New forest biomass carbon stock estimates in Northeast Asia based on multisource data

Forests play an important role in both regional and global C cycles. However, the spatial patterns of biomass C density and underlying factors in Northeast Asia remain unclear. Here, we characterized spatial patterns and important drivers of biomass C density for Northeast Asia, based on multisource data from in situ forest inventories, as well as remote sensing, bioclimatic, topographic, and human footprint data. We derived, for the first time, high‐resolution (1 km × 1 km) maps of the current and future forest biomass C density for this region. Based on these maps, we estimated that current biomass C stock in northeastern China, the Democratic People's Republic of Korea, and Republic of Korea to be 2.53, 0.40, and 0.35 Pg C, respectively. Biomass C stock in Northeast Asia has increased by 20%–46% over the past 20 years, of which 40%–76% was contributed by planted forests. We estimated the biomass C stock in 2080 to be 6.13 and 6.50 Pg C under RCP4.5 and RCP8.5 scenarios, respectively, which exceeded the present region‐wide C stock value by 2.85–3.22 Pg C, and were 8%–14% higher than the baseline C stock value (5.70 Pg C). The spatial patterns of biomass C densities were found to vary greatly across the Northeast Asia, and largely decided by mean diameter at breast height, dominant height, elevation, and human footprint. Our results suggest that reforestation and forest conservation in Northeast Asia have effectively expanded the size of the carbon sink in the region, and sustainable forest management practices such as precision forestry and close forest monitoring for fire and insect outbreaks would be important to maintain and improve this critical carbon sink for Northeast Asia.


| INTRODUC TI ON
Terrestrial ecosystems play a vital role in global carbon (C) cycles and the mitigation of global warming (Bonan, 2008;Piao et al., 2009). As the largest C reservoir of these terrestrial ecosystems, forest ecosystem comprises more than 80% and 40% of global terrestrial C pool above-and below-ground, respectively (Dixon et al., 1994;Pan et al., 2011). However, it has been observed that climate change and forest disturbances (logging, fires, and insects) are directly or indirectly converting forest ecosystems from net fixers to net sources of C to the atmosphere (Alam et al., 2012;Pugh et al., 2019;Stinson et al., 2011), causing substantial changes in forest age, structure, composition, and biomass across the world (Alexander et al., 2012;Zhang & Liang, 2014). Meanwhile, massive afforestation and forest restoration efforts have been conducted across the world over the past decades, especially in Northeast Asia, where 4.02 × 10 4 km 2 of new forests have been planted and maintained in China and Republic of Korea since year 2000 , but associated effects on global C cycles remain unknown. Therefore, accurate estimates of C stock under climate change has become increasingly important for scientific exploration of the Earth systems, as well as biological conservation and natural resource management.
Precision in estimating forest biomass C stock depends to a large extent on the modeling approaches and algorithms. The ordinary least squares (OLS) method is traditionally the most frequently used model in ecological studies and serves as a benchmark for the other model types (Dube & Mutanga, 2015;Lu, 2006). However, the underlying assumptions of OLS are not always met for multisource data, which can be highly nonlinear and often correlated with each other. To this end, the nonparametric and machine learning algorithms, such as random forests (RF), artificial neural network (ANN), support vector machine (SVM), and extreme gradient boosting (XGBoost) methods are developed to handle multisource data from forest ecosystems (Ahmed et al., 2015;Gao et al., 2018;Liang et al., 2016;Steidinger et al., 2019;Zhang et al., 2018). The RF regression is especially effective in fitting data through a set of decision tree models (Hong et al., 2019), with an advantage of processing large amounts of data with computational efficiency, and ranking explanatory variables by the contribution to the goodnessof-fit (Belgiu & Drăguţ, 2016;Breiman, 2001;Cracknell & Reading, 2014;Houghton et al., 2007). Meanwhile, the XGBoost model has exhibited better control against overfitting compared to prior gradient boosting algorithms (Chen & Guestrin, 2016), and is less demanding on computing resources than most of the other machine learning models (Torlay et al., 2017). Both RF and XGBoost have proven performance in forest C stock estimation (Carreiras et al., 2012;Tang et al., 2018).
Northeast Asia encompasses the Korean Peninsula and northeastern China. Characteristic of temperate climate conditions, this region is known for highly diverse (Qian & Ricklefs, 2000) and productive (FAO, 2010) forests. Since the end of the 1970s, Northeast Asia has seen an increased forest area and stand volume due to active reforestation and conservation undertakings, but little is known about forest biomass C stock in this region. Throughout the last decade, some local studies have been conducted on forest biomass and C stock in northeastern China Wei et al., 2013Wei et al., , 2014Zhang & Liang, 2014;Zhang et al., 2013), and the Republic of Korea (ROK; Choi & Chang, 2004;Fang et al., 2014;Lee et al., 2014;Li et al., 2010). However, little is known regarding the forest biomass C stock and spatial variations in the Democratic People's Republic of Korea (DPRK; Fang et al., 2014;Thurner et al., 2014).
The main objectives of this study were (a) to accurately estimate the current biomass C stock for the mixed temperate forests in Northeast Asia; (b) to understand the spatial patterns and important drivers of Northeast Asia's mixed temperate forest biomass C density; and (c) to project biomass carbon density and carbon stock of Northeast Asia in 2080 under climate change. To achieve these objectives, we compared different machine learning and statistical models based on multiple remote sensing and ground-based datasets across the region to map spatial patterns of biomass C density and their potential drivers in Northeast Asia.

