Assessing Assisted Natural Regeneration as a Cost‐Efficient Mitigation for Climate Change and Biodiversity Loss in China

Abstract

Global deforestation results in climate change and biodiversity loss. Assisted natural regeneration (ANR) emerges as a promising approach to achieving global forest restoration targets, yet its potential and benefits for climate and biodiversity in China remain underexplored. Here, we assessed ANR potential across China and modeled spatial prioritization strategies targeting climate mitigation, biodiversity conservation, and cost savings, individually and in combination, as well as strategies considering spatial constraints from current forest restoration projects and fragmentation mitigation. From 1995 to 2015, 3.40 million hectares of land naturally regenerated into forests, with an additional 5.11 million hectares identified as potential regeneration areas, which could contribute to 12.41% of China's restoration goal in 2035. Spatial prioritization revealed limited overlap among the three single‐objective ANR strategies, while a multi‐objective optimization strategy emerged as the most effective solution to achieve synergies among goals. The top 30% of prioritized areas under the multi‐objective strategy could sequester 46.54 gigatons of CO₂, reduce extinction risks of forest‐dependent species by 40.64%, and lower implementation costs by 31.55% compared to monoculture tree plantations. Our findings highlight that strategic spatial prioritization of ANR could mitigate climate change and biodiversity loss in a cost‐efficient manner and have the potential to reinforce current forest restoration projects.

Introduction

Rapid deforestation across the globe has led to substantial carbon emissions and biodiversity loss (Food and Agriculture Organization of the United Nations, 2020; Hua et al., 2022; Knoke et al., 2023). With growing awareness of the environmental hazards caused by deforestation, international initiatives such as the Bonn Challenge and the New York Declaration on Forests have mobilized 61 countries to commit to halting deforestation and restoring at least 200 million ha of degraded forest between 2020 and 2030, which is an area comparable to the size of Indonesia (Forest Declaration Assessment, 2021; International Union for Conservation of Nature, 2011; Lewis et al., 2019). These commitments mark a significant global campaign toward more sustainable and resilient development under climate change (Menz et al., 2013). The disclosure of proposed restoration plans in these commitments reveals that, however, monoculture tree plantations will account for 45% of the restoration efforts, making them the most commonly adopted approach in major countries such as China and Indonesia (Hua et al., 2016; Lewis et al., 2019). To meet these ambitious goals, the tree-planting rate is estimated to be at least 9 million ha/yr on average, whereas the currently ongoing projects of monoculture tree plantations amount to only 6.8 million ha/yr across the globe (Lewis et al., 2019). More importantly, with an average planting cost exceeding US$1,200 per hectare (Sinacore et al., 2023), the annual expenditure for monoculture tree plantations would reach US$10.8 billion. Hence, a significant gap remains between these pledges and implementation given the slow planting rates and tremendous cost of the currently adopted restoration pathway (Menz et al., 2013).

The assisted natural regeneration approach (hereafter ANR) has emerged as a promising, cost-efficient alternative to traditional restoration approaches, by integrating active interventions with passive restoration actions (Holl, 2017; Lewis et al., 2019; Poorter et al., 2016). Indeed, ANR leverages local community knowledge to eliminate obstacles and threats to forest regeneration on marginal lands, including removing invasive shrubs to foster native tree species growth, and building fences to protect saplings from livestock in intensively used grasslands (Alves et al., 2022; Shono et al., 2020). The effectiveness of ANR has been proven in many regions (Alves et al., 2022; Williams et al., 2024). For example, large-scale forests have been regenerating naturally following agricultural abandonment with active assistance in Europe and the United States, and this trend is now becoming evident globally (Chazdon et al., 2020; Shono et al., 2020).

As a leading participant in both the United Nations Framework Convention on Climate Change (UNFCCC) and the Convention to Combat Desertification (UNCCD), China has contributed 25% of the world's net increase in greening areas (Chen et al., 2019) and is committed to reforesting 10% of global pledged restoration areas by 2030 (China National Committee for the Implementation of the United Nations Convention to Combat Desertification, 2023; Sewell et al., 2020; United Nations Environment Programme, 2021). Despite such remarkable achievement, 98.8% of China's reforested areas comprise monoculture plantations, whose ecological benefits for climate and biodiversity as well as sustainability remain questionable, especially in arid ecosystems (Hasler et al., 2024; Hua et al., 2022). In this context, Chinese authorities are shifting their restoration pathway from extensive monoculture plantations to integrating artificial afforestation with ecosystem self-recovery, making ANR particularly pertinent to China's restoration efforts (Luo et al., 2023; Qi & Dauvergne, 2022).

ANR is not, however, a one-size-fits-all solution for reforestation endeavors across the whole of China. As with other nature-based climate solutions, its effectiveness varies across landscapes and socio-economic contexts (Hua et al., 2022). ANR works best in landscapes that are not highly degraded and embedded within forest remnants, where forest regeneration is more likely to occur (Alves et al., 2022). Moreover, farmers are unlikely to embrace ANR on their productive land unless they get substantial subsidies, such that the marginal land that is poorly suited for agriculture may be a preferred option for ANR (Alves et al., 2022; Shono et al., 2020). In China, large-scale farmland redistribution to marginal lands has occurred due to the inappropriate implementation of the “Requisition-Compensation Balance of Farmland” policy, resulting in a surge in fertilizer usage and habitat loss, and reduced crop yields (Kuang et al., 2022). Thus, these marginal lands represent an opportunity to support forest restoration without compromising food security (Alves et al., 2022). Despite its potential, there is still a lack of comprehensive knowledge of the factors driving forest regeneration and how to spatially prioritize ANR to optimize ecological outcomes. This knowledge gap hinders the effective deployment of ANR at the national scale in China (Lu et al., 2022; Xu et al., 2023).

