances, as it reflects not only the loss of carbon from biomass combustion but also the subsequent transfers of carbon to dead organic matter pools and the long-term changes in soil carbon. By tracking TEC, we can more accurately model the ecosystem’s response to changing fire regimes and mortality events over the 75-year simulation period.

While other metrics, such as net primary productivity or individual pool fluxes, are valuable for detailed biogeochemical analysis, TEC serves as the most robust and policy-relevant indicator for assessing the overall impact of a large-scale afforestation initiative on the landscape-level carbon balance.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

K.B.D. conducted the research, performed simulations, wrote the article, plotted results, and created illustrations. E.O. helped with ideation and literature review. R.B., J.M.-C., and Y.L. were involved in the acquisition of funding, editing the article and overall supervision. All authors contributed to discussion and conceptualization of arguments.

Peer review

Peer review information

Communications Earth and Environment thanks the anonymous reviewers for their contribution to the peer review of this work. Primary Handling Editors:[Nandita Basu]. [A peer review file is available].

Data availability

The data that support the findings of this study are publicly available to download and are referenced in the bibliography. Refer to the Methods section for more details. The data generated from the project can be found in the GitHub Repository⁹⁵.

Code availability

The code repository for this project, including data processing and simulations can be found in our GitHub Repository⁹⁵.

Competing interests

The authors declare no competing interests

Supplementary information

The online version contains supplementary material available at 10.1038/s43247-025-02822-z.