Introduction

Forests serve as primary carbon sinks and climate regulators in global environmental systems1. Since systematic monitoring began in the 1980s, forest ecosystems have undergone significant transformation, with old-growth carbon-rich forests increasingly replaced by younger, less diverse stands with reduced carbon storage capacity2,3. This stems from deforestation, degradation, and climate change effects, including intensified wildfires and pest outbreaks4,5. Global forest loss correlates with rising temperatures6, resource overexploitation7,8, and land-use changes from agricultural expansion and urbanization9. In China, forests have been incorporated into markets as tradable forest carbon products, which are marketable assets based on carbon sequestration capacity. Related studies often fail to capture complex, non-linear interactions and potential time-lagged effects in China’s forest product exchange system.

International research reveals patterns extending beyond simple linear assumptions. Lewis et al. demonstrated that while natural forest restoration represents the most effective carbon removal strategy, the timing and magnitude of sequestration benefits vary significantly across forest types and climatic conditions10. In temperate and boreal zones, halting deforestation/harvest after 2015 yields 19 Pg C by 2100, rising to 28 Pg C with longer-lived wood products; in tropical regions, secondary-forest recovery dominates (98 Pg C when shifting cultivation fallows return to forest)11. Chazdon et al.‘s analysis reveals young secondary forests (1–20 years) accumulate carbon rapidly at 0.526 Pg C/year initially (2008–2013), declining to 0.081 Pg C/year by 2043–2048, totaling 8.48 Pg C over 40 years12. Because the widely cited 4.4 Pg C/year figure refers to gross removals partly offset by legacy emissions, area gains may not translate into proportional near-term net benefits11. These findings suggest China’s extensive afforestation programs provide an ideal context for examining whether forest area expansion translates to proportional carbon benefits.

Forest coverage (area) is commonly used as a proxy for forest carbon sequestration potential, although forest age, species composition, and management practices strongly affect sequestration rates. This makes it difficult to infer near-term net carbon uptake from area changes alone, especially when economic activity and energy use dominate emissions dynamics over short horizons. Fang et al. demonstrated that forest expansion and regrowth contributed 60% and 40%, respectively, to China’s forest carbon sink from the 1970s to 2000s13.

China has implemented ambitious forest policies, fundamentally reshaping its landscape. The Three-North Shelter Forest Program (initiated in 1978) aims to create a 4500 km protective forest belt by 205014. This was followed by the Natural Forest Conservation Program (NFCP, 1998), piloted in 12 provinces and expanded to 18 by 2000. It reduced commercial logging and provided afforestation incentives, with approximately 61 billion yuan invested from 1998 to 2005 and 96.2 billion yuan designated from 2000 to 2010 toward an afforestation goal of around 31 million hectares by 201015. The Grain for Green Program (1999) converted over 30 million hectares of marginal cropland to forests16. These programs collectively drove China’s forest coverage from approximately 8.6% (1950s) to over 23% (2020)17. Following its 2060 carbon neutrality announcement, China has integrated forests into climate mitigation through carbon offset mechanisms within emissions trading systems18,19. The 14th Five-Year Plan (2021–2025) strengthened these commitments, incorporating forest carbon sinks as key climate strategy components18. This evolution reflects China’s shift from viewing forests primarily as ecological barriers to recognizing their value as carbon sinks20. However, international studies reveal complexity, including temporal delays and variable sequestration rates21,22, helping explain why China’s remarkable forest expansion has not translated to immediate, proportional emissions reductions.

The economic value of forests as carbon products depends on dynamic interactions between forest growth, carbon sequestration, and economic development. Atmospheric CO2 concentrations affect forest sequestration capacity while being influenced by economic activities23. Crespo Cuaresma et al. found income growth effects on forest cover strongest in early development stages, weakening in advanced economies24, suggesting forest potential as carbon products varies with development stage.

Forest carbon market integration has spurred innovative financial mechanisms globally, yet implementation faces significant challenges. Payments for ecosystem services (PES) programs transact US$36–42 billion annually, but most rely on simplified area-based payments failing to account for temporal dynamics25. REDD+ mechanisms attracted over US$9.8 billion in pledges26, but only one-third of projects successfully sold carbon credits by 2018, primarily due to measurement, reporting, and verification (MRV) challenges27. China’s forest carbon offset mechanisms face similar challenges. Among twelve registered forestry projects, only one had completed verification by 201628, underscoring practical difficulties in valuation. These implementation challenges point to a fundamental methodological limitation: the absence of analytical frameworks capable of capturing the complex, time-dependent relationships among forest management, economic development, and carbon outcomes.