Here, we assessed forest regeneration potential across China and developed spatially explicit prioritization strategies for ANR to achieve win-win outcomes for climate and biodiversity. Specifically, we addressed the following questions: (a) Which natural and socio-economic factors facilitate forest regeneration? (b) Where are the areas with potential for forest regeneration? (c) How can ANR maximize the ecological benefits for climate and biodiversity in a cost-efficient manner? Our study focused on estimating ANR potential at a broad scale rather than local solutions, aiming to provide a roadmap for operationalizing ANR as a national-scale restoration strategy.

Materials and Methods

To address the aforementioned questions, we first identified areas in which forest regeneration occurred between 1995 and 2015, based on the criteria proposed by Crouzeilles et al. (2020). Then, we constructed a random forest model to assess the relationships between forest regeneration and a wide range of natural and socioeconomic variables. This assessment provided a foundation for identifying areas suitable for ANR implementation (Crouzeilles et al., 2020; Pedregosa et al., 2011). Subsequently, we assessed the cost and benefits of implementing ANR across all potential regeneration areas (referred to as the “ANR4All” scenario). We also designed ANR optimization scenarios aligned with the objectives outlined in the UN Decade on Ecosystem Restoration (Strassburg et al., 2020; United Nations Environment Programme, 2021) with an optimization algorithm called non-dominated sorting genetic algorithm-II (a.k.a. NSGA-II). This optimization algorithm enabled spatially explicit prioritization of ANR, especially in the top 30% priority areas in line with the Global Biodiversity Framework Target 2 (Convention on Biological Diversity, 2022; Deb et al., 2002). We constructed three single-objective scenarios, each targeting climate change mitigation (i.e., higher carbon sequestration), biodiversity conservation (i.e., reduced extinction risks), or cost efficiency (i.e., lower cost), respectively; as well as three multi-objective scenarios that balanced these three objectives subject to varying spatial constraints. Lastly, we compared the ecological benefits and costs of different optimization strategies with those of the ANR4All scenario and random selection scenarios. The results were used to propose actionable recommendations for ANR-based restoration planning.

Identification of Forest Regeneration Types

Three forest regeneration types were considered in this study: regenerated areas, non-regeneration areas, and potential regeneration areas. Regenerated areas referred to areas where land cover was transformed from cropland or grassland to forest via natural regeneration, and were identified using the China land cover raster data set v1.0.0 (30-m resolution) (Yang & Huang, 2021), Global mangrove extent v2.0 (Bunting et al., 2022), and Global forest management data for 2015 (Lesiv et al., 2021). See Table S1 in Supporting Information S1 for detailed data sources. Following the criteria proposed by Crouzeilles et al. (2020), we identified regenerated areas over the past 20 years (1995–2015) as areas of (a) at least five contiguous grid cells previously belonging to grassland or cropland before conversion to forest for at least five consecutive years, (b) stable forest cover for at least three consecutive years as until 2015, and (c) excluding mangrove forests and those forests established by tree plantation or agroforestry. These criteria are suggested to mitigate potential grid cell misclassification resulting from immediate clearing for grazing and agriculture, temporary forest conversion, and alternative forest establishment (Crouzeilles et al., 2020).

For grid cells that remained as grassland or cropland throughout the past two decades in forest biomes, they were classified as non-regeneration areas (Crouzeilles et al., 2020). As for those grid cells that temporally transitioned from grassland/cropland to forest between 1995 and 2015 but recovered to grassland/cropland by 2015 in forest biomes (mangroves, tree plantations, and agroforestry excluded), they were considered potential regeneration areas. These areas might require human assistance to facilitate regeneration under environmental pressures or human modifications.

Estimation of Forest Regeneration Potential

We constructed a random forest model to identify factors influencing forest regeneration patterns, incorporating environmental, socio-economic, and landscape predictors. Key variables included: environmental conditions (e.g., climate, topography, and soil properties) that determine tree growth and succession (Curtis et al., 2018; Ewers et al., 2006; Flexas & Medrano, 2002), socio-economic indicators (e.g., population, economic gains from agricultural and livestock products) that influence land-use decisions and forest management practices (Ewers, 2006; Hansen et al., 2013; Wright & Muller-Landau, 2006), and landscape patterns (e.g., dominant land use types and land use intensity) that capture the interplays between environment and socio-economic activities (Crk et al., 2009; Curtis et al., 2018; Laurance et al., 2009; Molin et al., 2018). See Table S2 in Supporting Information S1 for the list of predictors and examples of their potential impacts on forest regeneration, and Text S2.1 in Supporting Information S1 for details on predictor preprocessing procedures.

To construct the training and validation data sets, we implemented the following procedures: First, we created a grid of 300 × 300 km to divide the territory of mainland China into 159 square polygons. Then, we generated 10,000 random points (at least 100 m away from each other) separately for regenerated areas and non-regeneration areas within each square polygon, and extracted values from raster maps of the aforementioned predictors. This resulted in 3.18 million sample points (i.e., 159 × 10,000 × 2), which were then split into 90% for model training and 10% for model validation. When training the random forest model, we used regeneration status (1 for the regenerated area; 0 for the non-regeneration area) as the response variable and those biophysical and socio-economic indicators as predictors and set the number of trees as 500. Lastly, we applied the random forest model to estimate the potential of natural regeneration within potential regeneration areas. We calculated the variable importance value for each predictor using the importance score, and then fitted generalized additive models to explore how these predictors influence forest regeneration potential (Felipe-Lucia et al., 2020), due to the complex relationships identified in prior studies (Crouzeilles et al., 2020; Williams et al., 2024). See Text S2.2 in Supporting Information S1 for details on model construction.