| Study area
We studied forested areas of Northeast Asia, located at the eastern margin of the Eurasian continent between the Da Xing'an Ling Mountains to the west and the East Sea to the east, the Amur River to the North and the Jeju Island to the South (Figure 1). The region covers approximately 1.74 × 10 8 ha of land area, of which 42% (5.39 × 10 7 ha) is covered by forests, accounting for approx. 0.01% of global forest area (FAO, 2010). Influenced by high latitude East Asian monsoons, the regional climate varies from warm temperate subregions in the south to cool temperate subregions in the north, and ranges from humid and semi-humid in the east to semi-arid in the west. The annual mean temperature ranges between −6.3 and 15.4°C and the annual precipitation ranges from 250 to 2,084 mm, as detailed in Table 1.

| Northeastern China Forest Inventory Network (FIN) data
An extensive forest inventory network (FIN) of northeastern China has been established across the entire mixed temperate forest region since 2017. Based on a systematic sampling scheme, FIN sample plots are evenly distributed with an approximately 30 km distance between any two nearest plots. Therefore, each FIN sample plot represents 30 km × 30 km (900 km 2 ) of forest area, with an exception that for a total of 50 plots, of which original locations fall outside the natural forest, alternative distances ranging from 20 to 60 km (with a mean of 30 km) have been applied so that actual plot locations are within a natural forest area.
The FIN is comprised of 456 permanent sample plots, each 0.1 ha in size, with a 17.85 m radius. For consistency with the Republic of Korea (ROK) plot size, we divided each 0.1 ha sampling plot into two non-overlapping semi-circle plots 0.04 ha in size (Otypková & Chytry, 2006). In total, we derived 912 0.04 ha permanent sample plots for northeastern China. Within each plot, all free-standing woody stems with a diameter at breast height (dbh) greater than 6 cm were geo-referenced, tagged, and recorded by species name, dbh, total height, crown width, and crown length. In addition, the elevation, slope, aspect, and soil depth of each plot were also recorded.

TA B L E 1 (Continued)
Mongolica, along with other deciduous species in the southern region ( Figure 1).

| ROK national forest inventory data
The latest ROK national forest inventory (NFI) data used for this study are based on a systematic cluster sampling design for surveys at an interval of 4 km along the longitude and latitude, and 1 or 2 km along the longitude and latitude for small forested areas.