To overcome these limitations, we propose an integrated framework combining temporal modeling with causal inference. This enables us to disentangle predictive relationships from causal mechanisms while accounting for spatial and temporal complexities that have constrained previous forest carbon finance analyses. We hypothesize that conventional valuation approaches inadequately capture time lags between afforestation and peak sequestration, potentially causing inaccurate carbon market pricing. For predictive modeling, we employ Long Short-Term Memory-Multilayer Perceptron (LSTM-MLP) with multispatial Convergent Cross Mapping (mCCM) for causal inference. Li and Zhang reported LSTM achieved an R² value of 0.984 for daily CO₂ predictions in China, outperforming commonly used statistical or machine learning models, though they noted that excluding key exogenous variables, such as GDP and sectoral energy consumption, and focusing only on daily horizons may limit the model’s policy relevance and overlook longer-term trends29. Mussa and Khalifa demonstrated robust emission driver identification through MLP architectures (R² = 0.972), with strategic variable selection improving both forecasting performance and policy insight30. Xu et al. applied LSTM to China’s provincial forest carbon sinks with impressive results (R² = 0.979), identifying GDP growth and urbanization as key drivers31. Building upon the proven strengths of existing approaches, our LSTM-MLP architecture systematically incorporates energy consumption variables alongside forest coverage, CO₂ emissions, and economic indicators.

While LSTM-MLP enables robust prediction, it cannot establish causal directionality essential for policy design. Therefore, we employ mCCM, a nonlinear causal inference method specifically designed for systems with short time series replicated across spatial units like China’s 30 provinces. Inspired by the extensions of a classic CCM that explicitly account for causal time delays, we extend the multispatial CCM (mCCM) with time-delayed analysis, evaluating cross-map correlations as lag functions to distinguish temporally structured causal relationships from contemporaneous associations or generalized synchrony32. Leveraging this framework and data spanning 30 Chinese provinces over 2000–2019, this study has two objectives: (1) quantify the relative significance of factors influencing forest-economy-emissions relationships, emphasizing temporal dynamics inadequately captured by existing mechanisms; and (2) construct and validate a causal network, providing empirical foundation for time-differentiated forest carbon pricing models. Our study provides novel insights into the “forest carbon paradox”, explaining why existing financial mechanisms struggle with verification and permanence.

Results and discussion

Theoretical framework of forest carbon sequestration

Understanding biological dynamics of carbon uptake throughout forest life cycles is essential for interpreting relationships between forest coverage expansion, CO₂ emissions, and economic growth. Trees’ carbon sequestration capacity varies significantly throughout life cycles, following patterns mirroring growth stages. As illustrated in Fig. 1A, growth progresses through distinct phases: pioneer, young, maturing, and steady state, each corresponding to different sequestration rates (Fig. 1B).

Fig. 1: Understanding the contribution of trees in carbon storage requires evaluating both aboveground growth and whole-cycle biomass over life time.
Fig. 1: Understanding the contribution of trees in carbon storage requires evaluating both aboveground growth and whole-cycle biomass over life time.
Full size image

A Growth trajectory across four developmental stages (Pioneer, Young, Maturing, and Steady State), with the dashed red line indicating a sigmoidal growth curve. B Carbon sequestration rate over the complete life cycle (Seeds, Pioneer, Young, Maturing, Steady State, Aging, and Dead). The pink curve traces the sequestration rate, which peaks at the Maturing stage and drops below zero (horizontal black line) during Aging and Death.

Trees exhibit rapid growth and high uptake in early pioneer stage. In temperate forests, aboveground net primary productivity (ANPP) ranges from 2.57 to 5.78 Mg C ha⁻¹ yr⁻¹ in younger managed forests, while older natural areas may have lower ANPP around 2.51 Mg C ha⁻¹ yr⁻¹ 33. As forests progress into young and maturing phases, biomass accumulation continues with carbon storage at 0.7–2.0 Mg C ha⁻¹ yr⁻¹ in boreal and temperate forests of Canada34. Multiple carbon pools must be considered: aboveground biomass (AGB), belowground biomass (BGB), soil organic carbon (SOC), dead woody debris, and litter. Distribution varies by region—in Africa, living biomass accounts for approximately 60% with soil carbon around 34%; in Europe, soil carbon dominates at 64% with living biomass only 25%35,36.

However, the carbon sequestration potential of forests is not unlimited. Sequestration rates peak during maturing phase, then plateau in steady state before declining (Fig. 1B). Forests are also increasingly vulnerable to disturbances, such as deforestation, degradation, and climate change impacts, which can rapidly release stored carbon. Eddy-covariance results at the Wind River old-growth Douglas-fir site report annual NEP of 217 ± 40 g C m⁻² yr⁻¹ in 1999, but the stand became a net carbon source (−100 g C m⁻² yr⁻¹) in the unusually warm year (2003), indicating that this old-growth forest’s net carbon balance is strongly climate-sensitive37. The 2015–2016 El Niño-induced drought released up to 2.3 PgC to the atmosphere38. Recent satellite-based estimates suggest that gross tropical forest carbon loss associated with forest conversion increased to around 2 PgC per year in 2015–201939. These dynamics imply delayed and heterogeneous climate benefits from added forest area. Our analytical approach tests whether China’s extensive afforestation translated into immediate sequestration benefits, or whether temporal dynamics created more nuanced relationships.

Spatiotemporal patterns and drivers of forest coverage, CO₂ emissions, and economic growth

Our analysis of 30 Chinese provinces from 2004 to 2019 reveals significant spatiotemporal variations in forest coverage, economic growth, energy consumption, and CO₂ emissions.