We compared the performance of the random forest model that included all variables with a model using only biophysical variables. The biophysical-only model was conducted with the predictors related to environmental conditions and landscape patterns (Williams et al., 2024). We then validated the random forest model including all variables using internal and external data sets (Araújo et al., 2005; Matutini et al., 2021). For internal validation, we used k-fold cross-validation (k = 10) on 3.18 million sample points, testing the model's robustness under various data splits and its adaptability across China's geophysical divisions. External validation was performed to test the model's generalization and to minimize the potential limits of data-splitting validation methods due to spatial and temporal autocorrelation (Matutini et al., 2021). We used a data set of China's additional natural forest maps from 1990 to 2020 (Cheng et al., 2024), focusing on the period from 1995 to 2015 to ensure temporal consistency with our study. Areas with additional natural forests were labeled as forest regeneration occurrences (1), and areas without as non-occurrences (0) (Chazdon et al., 2020). We applied the random forest model including all variables on the external data set to predict forest regeneration occurrences, and used k-fold cross-validation (k = 10) to test the model's generalization on the external data set (Matutini et al., 2021). Model classification performance was assessed using the area under the curve (AUC). Prediction performance was measured by accuracy, positive predictive rate (PPR), negative predictive rate (NPR), sensitivity (SEN), and specificity (SPE). See Text S1 in Supporting Information S1 for detailed validation. These processes were performed under the Python environment (ver.3.8) with several packages, including arcpy 3.0, numpy 1.21.5, pandas 1.3.5, scipy 1.10.1, and scikit-learn 1.4.2 (Pedregosa et al., 2011; Virtanen et al., 2020).

Prioritization of ANR

We aggregated the estimated natural regeneration raster to a coarser resolution of 100 m to improve computational efficiency over mainland China. This aggregation allowed us to calculate the proportion of forest cover within each 1-ha spatial unit for further assessments of total restoration cost, potential benefits for climate and biodiversity, as well as ANR prioritization (Cai et al., 2022; Strassburg et al., 2012, 2019, 2020).

Total Cost

We considered the total cost of ANR as the sum of implementation cost and opportunity cost, with the reduction in the total cost representing the benefits of cost savings: 1 $${total\,cost}_{i}={{implementation\,cost}_{i}+opportunity\,cost}_{i}$$ where $${total\,cost}_{i}$$, $${implementation\,cost}_{i}$$, and $${opportunity\,cost}_{i}$$ are the total cost of ANR, implementation cost, and opportunity cost (unit: US$ in 2010 as the benchmark, same below) in spatial unit $$i$$, respectively.

The implementation cost was estimated as the tree planting cost adjusted by the potential for forest regeneration following Strassburg et al. (2019): 2 $${implementation\,cost}_{i}=\left(1-{regenerated\,potential}_{i}\right)\times planting\,cost$$ where $${regenerated\,potential}_{i}$$ is the potential for forest regeneration in spatial unit $$i$$; $$planting\,cost$$ is the expense for tree planting per unit area (unit: US$/ha). Here, the planting cost was set as US$6,363.67/ha by dividing the total investment by the total plantation area in China in 2015, which was obtained from the China Forestry Statistical Yearbook in 2016 (National Forestry and Grassland Administration of China, 2017).

The opportunity cost was estimated as the agricultural or pastoral income from cropland or grassland if ANR is not implemented, as suggested by Strassburg et al. (2020): 3 $${opportunity\,cost}_{i}={agropastoral\,production}_{i}\times unit\,price$$ where $${agropastoral\,production}_{i}$$ is the production of each agropastoral product (unit: tons/ha) in spatial unit $$i$$; $$unit\,price$$ is the price per ton of the corresponding product (unit: US$/tons). Agricultural production was obtained from the Global Spatially Disaggregated Crop Production Statistics Data (Yu et al., 2020), pastoral production from Gridded Livestock of the World (Gilbert et al., 2018), and unit price from the China Rural Statistical Yearbook in 2016 (National Bureau of Statistics of China, 2017a). See detailed types of agropastoral products in Table S3 in Supporting Information S1.

Climate Change Mitigation

We considered carbon sequestration as the restoration benefit for climate change mitigation, and estimated it as the potential accumulation of aboveground biomass (AGB) and soil organic carbon (SOC) in China's forests following Cai et al. (2022): 4 $${carbon\,sequestration}_{i}={AGB}_{i}+{SOC}_{i}$$ where $${carbon\,sequestration}_{i}$$, $${AGB}_{i}$$, and $${SOC}_{i}$$ denote total carbon, AGB, and SOC (unit: Mg/ha) in spatial unit $$i$$, respectively.

AGB was estimated with the following equation (Cai et al., 2022): 5 $${AGB}_{i}=b\times \frac{{{B}_{\mathit{max}}}_{i}}{1+\left(\frac{{{B}_{\mathit{max}}}_{i}}{{B}_{{t}_{0}}}-1\right)\times \mathrm{exp}\left(-{{v}_{0}}_{i}\right)}$$ where $$b$$ is the coefficient converting biomass to carbon sequestration and set as 0.5 (Cai et al., 2022); $${B}_{{t}_{0}}$$ is the initial biomass (unit: Mg/ha) and set as 0.1 (Cai et al., 2022); $$t$$ denotes the accumulation period (here, $$t=21$$); $${{B}_{\mathit{max}}}_{i}$$ is the maximum biomass (unit: Mg/ha) and $${{v}_{0}}_{i}$$ is the maximum growth rate achievable under the optimal climatic conditions in unit $$i$$. See Equations S1–S4 in Supporting Information S1 for detailed AGB estimation.