| Predictor datasets
In addition to in situ forest inventory data, we also compiled 43 ecosystem, remote sensing and environmental variables as candidate predictors of forest biomass C density. These covariates were derived from published digital geospatial maps and ground-based survey data, and can be grouped into five categories: ecosystem (4 variables), remote sensing (14 variables), bioclimatic (21 variables), anthropogenic (4 variables), and geographic coordinates (longitude and latitude in the WGS84 datum). Among the aforementioned data, 39 covariates derived from raster layers share a common 1 km 2 native spatial resolution (Table 1).   (Simard et al., 2011). The forest canopy height data have been widely used for forest biomass mapping (Su et al., 2016;Zhang et al., 2019; Table 1).
We adopted the most current fourth version (Jarvis et al., 2008) of digital elevation model (DEM) from the Shuttle Radar Topography Mission (SRTM). The DEM data were available on GEE with a spatial resolution of 90 m and was resampled to a spatial resolution of 1 km (Table 1).
The four anthropogenic predictors related to the human footprint indices and roadless areas were derived from human footprint database (Venter et al., 2016) and the global roadless database (Ibisch et al., 2016) at a 1 km × 1 km spatial resolution, respectively ( Table 1).
All of the geospatial covariates were pre-processed using

| Data on planted forests and climate change
Planted forests have a great potential to contribute significantly to climate change mitigation, and can also provide future wood and energy supply as well as a range of wide social and environmental benefits in terms of ecosystem services. According to Bastin et  to high greenhouse gas concentration levels (Riahi et al., 2007). RCP4.5 represents a stabilization scenario in which total radiative forcing is stabilized shortly after 2100, without overshooting the long-run radiative forcing target level (Thomson et al., 2011). In addition, our baseline scenario assumed constant climate conditions from now till 2080.
We selected five CMIP5 models to represent the decent amount of uncertainty in climate model projections ( (Table S2).

| Community-level biotic attributes
Based on tree-level in situ data, we derived five key community-level (c) total woody biomass density (W); (d) total woody biomass C density (C d ); and (e) total woody biomass C stock (C s ) (Tables 1 and 2).
The total biomass of each individual tree, the sum of stem biomass, branch biomass, root biomass, and foliage biomass in kg, was estimated from its dbh and height using species-specific allometric equations (available for approximately 80% of the studied tree species), and a generic allometric equation for the remaining 20% species (Dong et al., 2014(Dong et al., , 2015Son et al., 2014;Wang, 2006; Table S3): where TB (kg) is the total biomass per tree; B (cm) indicates the dbh; H (m) is the tree height; and a, b, and c represent parameters from the species-specific or generic allometric equation (Table S3).
The total biomass of a plot (Mg/ha) was derived as the sum of biomass of all the individual trees on that plot, standardized to a one-ha basis: where W (Mg/ha) denotes the total plot biomass per unit area; TB i (Mg) is the biomass for the ith tree on the plot; N indicates the number of trees with dbh ≥ 6 cm; and A (ha) represents the plot area (0.04 ha).
It is known that plant C content tends to vary between leaf types and biomass (Thomas & Martin, 2012). However, variations between plant tissues are relatively minor. These factors were taken into account in this study as follows: where TC (kg) represents the total C stock for a tree; a is 0.488 for broadleaf and 0.508 for conifer; and TB (kg) is the total biomass of that tree (Thurner et al., 2014). The total biomass C density per unit area (C d , Mg/ha) can then be calculated as: where TC i (Mg) indicates the C stock for the ith tree on the plot, and A is the plot area (0.04 ha). The total biomass C stock (C s , Pg C) in the study area was calculated as the product of the total C density per unit area (C d ) and the total forest area (FA).

| Model calibration, evaluation, and comparison
For imputing the total forest biomass C density per unit area from the sample points to the entire study region, we compared three commonly imputation models, namely OLS, RF, and extreme gradient boosting (XGBoost). The first is a statistical regression model, whereas the latter two are machine learning models. ( TA B L E 2 Forested area, biomass C density, C stock, and C stock Percent for each region in Northeast Asia for the period ranging from 2016 to 2017. NF: Natural forest; PF: Plantation forest. C stock level: high (>85 Mg C/ha), medium (55-85 Mg C/ha), low (<55 Mg C/ha) There is no planted forest data from DPRK.