Forest coverage increased significantly from 2004 to 2019 (Fig. 2). Southern provinces led expansion, with Guangxi rising by 7.53 percentage points to 60.43% and Fujian increasing from 64.43% to 68.20%. Central China showed significant gains, with Hubei rising from 31.11% to 39.61% and Hunan from 44.81% to 49.71%. Northern provinces also progressed, as Hebei improved from 22.27% to 26.76%. Western provinces such as Sichuan and Yunnan experienced substantial increases, from 34.31% to 38.01% and from 46.18% to 53.48%, respectively. Most significant changes occurred between 2009 and 2014.

Fig. 2: Spatial patterns of forest coverage rate (%) in mainland China from 2006 to 2019.
Fig. 2: Spatial patterns of forest coverage rate (%) in mainland China from 2006 to 2019.
Full size image

A–N Correspond to years 2006–2019, respectively. Color gradients range from low coverage (yellow-green) to high coverage (dark purple), as shown in the legend of (N). Gray areas indicate provinces with missing data.

GDP data reveal China’s economic transformation (2006-2019) in Fig. 3. Eastern coastal provinces-maintained leadership that Guangdong’s GDP quadrupled from 2.60 to 10.80 trillion yuan; Jiangsu and Shandong rose from 2.12 and 1.90 trillion yuan to 9.87 and 7.05 trillion yuan, respectively. Central provinces showed substantial expansion that Henan grew from 1.20 to 5.37 trillion yuan. Western provinces demonstrated remarkable growth that Sichuan rose from 0.85 to 4.64 trillion yuan. Most rapid growth occurred 2010–2015.

Fig. 3: Spatial patterns of GDP (trillion yuan) in mainland China from 2006 to 2019.
Fig. 3: Spatial patterns of GDP (trillion yuan) in mainland China from 2006 to 2019.
Full size image

A–N Correspond to years 2006–2019, respectively. Color gradients range from lower GDP (orange) to higher GDP (dark purple), as shown in the legend of (N). Gray areas indicate provinces with missing data.

Energy consumption patterns (1997–2019) highlight regional disparities (Fig. 4). Eastern provinces dominated overall growth, with Shandong increasing by over an order of magnitude from 157.40 to 1768.72 million tons SCE and Jiangsu rising from 132.07 to 852.63 million tons SCE. Central provinces also experienced substantial growth, exemplified by Henan, where values increased from 105.27 to 455.75 million tons SCE. In western China, Xinjiang saw a dramatic increase from 40.44 to 627.18 million tons SCE. The most rapid growth occurred between 2002 and 2013.

Fig. 4: Spatial patterns of energy consumption (million tons SCE) in mainland China from 2006 to 2019.
Fig. 4: Spatial patterns of energy consumption (million tons SCE) in mainland China from 2006 to 2019.
Full size image

A–N Correspond to years 2006–2019, respectively. Color gradients range from lower consumption (yellow-tan) to higher consumption (dark blue), as shown in the legend of (N). Gray areas indicate provinces with missing data.

CO₂ emissions (1997–2019) revealed substantial growth with pronounced regional variations (Fig. 5). Eastern and central provinces consistently showed higher emissions, with Shandong increasing from 199.30 to 937.12 million tons and Jiangsu from 183.90 to 804.59 million tons. Industrial centers in central China also saw substantial increases, exemplified by Henan, where emissions rose from 154.00 to 460.63 million tons. Western regions experienced notable growth as well, with Xinjiang increasing from 63.20 to 455.27 million tons. The most rapid increase occurred between 2002 and 2011.

Fig. 5: Spatial patterns of CO2 emissions (million tons) in mainland China from 2006 to 2019.
Fig. 5: Spatial patterns of CO2 emissions (million tons) in mainland China from 2006 to 2019.
Full size image

A–N Correspond to years 2006–2019, respectively. Color gradients range from lower emissions (orange) to higher emissions (dark purple), as shown in the legend of (N). Gray areas indicate provinces with missing data.

These widespread changes across diverse geographical regions underscore China’s comprehensive economic development and its associated environmental challenges. The improvements in forest coverage, accompanied by substantial increases in GDP, energy consumption, and CO₂ emissions, highlight the complex interplay between economic growth and environmental sustainability in China, necessitating further investigation into the causal relationships among these factors.

Prediction and driving forces of CO2 emissions and GDP

The LSTM-MLP model demonstrated exceptional accuracy predicting both CO₂ emissions and GDP, providing foundation for understanding the relative importance of different factors in China’s economic environmental system. Energy consumption variables, especially electricity and natural gas, emerge as the primary explanatory factors for variations in CO₂ emissions and economic growth (Figs. 3 and 5). This corroborates research highlighting strong coupling between energy use, growth, and emissions in China40,41, but raises questions about immediate effectiveness of forest-based interventions. The significance of electricity consumption as a predictor reflects China’s rapid industrialization and urbanization42, evidenced by Shandong’s consumption surging from 157.40 million tons SCE (1997) to 1768.72 million tons (2019).

Intriguingly, forest coverage demonstrated minimal importance in predicting either CO₂ emissions or GDP despite China’s extensive afforestation efforts (2004–2019). Guangxi expanded coverage from 52.90% to 60.43%, while Hunan increased from 44.81% to 49.71%. However, these increases did not correspond to proportional emission decreases. In Guangxi, CO₂ emissions actually increased from 86.60 to 246.72 million tons during this period. Similarly, Hunan’s CO₂ emissions rose from 125.39 million tons to 310.64 million tons.