Soil organic carbon was estimated as the balance between inputs from litter humification and losses due to humus mineralization (Cai et al., 2022): 6 $${SOC}_{i}=\frac{{I}_{i}}{{k}_{i}}-\left(\frac{{I}_{i}}{{k}_{i}}-{{S}_{{t}_{0}}}_{i}\right)\times \mathrm{exp}\left(-{k}_{i}\times t\right)$$ where $${{S}_{{t}_{0}}}_{i}$$ is the initial SOC density (unit: Mg/ha) and was retrieved from China National Soil Information Grids (Liu et al., 2022); $$t$$ is the humification and mineralization period (here, $$t=21$$); $${I}_{i}$$ is the input from the litter humification and $${k}_{i}$$ is the decomposition rate in spatial unit $$i$$ (unit: Mg/ha). See Equations S5–S12 in Supporting Information S1 for detailed SOC estimation.

Biodiversity Loss Mitigation

We estimated the restoration benefit for biodiversity as the reduction in potential extinction risks of forest-dependent species due to habitat expansion after the implementation of ANR. We downloaded spatial data of species distribution ranges from the IUCN Red List of Threatened Species (International Union for Conservation of Nature, 2023), BirdLife International, and Handbook of the Birds of the World (BirdLife International & NatureServe, 2018), and retained species whose habitat type was forest and distribution ranges were overlapped with mainland China (Strassburg et al., 2012). As a result, a total number of 500 mammals, 536 amphibians, and 994 birds were retained. The potential extinction risk for each species was estimated using the adapted species-area relationship as in Strassburg et al. (2020): 7 $${{extinction\,risk}_{i}}_{s}=1-{\left(\frac{{{current\,habitat}_{i}}_{s}}{{{original\,habitat}_{i}}_{s}}\right)}^{z}$$ where $${{extinction\,risk}_{i}}_{s}$$ represents the extinction risk for species $$s$$ in spatial unit $$i$$; $$z$$ is the slope of the species-area relationships, and was set at 0.25 following Strassburg et al. (2020); $${{original\,habitat}_{i}}_{s}$$ is the forest extent in 1995 in spatial unit $$i$$ that falls within the distribution range of species $$s$$; $${{current\,habitat}_{i}}_{s}$$ is the forest extent in 2015 in spatial unit $$i$$ that falls within the distribution range of species $$s$$. We estimated the reduction in extinction risk for species $$s$$ as the difference between extinction risks of current and ANR implementation scenarios (see next section), and then we summed it across all species to represent the potential benefit of spatial unit $$i$$ to the overall mitigation of biodiversity loss.

Spatial Optimization

We designed three scenarios separately for single-objective and multi-objective ANR optimization strategies. The three single-objective scenarios aimed to (a) minimize total cost, (b) maximize carbon sequestration, and (c) maximize reductions in potential extinction risks of forest-dependent species, respectively. For the multi-objective optimization strategy, the three scenarios were: (a) taking total cost, carbon sequestration, and biodiversity loss into account (Equations 8 and 9); (b) incorporating spatial constraint from the extent of current forest restoration projects into the first multi-objective scenario (Equations 8–10); and (c) further incorporating fragmentation mitigation into the second multi-objective scenario (Equations 8–11) (Deb et al., 2002; Lagro, 1991): 8 $$\left\{\begin{array}{@{}l@{}}\min \sum\limits _{i=1}^{N}{{x}_{i}\times total\,cost}_{i}\\ \max \sum\limits _{i=1}^{N}{{x}_{i}\times {\increment}extinction\,risk}_{i}\\ \max \sum\limits _{i=1}^{N}{{x}_{i}\times carbon\,sequestration}_{i}\end{array}\right.$$ 9 $$\text{subject}\,\text{to}\,\sum\limits _{i=1}^{N}{x}_{i}\le potential\,regeneration\,areas,i\in N$$ 10 $${y}_{i}\in forest\,projects$$ 11 $$\mathrm{m}ax\sum\limits _{i=1}^{N}\frac{{{n}_{{\Omega }}}_{i}}{{{\Omega }}_{i}},i\in N$$ where $${x}_{i}$$ (ha) is the area of spatial unit $$i$$ (here, $${x}_{i}=1$$). The objective functions (Equation 8) encompass $${total\,cost}_{i}$$, $${carbon\,sequestration}_{i}$$, and $${{\increment}extinction\,risk}_{i}$$ in spatial unit $$i$$, as determined in the previous steps; $$N$$ is the total number of implementation units. The area constraint in Equation 9 ensures the total area of ANR implementation does not exceed that of the potential regeneration areas. $${y}_{i}$$ is the location in spatial unit $$i$$. The spatial constraint in Equation 10 confines implementation within the boundaries of current forest restoration projects retrieved from the Resource and Environment Science and Data Center of the Chinese Academy of Sciences (2022). The spatial constraint in Equation 11 signifies implementation targeting fragmentation mitigation and habitat contiguity of forests. $${{\Omega }}_{i}$$ (ha) represents the total area of the 4-neighborhood around spatial unit $$i$$ (here, $${{\Omega }}_{i}=5$$), while $${{n}_{{\Omega }}}_{i}$$ (ha) is the total forest area of the same 4-neighborhood (Lagro, 1991). We performed spatial optimization using the NSGA-II algorithm with numpy 1.21.5, gdal 3.4.3, and scipy 1.10.1 packages in Python (Gu et al., 2023; Michalek et al., 2023; Rouault et al., 2022; Virtanen et al., 2020). The NSGA-II algorithm was selected for its ability to handle conflicting objectives through Pareto ranking, eliminating the need for weight assignment in traditional multi-objective optimization methods such as linear programming (Deb et al., 2002; Strassburg et al., 2019).