| Ordinary least squares
OLS is a common statistical regression model which assesses the relationship between response variables and sets of explanatory variables by the principle of least squares (Goldberger, 1964): where y can be one of the three response variables: the total forest biomass C density per unit area (C d ), dominant height (DH), and mean dbh (D); x 1 , x 2 , …, x n are the predictor variables; a 0 is a constant; a 1 , a 2 , …, a n represent the regression coefficients associated with the response variables; n denotes the number of the predictor variables; and ε is the random error term with homoscedasticity and no autocorrelation.

| Random forests (RF)
RF regression applies the general technique of bootstrap aggregating (bagging) with a modified tree learning algorithm that selects, at each candidate split in the learning process, a random subset of the features (i.e., feature bagging) (Breiman, 2001).
Compared with statistical regression and other machine learning models such as ANN, SVM, and k-NN, RF model is less prone to the negative effects of overfit (Cracknell & Reading, 2014;Wang et al., 2015) and multicollinearity (Toloşi & Lengauer, 2011), and often has greater accuracy and better tolerance to noise and outliers in training data (Martens et al., 2007;Yaseen et al., 2019).
Therefore, RF has been used in processing high-dimensional datasets in ecological and forestry studies (Belgiu & Drăguţ, 2016;Liang et al., 2016;Steidinger et al., 2019). In this study, we derived training dataset using bootstrapping (drawing a random sample with replacement) from the observations, and used the remaining observation data for estimating the out-of-bag (OOB) errors. We fine-tuned the RF model for the following two hyper-parameters: the number of decision trees (ntree) and the minimum number of observations per tree leaf (mtry), while the minimum size of the terminal nodes (nodesize) was assigned with a default value of 5. We sought for the combination of the mtry and ntree values which resulted in the least root mean square error (RMSE) for OOB.
In the present study, importance of the variables was assessed using a permutation-based approach based on the increase in node purity. The ranking of variable importance from a permutation was repeated 20 times to generate the mean and variance of variable importance values (Strobl et al., 2007). Then, to better understand the biological drivers of C density and how they vary in the study area, we analyzed the partial dependence of C density across the five ecoregions of Northeast Asia (Olson et al., 2001). All of the random forest analyses were conducted in this study using the randomForest package (Liaw & Wiener, 2002) in R (version 3.5.1).

| Extreme gradient boosting (XGBoost)
XGBoost is a scalable machine learning system (Chen & Guestrin, 2016) that implements the gradient boosting decision tree algorithm highly efficient especially for high-quantity and high-dimension datasets (Nielsen, 2016). XGBoost estimates a function which projects a set of predictor variables into an output variable by minimizing a specified loss function, based on which it fits the regression tree model for the training data iteratively and merges the predictions from all iterations to obtain the prediction outcome via multiplying by the weights of the regression tree models (also known as learning rates). We examined the following hyper-parameters

| Model comparison and validation
We compared the foregoing models using a 90-10 cross-validation method (Stone, 1974 Then, based on the results, the best combination of hyper-parameters was selected for each machine learning model, based on which the best overall imputation model was selected.

| Imputation and mapping
For mapping current and future (2080 Baseline, RCP4.5, and RCP8.5) tree biomass C density (C d ) of the temperate forests across Northeast Asia, the entire study region was divided into a grid of 1 km pixels, each assumed to represent a relatively homogeneous landscape (Liang, 2012). Although this study's imputation models could technically be applicable to smaller pixels, the mapping was (6) y = a 0 + a 1 ⋅ x 1 + a 2 ⋅ x 2 +…+a n ⋅ x n + , done on the basis of 1 km pixels to match the finest resolution of the environmental predictors, as shown in Table 1.
From the 1,369 FIN and NFI permanent sample plots, we extrapolated current and future (2080 Baseline, RCP4.5, and RCP8.5) total tree biomass C density per ha (C d ) values to all the 1 km mapping pixels using the best imputation model, using observed point current and future (2080 Baseline, RCP4.5, and RCP8.5) tree biomass C density (C d ) data, as well as the values of the ecosystem, remote sensing, and environmental variables predictors extracted to each pixel.
We estimated the current and future (2080 Baseline, RCP4.5, and RCP8.5) total tree biomass C density per ha (C d ) pixel values at a 1 km resolution in Northeast Asia, and calculated the mean value and standard deviation of current and future (2080 Baseline, RCP4.5, and RCP8.5) tree biomass C density (C d ) for the entire Northeast Asia, northeastern China, DPRK, and ROK, respectively. Meanwhile, each pixel value represented a 1 km 2 forest area. Combined the aforementioned mean value and standard deviation of total tree biomass C density (C d ), we calculated the current and future (2080 Baseline, RCP4.5, and RCP8.5) total biomass C stock (C s ) in the entire Northeast Asia (Tables 2 and 4).