Low predictive power of forest coverage for CO₂ emissions stems from: (1) Time lag. Delayed manifestation of carbon sequestration benefits arising from temporally variable sequestration rates across ecosystem life cycles; (2) Scale mismatch. Apparently, increases in local vegetation cover or carbon uptake are offset by emissions operating at broader spatial scales43; (3) Forest age structure. Heterogeneous age distributions constraining landscape-level sequestration capacity and long-term carbon accumulation22; (4) Carbon pool complexity. Incomplete accounting when AGB, BGB, SOC, dead woody debris, and litter are not simultaneously considered34,35; and (5) Disturbance vulnerability. Susceptibility of accumulated carbon stocks to rapid release under disturbances such as fire, harvest, pests, or extreme climate events33.

Pan et al. reveals global forest carbon sink remained stable at around 3.5–3.6 Pg C yr⁻¹ from 1990s to 2010s despite regional variations44. Temperate and tropical regrowth forests saw 30% and 29% sink increases; boreal and tropical intact forests experienced 36% and 31% decreases. This stability masks substantial biome-level changes. These findings address our initial hypothesis regarding forest coverage impact on emissions, revealing complex reality where energy consumption variables overwhelmingly dominate as predictors, aligning with time-dependent nature of forest carbon sequestration described in our theoretical framework.

Energy consumption patterns and economic-environmental relationships

Strong predictive power of energy consumption variables for CO₂ emissions and GDP is evident in parallel increases observed across Chinese provinces (1997-2019, Fig. 4). Energy consumption, economic growth, and CO₂ emissions exhibit a strong coupling in Shandong. Over the study period, GDP increased from 1.90 to 7.05 trillion yuan (2006–2019), accompanied by a rise in energy consumption from 157.40 to 1768.72 million tons of standard coal equivalent (1997–2019), and a concomitant increase in CO₂ emissions from 199.30 to 937.12 million tons. This pattern replicates across provinces, particularly in economically dynamic eastern and central regions. These trends align with environmental Kuznets curve hypothesis45, though continuing rise in both GDP and emissions suggests China may not have reached the turning point. This is evident in rapid growth (2002-2013) when many provinces saw energy consumption more than double. Identification of electricity and natural gas as key predictors highlights critical role of energy transition in China’s sustainable development path. Natural gas is often viewed as a cleaner alternative to coal, yet its role as a “transition” fuel can still involve meaningful emissions and policy trade-offs, underscoring the complexities of moving to a low-carbon economy46.

Continuing rise in both GDP and emissions despite substantial forest coverage increases challenges linear expectations about afforestation’s mitigation potential. While Law et al. demonstrated forest management could mitigate up to 15% of emissions in temperate regions47, our findings suggest without concurrent energy system transformation, even extensive afforestation programs may yield limited immediate climate benefits. This tension exemplifies the “forest carbon paradox”. Energy consumption strongly drives both emissions and GDP in China, consistent with an industrial structure dominated by energy-intensive sectors during 2000–2019 and with prior evidence linking carbon emissions to economic growth and energy use48. Natural gas showed significant predictive power, reflecting China’s gradual shift toward cleaner fossil fuels49. However, coal still dominated at approximately 56.8% of primary energy consumption in 2020 versus 68% in 200050.

Emissions and GDP across five factors

To elucidate relative influence of energy consumption and forest coverage, we employed our LSTM-MLP model across five scenarios with comprehensive feature importance analysis. As shown in Figs. 6 and 7, CO₂ emissions and GDP are driven by a set of distinct underlying factors.

Fig. 6: Feature importance analysis for CO₂ emissions and GDP using different models.
Fig. 6: Feature importance analysis for CO₂ emissions and GDP using different models.
Full size image

Feature importance for CO₂ emissions (A) and GDP (B), respectively, evaluated across multiple predictors; Feature importance for CO₂ emissions (C) and GDP (D), respectively, evaluated by the new LSTM-MLP model for select predictors.

Fig. 7: Deep learning prediction results for CO2 emissions and GDP across five scenarios.
Fig. 7: Deep learning prediction results for CO2 emissions and GDP across five scenarios.
Full size image

A CO2 emission predictions and B GDP predictions, each comparing predicted values against actual values with different symbols representing each scenario. Scenario 1 includes all features; Scenario 2 drops forest coverage; Scenario 3 includes only electricity; Scenario 4 includes only gasoline; and Scenario 5 removes all features but forest coverage. The dashed diagonal line indicates perfect prediction (predicted = actual).

For CO₂ emissions (Fig. 6A, C), electricity consumption was most critical, showing largest MAE increase when removed. Natural gas and gasoline followed in importance. Forest coverage showed minimal importance with nearly zero MAE change when dropped. Similar patterns emerged for GDP prediction (Fig. 6B, D), with electricity most crucial, followed by gasoline and natural gas consumption. Forest coverage consistently demonstrated limited importance (score around 0).