We compared the total costs and ecological benefits across different objective scenarios with those of the ANR4All scenario and a random selection scenario. The comparison focused on different priority levels, particularly the top 30% of priority areas, consistent with Global Biodiversity Framework Target 2 (Convention on Biological Diversity, 2022). In the random selection scenario, grid cells within potential regeneration areas were selected randomly, and the scenario performance was measured as the mean total cost and ecological benefits of 100 repeated selections. The random selection procedure was implemented using Python's random package.

Results

Model Validity

The random forest model incorporating both biophysical and socioeconomic variables achieved higher classification (AUC = 0.84) and prediction accuracy (81%; Figure S1 in Supporting Information S1) compared to the biophysical-only model (AUC = 0.82, Accuracy = 71%; Figure S2 in Supporting Information S1), and was therefore selected for estimating forest regeneration potential. During internal validation, this model exhibited robustness across different data splits (Standard error <0.0011; Figure S3 in Supporting Information S1) and consistent classification performance across China's geophysical divisions (AUC = 0.80 ∼ 0.89, Accuracy = 75 ∼ 93%), though slightly reduced performance was observed in arid and cold temperate zones (Figure S4 in Supporting Information S1). In precipitation-based divisions, the model performed best in the semi-humid zone (III, AUC = 0.89), followed by the semi-arid (II, AUC = 0.83), humid (IV, AUC = 0.82), and arid zones (I, AUC = 0.71). It showed high sensitivity for detecting the regenerated areas in the humid zone (SEN = 0.89) and specificity for identifying the non-regeneration areas in the arid zone (SPE = 0.99). Across temperature-based zones, the model performed best in Tibetan Plateau (⑥, AUC = 0.88), followed by tropical (⑤, AUC = 0.86), temperate (②, AUC = 0.82), warm temperate (③, AUC = 0.80), subtropical (④, AUC = 0.81), and cold temperate zones (①, AUC = 0.66). It excelled at identifying the regenerated areas in the cold temperate zone (SEN = 0.89) and non-regeneration areas in the tropical zone (SPE = 0.98). Furthermore, the model showed an acceptable consistency with the external data set from Cheng et al. (2024) (AUC = 0.76, Accuracy = 63%; Figure S5 in Supporting Information S1), and notably achieved a 98% success rate in predicting new natural forests from 1995 to 2015 (PPR = 0.98, Figure S6 in Supporting Information S1).

Spatial Patterns of Forest Regeneration Types

From 1995 to 2015, an estimated 3.40 million hectares of forest were regenerated from cropland and grassland in China, with cropland contributing 2.02 million ha (59.38%) and grassland 1.38 million ha (40.64%). The regenerated areas were primarily located in central and southern China, where climates were humid subtropical and semi-humid warm temperate (Figure 1). These zones contained 61.99% and 12.19% of the total regenerated areas, respectively. As for the non-regeneration areas, grasslands and cropland respectively occupied 256.47 million (61.16%) and 162.90 million ha (38.84%). These non-regeneration areas were mainly distributed in northwest China and the Tibetan Plateau. The humid subtropical and semi-humid warm temperate zones contained 17.13% and 12.02% of the total non-regeneration areas, respectively.

[IMAGE OMITTED. SEE PDF]

Driving Factors of Forest Regeneration Potential

Among the focal biophysical and socio-economic variables, six variables relevant to land use (i.e., proportions of cropland, forest, and grassland), climate (i.e., mean temperature and precipitation seasonality), and soil (i.e., soil cation exchange capacity) had variable importance values summing up to 71% in the random forest model (Figure 2). Conversely, the remaining variables such as demographic and social-economic drivers collectively contributed to a total of 29% in variable importance values. Specifically, mean temperature and annual precipitation showed positive correlations with forest regeneration potential, while the impervious proportion was negatively correlated (p < 0.001; Figure S7 in Supporting Information S1). Other variables exhibited non-monotonic relationships with forest regeneration potential. For example, proportions of cropland, forest, and grassland, as well as soil cation exchange capacity, bulk density, slope, and rural population density exerted inverted U-shaped influences on forest regeneration potential. In contrast, fixed-asset investment displayed a U-shaped relationship. The remaining variables showed non-monotonic relationships with multiple inflection points, suggesting more complex relationships with forest regeneration potential.

[IMAGE OMITTED. SEE PDF]

It was estimated that an area of 5.11 million ha of land within current cropland and grassland had a relatively high potential for forest regeneration, while regeneration had yet to occur in these areas over the past 20 years, indicating that human assistance might be needed to facilitate regeneration. If successfully transformed into forest via ANR by 2035, these areas could contribute to 12.41% of China's promised restoration area in the UNCCD. These high potential areas were likely to be distributed in central, southern, and northeastern China, where climates were humid subtropical, semi-humid temperate, and humid temperate (Figure 2). According to the precipitation-based geophysical division, the humid zone might have an area of 3.76 million ha of potential forest regeneration, accounting for 73.59% of the total potential areas. On the other side, the subtropics were likely to contain 58.90% of the total potential areas, covering an area of 3.01 million ha. Forest regeneration potential remained low in arid zones, especially in northwest China and the Tibetan Plateau.