| Projecting future forest biomass carbon stock under climate change
To address the effect of climate change, we estimated future total biomass C density (C d ) and C stock (C s ) in 2080 under three climate change scenarios (Baseline, RCP4.5 and RCP8.5, see Section 2.2.4). We first estimated plot-level total stand volume in 2080 using climatesensitive stand forest growth models developed for the five main for-  a 0 , a 1 , a 2 , a 3 , b 0 , b 1 , b 2 , and b 3 represent the estimated coefficients.
We then estimated total biomass C density (C d , Mg/ha) in 2080 using the biomass expansion factor (BEF) method (Brown, 2002;Fang et al., 2001): where the coefficient r is 0.488 for broadleaf carbon content and 0.508 for conifer carbon content; V (m 3 /ha) is the stand volume. BEF = a + b/V represents the ratio of stand biomass to stock volume, with values provided by Fang et al. (2001).

| Model comparison and validation
Cross-validation shows that the RF regression model in general had higher goodness-of-fit and lower prediction errors than the other two candidate models (XGBoost and OLS). In terms of coefficient of determination (R 2 ), the RF regression model accounted for 60%, 86%, and 59% of the variance in the biomass C density (C d ), dominant height (DH), and mean dbh (D), respectively. These R 2 values were 3.4%, 3.6%, and 15.7% higher than those of the XGBoost, and 5.3%, 43.3%, and 353.8% higher than those of the OLS, respectively ( Figure 2). In terms of the prediction errors, the RF regression model had RMSE values of 18.2, 1.9, and 2.9 for estimating C d , DH, and D, respectively, which were 1.67%, 2.3%, and 4.0% lower than those of the XGBoost, and 3.5%, 39.5%, and 30.3% lower than those of the OLS, respectively. Similarly, the MAE values for the RF regression model were 2.7%, 12.7%, and 9.4% lower than those of the XGBoost, and 3.6%, 40.6%, and 29.7% lower than those of the OLS model, respectively ( Figure 2; see Figure S1 for regression models of biomass density).

| Current forest biomass C stock
For the entire Northeast Asia, we estimated the mean biomass C density (C d ) to be approximately 60.77 ± 14.63 Mg C/ha (mean ± standard deviation). For planted forest (PF), C d was estimated to be 59.27 ± 8.60 Mg C/ha, whereas for natural forest (NF), C d was 60.97 ± 14.92 Mg C/ha (Table 2). Among the three regions examined in this study (northeastern China, DPRK, and ROK), the highest C d was in DPRK (63.71 ± 14.44 Mg C/ha), whereas the lowest C d was in ROK (54.19 ± 6.01 Mg C/ha).
The total biomass C stock (C s ) across the Northeast Asia region was estimated to be 3.28 ± 0.76 Pg C, with 53.89 million ha of total forested area. Planted forests (PF) in Northeast Asia accounted for 6.98 million ha in area (13.0%) and 0.41 ± 0.06 Pg C in biomass C stock (12.5%) while natural forests (NF) accounted for 46.91 million ha (87.0%) in area and 2.86 ± 0.70 Pg C in biomass C stock (87.5%).
Across the region, northeastern China accounted for the largest carbon pool, with 2.53 ± 0.63 Pg C for a total forest area measuring 41.24 million ha. In northeastern China, PF stored 0.34 ± 0.08 Pg C (13.4%), whereas NF stored 2.86 ± 0.70 Pg C (86.6%). DPRK (0.40 ± 0.09 Pg C) and ROK (0.35 ± 0.04 Pg C) were found to have similar total biomass C stock (Table 2).