To further investigate these relationships, we tested our LSTM-MLP model across five scenarios, with results presented in Fig. 7. For CO₂ emissions prediction (Fig. 7A), Scenario 1 (all features) achieved highest R² of 0.9864. Notably, Scenarios 3 and 4 (electricity and gasoline only) demonstrated robust performance with R² of 0.9597 and 0.9599. Scenario 5 (forest coverage only) showed significantly lower R² of 0.8756. Similar patterns emerged for GDP prediction (Fig. 7B) that Scenario 1 performed best (R² = 0.9912), closely followed by Scenarios 3 and 4 (R² = 0.9860 and 0.9542). Scenario 5 showed weakest performance (R² = 0.7326). Prediction error distributions confirmed Scenarios 1–3 yield the most accurate predictions, while Scenarios 4 and 5 exhibit increasing bias and dispersion (Supplementary Fig. 1.2).

The contrast challenges simplistic assumptions about afforestation as immediate mitigation. While Fajardy et al. and Smyth et al. have highlighted potential long-term benefits of afforestation51,52, immediate mitigation efforts may need a different focus on energy sector interventions. Moreover, it is important to note that while these findings provide valuable insights, the permutation-based feature importance techniques used in this analysis cannot offer valid statistical testing of the results53. This limitation underscores the need to explore relationships using robust statistical models. To address model sensitivity, we implemented rigorous hyperparameter optimization using grid search across learning rates (0.001–0.01), hidden layer dimensions (32–128 nodes), and dropout rates (0.1–0.5). Final model configuration demonstrated stable performance across test sets, with R² variations below 0.05. To mitigate overfitting, we employed early stopping (patience = 10 epochs). Consistency between training and testing performance (average RMSE difference less than 5%) indicates successful regularization. Our energy consumption variables exhibit moderate intercorrelations (mean r = 0.29). We performed repeated permutations (n = 20) and report mean importance scores with standard errors, providing a more robust assessment of feature significance.

Unraveling causal relationships of forest coverage, economic growth, and environmental impact

To deepen understanding of interactions between energy consumption, growth, CO₂ emissions, and forest coverage, we employed advanced causal inference techniques: mCCM and time-delayed mCCM. Initial mCCM uncovered significant bidirectional causality between electricity and gasoline consumption. However, time-delayed model suggested more nuanced relationships (Supplementary Fig. 2.1D): gasoline consumption likely forces electricity consumption, indicated by the negative optimal cross-mapping lag from gasoline to electricity consumption and a vague optimal lag in the reverse direction.

We first examined the relationship between forest coverage and GDP. Initial mCCM analysis showed that GDP demonstrated higher cross-mapping skill on forest coverage than the reverse (Fig. 8A), but this direction is not statistically significant (P = 0.374), so we do not claim a robust causal effect. Time-delayed mCCM shows peak skill when GDP leads forest coverage (Fig. 8B), consistent with a policy-driven pathway, but we interpret this timing pattern cautiously given the lack of statistical support. The weak reverse effect indicates that forest coverage changes have limited immediate impact on aggregate economic output.

Fig. 8: Causal Analysis among Forest Coverage, GDP, and CO₂ emissions.
Fig. 8: Causal Analysis among Forest Coverage, GDP, and CO₂ emissions.
Full size image

Pairing results are shown of mCCM (left panel) and time-delayed mCCM (right panel). Forest coverage vs GDP (A, B), CO₂ emissions vs GDP (C, D), and CO₂ emissions vs forest coverage (E, F). For plots on the left panel, lines and shaded regions show mean cross map skill and standard deviation over 420 libraries (number of each variable sample and all spatial replicates in the corresponding composite time series). The number of bootstrapped iterations is set to 1000. For plots on right panel, the effect of variable x on y is extended with lag of ± 8 years. The cross map skill at the lag of 0 years corresponds to the convergent value (at sufficiently large library length) from plots on left panel.

Energy consumption strongly drives economic growth: both electricity and gasoline show causal effects on GDP with comparable strength of 0.85 (Supplementary Fig. 2.1A and 2.2A). The relationship between GDP and CO₂ emissions revealed an interesting pattern. While initial analysis showed trend suggesting bidirectional forcing effects, it was not statistically significant (Fig. 8C). Time-delayed mCCM subsequently confirmed bidirectional causal relationship (Fig. 8D).

Given China’s extensive afforestation, we next examined forest coverage role. Contrary to expectations, forest coverage suggested a relatively weak causal relationship with CO₂ emissions short-term. Analysis revealed strong causal effect of CO₂ emissions on forest coverage (0.88), but much weaker reverse effect (0.49) (Fig. 8E). Time-delayed mCCM provided further insight, showing upward trend in predictive skill when CO₂ emissions cross-map forest coverage, and downward trend opposite direction (Fig. 8F). Combined with negative optimal cross-mapping lag from forest coverage to CO₂ emissions and positive optimal lag from CO₂ to forest coverage, these results suggest a unidirectional causal forcing from CO₂ emissions to forest coverage.