Priority Area for Assisted Forest Regeneration

If the assisted natural regeneration (ANR) could be implemented across all potential regeneration areas (the ANR4All scenario), it was estimated that 100.26 gigatons of CO₂ emissions could be sequestered and potential extinction risks of forest-dependent species could be reduced. The total cost was projected to be US$656.06 billion, which included an opportunity cost of US$524.20 billion and an implementation cost of US$131.86 billion—19.78% lower than that of monoculture tree planting.

Among the three single-objective scenarios, the top 30% ANR priority areas under the cost-saving scenario were located in China's humid zones, especially the Greater Khingan Range in northeast China, eastern Inner Mongolia and southern Xinjiang in northwest China, and Yunnan in southwest China (Figure 3a). The total cost of ANR in these prioritized areas was estimated to be US$61.07 billion, namely only 9.31% of the ANR4All scenario. The implementation cost for these areas was US$30.49 billion, 38.18% lower than that of monoculture tree plantations. Compared to the random scenario, the cost-saving scenario reduced the total cost by 72.87%. Under the climate-mitigation scenario, the top 30% priority areas were concentrated in south and southwest China (Figure 3b), where climates were humid subtropical. The implementation of ANR within the top 30% priority areas under this scenario was predicted to sequester 60.64 gigatons of CO₂ emissions, accounting for 60.48% of the total sequestration from the ANR4All scenario. This scenario increased CO₂ sequestration by 53.52% compared to the random scenario. As far as the biodiversity-conservation scenario, the top 30% priority areas were primarily distributed in the Greater Khingan Range in northeast China and southern Xinjiang in northwest China (Figure 3c), and implementing ANR within the top 30% priority areas could reduce the potential extinction risks of forest-dependent species by 48.45%. Compared to the random scenario, the biodiversity-conservation scenario would decrease the potential extinction risks by 33.73%. These three scenarios had, however, very limited overlap in space and ecological benefits under the same priority scheme (Table 1; Figure 4). Under the top 30% priority scheme, the cost-saving scenario had only 14.86% of the prioritized areas overlapped with their counterparts under the climate-mitigation scenario, and 18.99% with their counterparts under the biodiversity-conservation scenario. In addition, the cost-saving scenario would sequester 25.09 gigatons of CO₂ and reduce the potential extinction risks by 30.68%, reaching only 63.52% of the sequestration and 84.68% of the reduction in the random scenario. The climate-mitigation scenario could reduce extinction risks by 20.71%, achieving only 57.16% of the reduction in the random scenario. Meanwhile, the biodiversity-conservation scenario would contribute just 25.05% of the CO₂ sequestration, reaching only 63.57% of the sequestration in the random scenario as well.

[IMAGE OMITTED. SEE PDF]

Table 1 Costs and Restoration Outcomes of Assisted Regeneration Under the Top 30% Priority Scheme in Different Optimization Scenarios

Optimization scenario	Total cost	Carbon sequestration	Extinction risk
Value (US$ billion)	Proportion (%)	Value (gigaton CO2)	Proportion (%)	Value (%)	Proportion (%)
Single objective
Cost savings	61.07	9.31	25.09	25.03	−30.68	30.68
Climate mitigation	159.80	24.36	60.64	60.48	−20.71	20.71
Biodiversity conservation	203.57	31.03	25.11	25.05	−48.45	48.45
Multiple objectives
Three objectives combined	215.91	32.91	46.54	46.42	−40.64	40.64
Three objectives combined, subject to the current project extent	180.76	27.55	37.60	37.50	−34.96	34.96
Three objectives combined and Fragmentation mitigation, subject to the current project extent	183.74	28.01	30.46	30.38	−40.22	40.22
Random selection	225.08	34.31	39.50	39.39	−36.23	36.23

[IMAGE OMITTED. SEE PDF]

In contrast, the multi-objective optimization strategy could simultaneously enhance benefits to both climate and biodiversity whilst lowering the total cost (Figure 3d). Under the top 30% priority scheme, prioritized areas were mainly located in the Greater Khingan Range in northeast China, southern Xinjiang in northwest China, and Yunnan in southwest China. This priority scheme could sequester 46.54 gigatons of CO₂ emissions (46.42% of the total sequestration from the ANR4All scenario) and reduce extinction risks by 40.64%, with a total cost of US$215.91 billion (32.91% of the ANR4All scenario), of which the implementation cost was projected to be US$33.76 billion—31.55% lower than that of monoculture tree plantations. Compared to the random scenario, the multi-objective optimization strategy could reduce the total costs by 4.07%, increase CO₂ sequestration by 17.82%, and decrease the potential extinction risks by 12.17%, simultaneously. When performing multi-objective optimization subject to the extent of current forest restoration projects (Figure 3e), the top 30% prioritization areas were predicted to be located in the Three Shelterbelt Initiatives along the Liaohe River in northeast China, the Three-North Forestation Project in northwest China, and the upper Yangtze River Basin in southwest China. This prioritization scheme was predicted to sequester 37.60 gigatons of CO₂ emissions (37.50% of the ANR4All scenario) and reduce potential extinction risks by 34.96% at a total cost of US$180.76 billion (27.55% of ANR4All), while the implementation cost in these prioritization areas could decrease by 30.50% from US$38.86 to US$27.01 billion compared to that of monoculture tree plantations. Lastly, the multi-objective restoration with fragmentation mitigation scenario (Figure 3f) was effective in reducing potential extinction risks of forest-dependent species by 40.22%, though it would sequester less CO₂ emissions—30.46 gigatons (30.38% compared with ANR4All). The total cost would be higher than that of the previous strategies, reaching US$183.74 billion (28.01% of the ANR4All scenario), with an implementation cost of US$27.71 billion (28.71% lower than that of monoculture tree plantations).