| Future forest biomass C stock
For Northeast Asia, we estimated the mean biomass C density (C d )

| Spatial patterns of the current forest biomass C densities
Forest biomass C densities (C d ) were found to vary greatly across the Northeast Asia (Figure 3

| Spatial patterns of the future forest biomass C densities
For the entire Northeast Asia, we mapped spatial patterns of biomass C densities (C d ) in 2080 under three climate change scenarios (Baseline, RCP4.5, and RCP8.5). Our results show that forest biomass C densities (C d ) varied greatly across the region (Figure 6).

| Important predictors of the current forest biomass C density
We ranked all the predictor variables for the current biomass C density (C d ) according to the variable importance values derived from the RF. Besides the most closely associated mean dbh (D, 32%) and dominant height (DH, 13%), elevation (Elev) was the next most important predictor with a standardized relative importance score of 5%, followed by the human footprint (H 1 , 4%), latitude in WGS84 datum (LAT, 3%), and annual mean temperature (T 1 ) (Figure 7).
To better understand how these predictors may influence current biomass C density (C d ), we analyzed the partial dependence of current biomass C density (C d ) on each predictor variable from four most important predictors, that is, mean dbh (D), dominant height (DH), elevation (Elev), and the human footprint (H 1 ). We further compared  Figure 1). Our results show that forest biomass C density (C d ) was positively associated with all four variables in general. More specifically, the association between C d and D was smoother for Eco3, Eco4, and Eco5, whereas the association was sigmoidal (close-to-flat on both ends with a steep incline in the middle, Figure 8) for Eco1 and Eco2. Similarly, the association between C d and DH was smoother for Eco5, whereas sigmoidal curves were observed for all other ecoregions. C d increased sharply with elevation (Elev) in low-altitude areas of Eco1 and Eco2, but this positive effect leveled off after elevation reached 250-500 m. For other ecoregions, C d increased sharply in mid elevation range but remained relatively stable in low and high altitudes. Finally, in areas where human footprint (H 1 ) was low, C d increased with H 1 in all ecoregions but Eco4 and Eco5. In Eco4, C d first declined with H 1 before a sharp increase, whereas, in Eco5, C d was largely independent from H 1 (Figure 8). We further found that both elevation (Elev) and human footprint (H 1 ) had a positive effect on mean dbh (D) (Figure 9).