This asymmetrical relationship likely reflects China’s policy-driven afforestation response to environmental degradation. Rising emissions and environmental concerns prompted increasingly ambitious forest expansion policies, such as the expansion of the Three-North Shelter Forest Program in its later phases (2001–2020) and the extension of the Natural Forest Protection Program in response to escalating environmental challenges. Related policy-design work by Guo et al. shows that integrating provincial carbon sinks into China’s carbon quota and transfer-payment framework channels resources from high-emission, developed provinces to sink-rich regions, linking carbon-sink development to regional emissions responsibilities54. Meanwhile, a weaker reverse causal effect reflects the biological reality that newly established forests require decades to reach full sequestration potential, creating a significant time lag between policy implementation and measurable emissions reduction55,56. To provide a national-scale context for these findings, we aggregate provincial totals to national series for 2004–2019 and overlay forest coverage with national CO₂ emissions, annotated by major policy periods (Fig. 9). Both metrics increase concurrently. This pattern aligns with our mCCM asymmetry findings. The stronger CO₂ leading forest direction is consistent with a policy-response channel in which rising environmental pressure and emissions-related concerns are followed by intensified forest programs and coverage increases on relatively short horizons (years to a few years). By contrast, the weaker forest leading CO₂ direction is consistent with the fact that national emissions during 2004-2019 are dominated by energy-system drivers, and that forest coverage is an area proxy rather than a direct measure of net carbon uptake. Accordingly, even if forests contribute to mitigation, their aggregate signal in national emissions may be difficult to detect within the study window and with area-only measures. Additionally, forest coverage does not directly encode carbon stock, density, or management quality; therefore, coverage increases need not translate proportionally into near-term net carbon uptake. Finally, biophysical processes governing carbon accumulation typically involve lags and variability that may extend beyond the short horizons over which national emissions are commonly evaluated. We frame these mechanisms as plausible explanations aligned with observed timing, while avoiding attribution beyond what the data supports.

Fig. 9: National-scale temporal context for forest coverage and CO₂ emissions during China’s major environmental policy periods (2004–2019).
Fig. 9: National-scale temporal context for forest coverage and CO₂ emissions during China’s major environmental policy periods (2004–2019).
Full size image

Aggregate national forest coverage (green line, left y-axis) and aggregate national CO₂ emissions (orange line, right y-axis) derived from provincial totals. Shaded regions indicate major policy implementation periods: Three-North Shelter Forest Program Schedule 4 (2002–2010, light green), 11th Five-Year Plan (2006–2010, light green), 12th Five-Year Plan (2011–2015, tan), and 13th Five-Year Plan (2016–2020, light purple).

Our findings provide important insights into forest carbon markets globally, highlighting three key challenges. First, temporal mismatch between forest establishment and peak carbon uptake creates fundamental biological constraint that current market mechanisms inadequately address. Second, static pricing models fail to capture dynamic nature of forest carbon sequestration—even at higher carbon prices ($25/t–$54/t), forest carbon sinks would only offset 2.3–5.6% of China’s total emissions from 2021 to 206056. Third, verification challenges further limit effectiveness—among twelve registered forestry projects, only one completed verification and received credits by 201628. These findings suggest forest carbon markets would benefit from time-differentiated valuation approaches better reflecting actual sequestration patterns.

To synthesize relationships, we constructed a causal network among five factors studied by mCCM and time-delayed mCCM (Fig. 10). Gasoline and electricity consumption are identified as strongly significant drivers of GDP (causal strengths 0.84 and 0.83). Gasoline consumption also demonstrated a significant causal effect (0.61) on CO₂ emissions (Supplementary Fig. 2.1C). GDP exerted a moderate causal effect (0.55) on CO₂ emissions, with a slightly weaker reverse relationship (0.48). Notably, CO₂ emissions showed a strong causal effect (0.88) on forest coverage. The causal network also reveals indirect pathways between gasoline consumption, CO₂ emissions, and forest coverage (Supplementary Note 2). We found two paths: a transitive causal chain where gasoline causes CO₂ emissions, which in turn causes changes in forest coverage, and a direct path where gasoline affects forest coverage. The stronger link from CO₂ to forests (0.88) versus gasoline to forests (0.81) suggests CO₂ emissions are the more direct causal driver of forest coverage changes. A strong bidirectional relationship exists between gasoline and electricity consumption (causal strengths 0.76 and 0.72), although the validity of the direction from electricity to gasoline is challenged via time-delayed mCCM. Robustness and validity are further supported by diagnostic plots (Supplementary Note 3) confirming assumptions of non-linearity and non-randomness underlying mCCM models (Supplementary Fig. 3A–C).

Fig. 10: Estimated causal network across five factors.
Fig. 10: Estimated causal network across five factors.
Full size image

(Note, the solid lines represent the relationships that have clear convergence in mCCM and were confirmed by extended mCCM with negative optimal lags. The dotted lines represent relationships that show only one of the aforementioned properties in solid lines).

The strong causal links from energy consumption to both GDP and CO₂ emissions, coupled with a weaker short-term forest coverage impact, suggest China’s current trajectory still prioritizes energy-intensive growth57. The bidirectional GDP-CO₂ emissions relationship challenges simplistic environmental Kuznets curve interpretations58. The asymmetrical relationship between CO₂ emissions and forest coverage is qualitatively consistent with provincial-scale studies that highlight strong spatial heterogeneity in emissions responsibilities and emphasize the need to couple carbon sinks with regional emission obligations54,59.