Discussion

Global deforestation continues to drive both climate change and biodiversity loss, compelling political leaders around the world to commit to ambitious forest restoration goals. Our study examined the potential of assisted natural regeneration (ANR)—a complementary approach to monoculture tree plantations—in China's reforestation efforts. Aligned with China's commitment to achieving 26% forest coverage by 2035, we developed a spatial ANR prioritization framework to balance the goals of climate change mitigation, biodiversity conservation, and cost efficiency. This framework outlined a national ANR roadmap to facilitate China's transition from extensive plantations toward integrated restoration strategies that leverage ecosystem self-recovery. We identified 5.11 million hectares of cropland and grassland with forest regeneration potential, especially in humid zones. The area represents 12.41% of China's restoration commitments at the UNCCD (United Nations Environment Programme, 2021). We also found that the multi-objective optimization strategy, which systematically accounted for biodiversity, climate, and cost, emerged as the most effective solution to maximize ecological benefits. Prioritizing the top 30% of areas under this strategy could sequester 46.54 gigatons of CO₂ emissions, reduce extinction risks of forest-dependent species by 40.64%, and lower the implementation cost by 31.55% than monoculture tree plantations. Compared to the random scenario, the multi-objective optimization strategy could reduce the total costs by 4.07%, increase CO₂ sequestration by 17.82%, and decrease extinction risks by 12.17%.

Our findings underscore the importance of climate, soil, and landscape patterns as primary drivers of forest regeneration in China. The humid zones were key regenerated areas, such as the Tsinling Mountains, the upper Yangtze River Basin, and the Greater Khingan Range (Xu et al., 2023). These areas mainly benefit from annual precipitation of over 800 mm which is sufficient to support biomass accumulation (Poorter et al., 2016), and from vast and continuous forests that can serve as seed sources for native vegetation (Lu et al., 2018). Conversely, arid regions like northwest China and the Tibetan Plateau exhibited limited regeneration potential due to harsh climatic and soil conditions. Afforestation in these regions often faces challenges such as groundwater depletion and dependence on irrigation (Zhang et al., 2022). To successfully implement ANR in these arid and semi-arid zones, particularly within the Three-North Shelterbelt Forest, substantial water or nutrient inputs would be required (Xu et al., 2023).

Comparisons across optimization strategies revealed trade-offs in ANR benefits. For instance, under the cost-saving scenario, ANR in the top 30% priority areas could lower implementation costs by 38.18% compared to monoculture tree plantations. However, this scenario provided limited co-benefits, sequestrating 25.09 gigatons of CO₂ and reducing extinction risks by 30.68%, which was equivalent to 63.52% and 84.68% of the random scenario. Meanwhile, prioritized areas also varied significantly across the three single-objective scenarios, with spatial overlaps ranging from 14.86% to 18.99%. In contrast, the multi-objective optimization strategy offered balanced outcomes, achieving 46.54 gigatons of CO₂ sequestration and reducing extinction risks of forest-dependent species by 40.64% with a total cost of US$215.91 billion under the top 30% priority scheme. Compared to the random scenario, the multi-objective optimization strategy could reduce the total costs by 4.07%, increase CO₂ sequestration by 17.82%, and decrease the potential extinction risks by 12.17%. When applied within the extent of current forest restoration projects, this strategy could reduce extinction risks by 34.96% and achieve 37.60 gigatons of CO₂ sequestration with a total cost of US$180.76 billion. Further incorporating fragmentation mitigation, this strategy would reduce extinction risks by 40.22%, though CO₂ sequestration would be lower (30.46 gigatons) and the total cost would be higher (US$183.74 billion). These results align with the general findings from multi-objective ecosystem restoration planning, where cost-saving strategies perform poorly in terms of mitigating climate change and biodiversity loss, with trade-offs between carbon sequestration and biodiversity conservation (International Institute for Sustainability Australia, 2021; Strassburg et al., 2020). The multi-objective optimization strategy outperforms the random scenario, offering a more balanced approach (International Institute for Sustainability Australia, 2021; Liévano-Latorre et al., 2025; Strassburg et al., 2020). These results highlight the potential of ANR to address multiple sustainable development goals simultaneously (Chausson et al., 2020; Griscom et al., 2017). We recommend implementing ANR in priority areas that can balance these co-benefits, such as in the Greater Khingan Range, southern Xinjiang, and Yunnan. These efforts could complement the Shelterbelt Forest along the Liaohe River, the Three-North Forest, and the Upper Yangtze River, thereby reinforcing China's ecological barriers cost-effectively (Lu et al., 2022; Xu et al., 2023).

It is important to note that the climate effect of assisted natural regeneration depends on how radiant and turbulent energy fluxes over these forests modify surface temperature (Peng et al., 2014). While forest regeneration often promotes daytime cooling through evapotranspiration, it may lead to nighttime warming due to reduced atmospheric turbulence (Lee et al., 2011). In arid regions, for instance, reduced albedo may offset the climate benefits of carbon sequestration (Hasler et al., 2024). Indeed, earlier research has reported that tree plantations resulted in a net warming effect in drier regions of China where annual precipitations were less than 600 mm (Peng et al., 2014). Nevertheless, our analysis indicates that key regeneration priority areas are primarily located in humid regions where these trade-offs are less pronounced (Hasler et al., 2024). It is also noteworthy that potential ANR areas may experience cyclic conversions between forests, croplands, and grasslands, which will influence albedo and overall climatic impacts (Shen et al., 2019). Therefore, successful ANR requires context-specific planning and implementation at the local scale, considering multiple impacts and their potential trade-offs (Xu et al., 2023).