| D ISCUSS I ON
Based on the extensive in situ forest inventories, we estimated the total forest biomass C stock for the mixed temperate forests in Northeast Asia to be 3.28 Pg C, 12.5% of which was contributed by planted forests (0.41 Pg C). ROK and northeastern China F I G U R E 5 Estimated tree mean diameter at 1.3 m above-ground on a 0.04 ha basis (D, cm) across the mixed temperate forest regions of Northeast Asia [Colour figure can be viewed at wileyonlinelibrary.com] F I G U R E 6 Estimated forest biomass C density (C d , Mg/ha) in 2080 under three climate change scenarios (Baseline, RCP4.5, and RCP8.5) across the mixed temperate forest regions of Northeast Asia [Colour figure can be viewed at wileyonlinelibrary.com] had 0.35 and 2.53 Pg C of forest biomass C stock, of which 20% (0.07 Pg C) and 13.4% (0.34 Pg C) was contributed by planted forests, respectively. DPRK had a total forest biomass C stock of 0.40 Pg C ( Table 2). The entire Northeast Asia stored approximately 0.91% of the total global forest biomass C stock (Pan et al., 2011) with only 0.01% of the global forest area. For northeastern China, compared to previous estimates of 1.79 Pg C for 1994-1998 (Fang et al., 2001) and 2.17 Pg C for 1997-1999(Tan et al., 2007, we found 16.59%-41.3% increase in forest biomass C stock over the recent 20-25 years, of which 45.95%-94.44% was contributed by planted forests. For DPRK, compared to a previous estimate of 0.23 Pg C for year 2000 , our estimates correspond to 73.91% increase in forest biomass C stock over the recent 20 years. For ROK, compared to previous estimates of 0.24 Pg C for year 2000 , our estimates correspond to 45.83% increase over the recent two decades, of which 63.64% was contributed by planted forests. Overall, we estimated that forest biomass C stock in Northeast Asia increased by approximately F I G U R E 7 Standardized variable importance values (%) determined by the Random Forests (RF) method: D (mean dbh), DH (dominant height), Elev (elevation), H 1 (the human footprint), LAT (latitude in WGS84 datum), and T 1 (annual mean temperature) were the most important independent variables for C d ; T 2 (mean diurnal range), T 6 (minimum temperature of coldest month), T 1 (annual mean temperature), T 11 (mean temperature of coldest quarter), P 1 (annual precipitation), and LON (longitude in WGS84 datum) were the most important independent variables for DH; Elev (elevation), FC (percent crown cover); LAT (latitude in WGS84 datum), LON (longitude in WGS84 datum), H 1 (the human footprint), and P 9 (indexed annual aridity) were the most important independent variables for D. See Table 1  20%-45% over the past 20 years, of which 40%-76% was contributed by planted forests (Table 3).
In general, our estimate of the Northeast Asia's mixed temperate forest biomass C density (C d , 60.77 Mg C/ha) was similar to the estimate of Thurner et al. (2014) for temperate broadleaf and mixed forests (53.8 Mg C/ha), but was higher than the general estimate (45.5 Mg C/ha) for East Asia   (Table 3). For ROK, this study's estimate of the biomass C density (C d , 54.2 Mg C/ha) was very similar to that obtained by Lee et al. (2014) (54.9 Mg C/ha), but almost 50% higher than the estimate by Li et al. (2010) and (Fang et al., 2014) (38.6 Mg C/ha). For DPRK, our estimate of the biomass C density (C d ) was 73% higher than that previously obtained by Fang et al. (2014) (63.7 vs. 36.8 Mg C/ha). Furthermore, our estimate of the biomass C density (C d ) in northeastern China (61.37 Mg C/ha) was 32% higher than previous estimates by Tan et al. (2007) (46.3 Mg C/ha), but was similar to a more recent estimate by Tang et al. (2018) (55.7 Mg C/ha).
There are three possible reasons behind the aforementioned differences between our current estimates and those from previous studies. First, the increased forest biomass C stock (C s ) can result from extensive afforestation, reforestation, and conservation efforts. Our results indicate that afforestation campaigns in Northeast Asia since the end of the 1970s (Fang et al., 2001;Lee et al., 2014;Xiao, 2005) made a strong positive influence on biomass C stock (C s ) (Table 2; Figure 8). More specifically, China over the past two decades has achieved one of the highest afforestation rate in the world , with 567,420 km 2 of afforestation-an area larger than Spain. In addition, since 1998, forest management policies in China have been redirected toward afforestation and reforestation, and logging have been largely prohibited or restricted to the levels of controlled harvesting (Yu et al., 2011), which has successfully promoted natural forest resource rehabilitation and recovery in northeastern China (Deng et al., 2012;Wei et al., 2014;Yu et al., 2015).