Our comprehensive analysis reveals the “forest carbon paradox”. Despite China’s remarkable success in expanding forest coverage, these efforts demonstrate limited immediate impact in mitigating CO₂ emissions or influencing economic growth patterns. This extends beyond China. Growing global evidence shows forest-based carbon credits often overestimate near-term climate benefits. Systematic assessment across 2346 projects found fewer than 16% of issued credits correspond to real emission reductions60. Corporate offset portfolios are dominated by low-quality credit, with 87% carrying a high risk of being non-additional or otherwise low-integrity61. Evaluations of avoided-deforestation projects identify systematic over-crediting62,63. Carbon accumulation from reforestation unfolds over decades with wide spatial variation globally, and managed temperate forests can even contribute to warming rather than deliver expected climate benefits64,65. Reflecting these constraints, the IPCC gives persistent concerns about additionality and permanence for land-based mitigation66. Taken together, these global patterns reinforce our core finding that forest cover has limited short-run predictive power for emissions and economic growth in China. This pattern aligns with structural features of forest carbon markets and supports the need for time-differentiated valuation and closer integration of forest strategies with energy-sector mitigation.

Policy implications and future directions

Our results reveal a fundamental timing mismatch in forest-based climate mitigation. National-scale evidence shows forest coverage and CO₂ emissions both increased during 2004 and 2019 (Fig. 9), indicating area expansion builds long-term sink potential but does not produce proportional near-term emissions reductions. This necessitates a portfolio approach: protecting mature, high-carbon-stock forests delivers immediate climate value by preserving existing stocks and reducing reversal risk, while expansion programs should be designed for permanence through enhanced survival, management, and monitoring.

Four actionable recommendations emerge. First, time-sensitive crediting: carbon credit issuance and pricing must account for delayed sequestration benefits rather than assuming immediate mitigation from area gains. Second, conservation prioritization: mature forest protection should receive equivalent policy attention to expansion programs, given its immediate climate value. Third, improvements in MRV systems, achieved by combining scalable remote sensing with targeted field validation, can lower transaction costs and address the shortcomings of area-only metrics. Fourth, energy-sector complementarity is critical, as energy-related drivers dominate emissions dynamics and require forest-based strategies to complement, rather than replace, rapid decarbonization across electricity, industry, and transport. Future research should integrate direct carbon stock and biomass measures to strengthen inference about sequestration magnitude and timing.

Methods

Data collection

Our analysis covers 30 mainland China provinces (2000–2019), excluding Hong Kong, Macao, Taiwan, and Tibet for data consistency. Forest coverage data were sourced from three consecutive National Forest Resource Inventories: 7th (2004–2008), 8th (2009–2013), and 9th (2014–2018) inventories. Additional data for 2019-2020 were obtained from China National Forestry and Grassland Administration reports. Provincial land area data derived from China National Bureau of Statistics’ provincial area reports. Carbon emission data obtained from China Carbon Emissions Database (CEAD). All other raw data were extracted from China Statistical Yearbook. All economic data were deflated to 1999 constant prices.

The period 2000–2019 was selected to capture rapid economic development and afforestation policy implementation, including operationalization of China’s National Forest Protection Program (NFPP) and Grain for Green Project (GGP). This timeframe encompasses critical expansion phase of China’s major ecological restoration initiatives. Data availability and consistency across all 30 provinces made this period particularly suitable for our analytical approach, requiring complete time series for LSTM-MLP modeling and convergent cross-mapping analyses.

While forest coverage serves as a useful proxy for forest carbon stocks, we acknowledge limitations in fully representing sequestration dynamics. Forest age significantly influences carbon uptake rates67. Forest type composition also substantially affects sequestration potential: empirical estimates of aboveground carbon accumulation in China range from about 23.6 to 82.9 Mg C ha⁻¹ yr⁻¹ across forest types, corresponding to roughly 3–4-fold differences in early-stage carbon uptake capacity68. Despite limitations, we selected forest coverage data from China’s National Forest Resource Inventories due to consistency across all provinces throughout our study period, direct relevance to China’s afforestation policy targets, and reliability as ground-verified information.

Hybrid models and feature importance assessment

We developed a hybrid deep learning model combining Long Short-Term Memory Networks (LSTMs) with Multi-Layer Perceptrons (MLPs) to analyze temporal dependencies. This LSTM-MLP architecture predicted both GDP and total CO₂ emissions. LSTMs were selected for their exceptional ability to capture long-term dependencies in sequential data69. Unlike traditional RNNs struggling with the vanishing gradient problem, LSTMs utilize intricate cell structures with gating mechanisms that effectively regulate information flow.

The LSTM component processes sequential data through the following equations:

$${f}_{t}={\rm{\sigma }}\left({W}_{f}\cdot \left[{h}_{t-1},{x}_{t}\right]+{b}_{f}\right)$$
(1)
$${i}_{t}={\rm{\sigma }}\left({W}_{i}\cdot \left[{h}_{t-1},{x}_{t}\right]+{b}_{i}\right)$$
(2)
$$\widetilde{{C}_{t}}=\tanh \left({W}_{C}\cdot \left[{h}_{t-1},{x}_{t}\right]+{b}_{C}\right)$$
(3)
$${C}_{t}={f}_{t}\odot {C}_{t-1}+{i}_{t}\odot \widetilde{{C}_{t}}$$
(4)
$${o}_{t}={\rm{\sigma }}\left({W}_{o}\cdot \left[{h}_{t-1},{x}_{t}\right]+{b}_{o}\right)$$
(5)
$${h}_{t}={o}_{t}\odot tanh\left({C}_{t}\right)$$
(6)

⊙ denotes element-wise multiplication, σ and tanh denote sigmoid and hyperbolic tangent functions. The structure is shown in Supplementary Fig. 1.1. Input layers process time series data from various energy consumption sources and forest coverage70. Forest coverage was included as a primary feature due to potential nonlinear interactions with target variables. Each series is processed by an individual LSTM layer, outputs concatenated, then fully connected to an MLP, generating the final prediction.