Extensive marginal farmland in China, often abandoned due to low productivity, presents a promising opportunity for ANR implementation (Kuang et al., 2022; Shono et al., 2020). Rural labor loss has also led to widespread agricultural abandonment, leaving the most remote villages in China at risk of vanishing (Liu & Li, 2017). Meanwhile, the inappropriate implementation of the “Requisition-Compensation Balance of Farmland” policy has further accelerated the expansion of marginal farmland, lowering crop yields (Kuang et al., 2022). Fortunately, policy shifts emphasizing ecosystem self-recovery align well with ANR principles, providing an avenue for restoring abandoned farmland and degraded forests. For example, nature-based restoration is now a key principle of the Overall Plan for Major Projects for the Protection and Restoration of Important National Ecosystems (2021–2035) in China and has been applied to restore many key ecological areas such as the Three-North Forest and national nature reserves. However, effective implementation of ANR requires stakeholder engagement, particularly among local farmers in priority areas (Holl & Brancalion, 2020). Lessons from initiatives such as REDD + underscore the need for social safeguards for Indigenous peoples (Barletti & Larson, 2017), such as those who lack secure land tenure and face elevated risks of displacement (Liévano-Latorre et al., 2025; Ribot, 2014). To mitigate such risks, measures such as improved subsidies and job guarantees (e.g., integrating ANR with enrichment planting and agroforestry) are crucial to protect the interests of the affected farmers (Shono et al., 2020; Wang et al., 2023). Furthermore, ANR could also contribute to the restoration of degraded forests in addition to forest land previously converted to cropland or grassland (Shono et al., 2020; Williams et al., 2024). It also provides an opportunity to facilitate the regeneration of native vegetation in exotic and native tree plantations, thereby enhancing the ecological value of large-scale reforestation initiatives from previous decades (Williams et al., 2024). In practice, this approach has been successfully applied in various local contexts following agricultural use and deforestation, and has the potential to achieve considerable carbon sequestration and biodiversity conservation if scaled up in China (Shono et al., 2020; Williams et al., 2024).

Conclusions

This study highlighted the significance of nature-based climate solutions in operationalizing global initiatives such as the UN Decade on Ecosystem Restoration. Our methodological framework for ANR potential assessment and spatial prioritization identified the multi-objective optimization as the most effective strategy balancing ecological benefits and cost efficiency. The top 30% of priority areas, concentrated in the Greater Khingan Range, southern Xinjiang, and Yunnan, were predicted to sequester 46.54 gigatons of CO₂ emissions and reduce extinction risks of forest-dependent species by 40.64%, at a total cost of US$215.91 billion. Our findings confirm the feasibility and sustainability of ANR as a critical component of large-scale reforestation efforts. We advocate for integrating ANR into national restoration strategies to enhance climate resilience, conserve biodiversity, and promote cost-efficient ecological recovery.

Acknowledgments

This research was supported by the National Natural Science Foundation of China (42171414) and the Fundamental Research Funds for the Central Universities (2042024kf0029).

Data Availability Statement

Data used to identify forest regeneration types are available from Yang and Huang (2021), Bunting et al. (2022), and Lesiv et al. (2021). Environmental condition predictors are derived from Liu et al. (2022), Peng et al. (2019), Fick and Hijmans (2017), and the Earth Resources Observation and Science Center (2018). Socio-economic predictors are obtained from Chen and Gao (2021), Kummu et al. (2017), Xu (2023), and the National Bureau of Statistics of China (2017a, 2017b, 2017c). Landscape pattern predictors are obtained from supplementary information files of Curtis et al. (2018), Hansen et al. (2013), and Tyukavina et al. (2022), along with data sets from Yang and Huang (2021), Lesiv et al. (2021), Gilbert et al. (2018), Yu et al. (2020), the National Forestry and Grassland Administration of China (2017), and the Resource and Environment Science and Data Center of the Chinese Academy of Sciences (2020, 2022). Data for cost assessment are available from Gilbert et al. (2018), Yu et al. (2020), the National Bureau of Statistics of China (2017a), and the National Forestry and Grassland Administration of China (2017). Data for evaluating carbon sequestration are from Liu et al. (2022), Lu et al. (2021), and Peng et al. (2019). Data sets for estimating extinction risk are provided by the International Union for Conservation of Nature (2023) and BirdLife International & NatureServe (2018). Validation data for the random forest model can be accessed from the Cheng et al. (2024). The statistical yearbooks and administrative boundary data are currently available only in Chinese and can be accessed via the China National Knowledge Infrastructure (National Bureau of Statistics of China, 2017a, 2017b, 2017c; National Forestry and Grassland Administration of China, 2017) and the Resource and Environment Science and Data Center of the Chinese Academy of Sciences (Xu, 2023), with user registration and fee required. Software used includes Python packages scipy 1.10.1 (Virtanen et al., 2020), scikit-learn 1.4.2 (Pedregosa et al., 2011), gdal 3.4.3 (Rouault et al., 2022), and the packages installed in ArcGIS Pro 3.0.1 (Environmental Systems Research Institute, 2020), as well as Office 2021 (Microsoft, 2021) and ChiPlot (available at ).

References

Alves, J., Oliveira, M., Chazdon, R., Calmon, M., Pinto, A., Darvin, E., & Pereira, B. (2022). The role of assisted natural regeneration in accelerating forest and landscape restoration: Practical experiences from the field. Retrieved from https://www.wri.org/research/assisted‐natural‐regeneration‐case‐studies

Araújo, M. B., Pearson, R. G., Thuiller, W., & Erhard, M. (2005). Validation of species–climate impact models under climate change. Global Change Biology, 11(9), 1504–1513. [DOI: https://dx.doi.org/10.1111/j.1365-2486.2005.01000.x]