After the Korean War (1950)(1951)(1952)(1953), the governments of both DPRK and ROK have implemented massive programs aimed at restoring forests (Tak et al., 2007;UNEP, 2003). These forest restoration efforts have increased C sink over the past three decades Li et al., 2010). Second, such increase in forest biomass C stock (C s ) can be partially attributed to natural forest succession and climate change ( and Tan et al. (2007) are based on the 1970and 1982-1999 data, respectively. These data are dated two to five decades earlier than the dates of the forest inventories used in this study (2017for northeastern China, and 2011-2015. Forest biomass C stock (C s ) for the region over this period can increase significantly due to maturation of stand and natural succession , Li et al., 2016. Meanwhile, the impact of natural forest succession on biomass C stock (C s ) for this region can be amplified or altered by climate change . Over the past two decades, it has been observed that the average temperature in the forested regions of northeastern China has increased by 0.07℃/year (Piao et al., 2004), which can significantly increase forest biomass C stock (C s ) (Hararuk et al., 2015;Tang et al., 2018). Moreover, our results also show that continued climate warming would accelerate the increase in forest biomass C stock (C s ) (Table 4; Figure 6). Third, the differences in the methodology used to estimate the biomass C stock (C s ) can contribute to the differences in estimation results. The species-specific biomass allometric equations used in this study, consistent with those used in another recent study (Tang et al., 2018), can capture the geospatial differences in tree species composition better than a majority of the previous studies (Choi & Chang, 2004;Fang et al., 2014;Li et al., 2010) which calculate the biomass C stock (C s ) by multiplying the stem wood volumes by biomass expansion factors (BEF) that are not species specific and could result in an overestimation of the biomass C stock (C s ) for early successional forests (Guo et al., 2010;Lee et al., 2014).
This study shows that the mean dbh ( were also largely distributed in the same areas ( Figure 5). These findings were consistent with previous studies (Tan et al., 2007;Tang et al., 2018), which find that high precipitation and warm temperature conditions C d in these areas. Meanwhile, we also found a consistent positive effect of dominant height (DH) on C d in all ecoregions of Northeast Asia (Figure 8).
Our finding of a positive effect of elevation on C d in all ecoregions of Northeast Asia, consistent with two recent studies (Wang et al., 2008, can be attributed to the fact that high-elevation mountainous regions are relatively less affected by historical logging than the low elevation plains (Tang et al., 2006) due to excessive slope gradient and small compartment surface (McEwan et al., 2020).
Meanwhile, our finding of a positive effect of human footprint (H 1 ) on C d in all ecoregions but Eco5 (Figure 8) indicates that the influence of human interference on forest biomass C stock can go both ways.
In addition to a negative impact of logging on C d as described above, planting trees can have an opposite impact. The positive association between the human footprint (H 1 ) and mean dbh (D) in all ecoregions but Eco5 (Figure 9) as well as the Spatial Database of Planted Trees (SDPT v.1;Harris et al., 2020) imply that the ages of planted forests increased with the proximity to human settlements and metropolitan areas. This positive association in Northeast Asia is most likely caused by the fact that many reforestation and forest plantation projects in this region started near human settlements and metropolitan areas and extended outward in later stages (Yu et al., 2011).
Our estimates of the biomass C density (C d ) and C stock (C s ) involve two types of uncertainties. The first type of uncertainty was from the estimates of the forested areas. In this study, forest areas were defined as those with more than 20% crown cover (FC). As a result, our estimate of the forest areas in DPRK was similar to that of Fang et al. (2014) (6.26 × 10 6 ha vs. 6.30 × 10 6 ha), but higher than that of Li et al. (2010) (6.39 × 10 6 ha vs. 6.22 × 10 6 ha) in ROK, as the latter involves a higher crown cover threshold. The second uncertainty stemmed from a lack of in situ data from DPRK due to historical and political reasons. Nevertheless, trained from in situ data from northeastern China and ROK as well as local remote sensing and bioclimate data, our estimates for DPRK represent the most locally relevant biomass C density and stock estimates for this country.
Our results suggest that reforestation and forest conservation in Northeast Asia have effectively expanded the size of the carbon sink in the region , contributing to the climate change mitigation (Canadell & Raupach, 2008), as well as soil conservation and water quality improvement (Farley, 2007;Jackson et al., 2005;Thuille & Schulze, 2006). To this end, sustainable forest management practices such as precision forestry and close forest monitoring for fire and insect outbreaks Yu et al., 2015;Zhang et al., 2011) would be an effective tool to maintain and improve this critical carbon sink for Northeast Asia. Furthermore, our finding of a positive effect of climate change on forest biomass carbon stock for the entire region except ROK ( Stronger measures of sustainable forest management need to be taken to curb the negative effect of climate change on forest biomass carbon stock in ROK.

ACK N OWLED G EM ENTS
We

CO N FLI C T O F I NTE R E S T
None. Implications of increased deciduous cover on stand structure and