To quantify the relative importance of input features, we employed permutation-based importance assessment71. This approach measures the increase in prediction error when a feature’s values are randomly shuffled;

$$P{I}_{j}=\frac{1}{K}\mathop{\sum }\limits_{k=1}^{K}\left[L\left(Y,\,f\left({X}_{{perm},j}^{\left(k\right)}\right)\right)-\,L\left(Y,\,f\left(X\right)\right)\right]$$
(7)

where \(P{I}_{j}\) is the permutation importance of feature \(j\), \(K\) is the number of permutation iterations, \(L\) represents the loss function, \(Y\) denotes the true values, \(f\) is the trained LSTM-MLP model, and \({X}_{{perm},j}^{\left(k\right)}\) indicates the feature matrix with the \(j\)-th feature randomly shuffled in \(k\)-th iteration. This approach posits that the larger the increase in prediction errors following feature permutation, the greater the impact that feature is deemed to have on the outcome of interest72.

Our variable selection strategy is grounded in established work from energy economics and environmental policy. Energy consumption is both a core input to economic production and the main proximate driver of CO₂ emissions, so multiple energy carriers are included alongside GDP. The Environmental Kuznets Curve (EKC) literature further motivates analyzing growth and environmental pressure jointly rather than in isolation. Forest coverage is included because it summarizes large-scale afforestation and conservation efforts that are central to China’s climate policy and to our research question, while our analysis is designed to test rather than assume its short-run influence on aggregate emissions.

Our analysis incorporates 20 energy consumption variables systematically categorized based on distinct roles in China’s energy-economy system (Supplementary Table 1). Coal-related products are theoretically essential given coal’s dominance at 57.32% of China’s primary energy consumption in 2019. Industrial process fuels capture China’s heavy industrial structure. Petroleum products are disaggregated based on distinct economic functions. Additional energy forms (natural gas, heat, and electricity) represent critical components of China’s evolving energy system.

For subsequent causal analysis, we implemented a systematic variable selection procedure to identify key drivers. Five scenario-based experimental frameworks were constructed to evaluate relative explanatory power: (1) all variables combined; (2) forest coverage with selected energy variables; (3) electricity consumption alone; (4) gasoline consumption alone; and (5) forest coverage alone. This variable selection strategy is supported by established energy–economy literature documenting robust causal linkages between energy consumption and economic growth in China73,74. From emissions perspective, power sector (encompassing electricity and heat) accounted for approximately half of China’s energy-related CO₂ emissions in 2019, while oil-based fuels comprise over 70% of transport energy consumption75.

To rigorously evaluate model accuracy, we employed Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and coefficient of determination (R²) metrics (Supplementary Table 2). For assessing feature importance, we implemented permutation-based techniques quantifying changes in prediction errors71. However, this approach cannot offer valid statistical testing on feature importance76. We applied subsequent statistical models to validate and extend the results obtained from LSTM–MLP model.

Causality analysis

To elucidate causal relationships within China’s interrelated economic-environmental system, we employed multispatial Convergent Cross Mapping (mCCM), an advanced variant specifically designed for spatially replicated time series with short observational sequences53. This method leverages spatial replication across provinces to compensate for the limited temporal extent of each individual series.

Each of China’s 30 provinces serves as an observational unit. The mCCM tests are based on 14 annual observations (2006–2019) for each province, resulting in substantial library length (L) of 420. The mCCM analysis proceeded through: (1) determined optimal embedding dimensions; (2) tested for nonlinearity; (3) generated diagnostic plots; (4) conducted cross-mapping analysis; (5) performed time-delayed analysis (Supplementary Fig. 3.1–3.3); (4) Apply bootstrapping with replacement; (5) Employ nonparametric bootstrapping to test for significant causal relationships; and (6) consider time-delayed effect to validate inference on causal directions. While mCCM offers significant advantages, it inherits certain limitations from standard CCM technique, including challenge of distinguishing true bidirectional causality from synchrony77.

To enhance robustness, we implemented extended version of mCCM that explicitly considers time lags78. In case of synchrony caused by strong unidirectional forcing, this approach detects negative lag for cross mapping in true causal direction and positive lag in other directions. This allows us to determine whether driving variable impacts response variable with yearly delays, enabling more precise inferences about true causal directions.

The combination of mCCM with our LSTM-MLP models allows us to identify predictive relationships and establish causal mechanisms, providing a more complete understanding of how forest coverage interacts with economic and environmental variables across China’s diverse regions. Detailed model architecture is provided in Supplementary Note 1.