1. Introduction
Among the studies on global climate change and the carbon cycle, research on the quantity, distribution, and dynamics of forest carbon stocks is popular and remains a high priority for predicting the growth and yield of forests [
1,
2,
3]. Since the carbon concentrations in a tree or stand components are relatively constant (approximately 50%), most studies focus on forest biomass estimations rather than carbon storage estimations. Thus, the calculation of accurate forest biomass estimations has become one of the most crucial steps for successfully implementing the Reducing Emissions from Deforestation and Forest Degradation (REDD+) project as well as for the conservation and enhancement of forest carbon stocks and the sustainable management of forests. These initiatives provide a framework that benefits developing countries by rewarding them financially to reduce carbon emissions [
4,
5]. To date, different countries have used various methods for their national-scale carbon assessments, which differ in the degree of accuracy and the amount of field inventory [
6,
7,
8,
9,
10].
Forest stand biomass can be estimated at either the tree or stand level. At the tree level, the most commonly used method is the allometric equation, which estimates tree biomass using simple tree variables, such as the tree diameter at breast height (
D) and/or tree height (
H) [
11,
12,
13,
14]. Tree allometric equations require inventories that have high cost and time requirements, although they yield high accuracies for estimating tree biomass. The sum of the biomasses of individual trees provides the biomass for a plot or stand. Alternatively, the stand biomass can be estimated using either stand biomass models or biomass expansion and conversion factors (BCEFs), which represent the ratio of the stand biomass to stand volume. Historically, most biomass studies in the literature have focused on tree biomass models, while efforts related to stand biomass models have been limited or lacking [
6,
7,
9,
10,
15]. For the stand biomass estimations, some researchers believe that the use of stand biomass models can more easily estimate the stand biomass and prevent having to address complex error propagation procedures at different spatial scales [
6,
7].
In recent years, with the rapid accumulation of forest biomass data throughout the world, researchers have attempted to improve forest biomass estimations and proposed various stand biomass estimation methods [
7,
10,
16,
17,
18]. Studies have shown that stand biomass is closely related to some easily measured stand variables, such as the quadratic mean diameter, average height, and basal area of the stand [
7,
8,
19]. In addition, the total and individual components of stand biomass have a strong correlation with stand volume, which has been used as a predictor in stand biomass models [
20,
21,
22]. Recent methodology guidelines from the IPCC [
23] provided a set of species-specific default values for BCEFs. It is well known that using a constant BCEF to quantify stand biomass produces certain biases or errors because BCEFs vary depending on the growth conditions and stand development stage, such as the stand age, stand size and stand density [
15,
16,
21]. Studies have noted that BCEFs are not constant, and linear models have been established between stand biomass and volume [
20,
24,
25]. Other researchers constructed nonlinear models (e.g., hyperbolic functions, reciprocal equations, and power functions) to express the relationships between stand biomass and BCEFs [
7,
10,
15,
16,
17,
26]. Overall, using stand biomass models with stand volume as the sole predictor or expanding the stand volume with an available BCEF is a relatively simple method for estimating stand biomass [
7]. However, few comparative evaluations have been performed for different stand biomass estimation methods in the literature, especially for predicting stand biomass across a large geographic region.
When more than one biomass component is found in the same sample plot, the stand biomass equation is used to fit the total and component biomass data simultaneously, which explains the inherent correlations among the biomasses of the stand components in the same sample plot [
6,
12,
13]. Consequently, the sum of stand biomass predictions from the component biomass models and the total biomass model are the same. To achieve the additivity of stand biomass equations, various parameter estimation methods and model specifications are used in linear and nonlinear models [
11,
12,
27,
28,
29]. Among these methods of parameter estimation, nonlinear seemingly unrelated regression (NSUR) and seemingly unrelated regression (SUR) are the most widely used. An advantage of SUR and NSUR is the low variance of the total stand biomass model because of their own predictor variables and weighting function account for heteroscedasticity, which makes SUR and NSUR popular methods for parameter estimation in nonlinear and linear stand biomass equations [
11,
12,
13,
14,
30]. Although several researchers have proposed the inclusion of additivity, it has often been ignored in some stand biomass models [
8,
10].
Additionally, similar to tree biomass models, stand biomass models commonly show heteroscedastic model residuals. To overcome the heteroscedasticity of the stand biomass model residuals, logarithmic transformation or a weighted regression should be performed before the construction of each carbon model [
12,
13]. To acquire an ideal result from logarithmic regression, a correction is necessary after the antilog transformation, i.e., the predicted values are multiplied by a correction factor [
31,
32]. However, when determining the total and component equations of stand biomass, after applying the correction factor to the logarithmic equations of the additive system, realizing additivity is difficult [
32]. Thus, the weighted regression successfully overcomes the heteroscedasticity of the total and component biomass model residuals in an additive system [
12,
13].
Asia is one of three primary locations of temperate mixed forests worldwide (i.e., northeastern North America, Europe, and eastern Asia), and they are mainly distributed in the forest regions of northeastern China. Chinese temperate forests are widely distributed in the Eastern Da Xing’an Mountains. These temperate forests play a crucial role in the Chinese national carbon budget and climatic system. Unfortunately, the forest resource inventory data in this region have not been fully utilized, and only limited stand biomass models are available. Thus, the objectives of this study were to (1) develop three alternative additive systems of stand biomass equations (i.e., stand biomass models using stand variables (M-1), stand biomass models using stand volume (M-2), and stand biomass models using both stand volume and BCEF (M-3)) for estimating the stand biomass of major forest types (including white birch forest, larch forest, poplar-birch forest, deciduous broadleaf mixed forest, coniferous and broadleaf mixed forest, and coniferous mixed forest) in the Eastern Da Xing’an Mountains, Northeast China, (2) use the jackknife method to validate the performance of those stand biomass models, and (3) evaluate the predictive ability of four alternative methods (i.e., M-1, M-2, M-3, and constant BCEF (M-4)) to estimate the total and component biomasses in each forest stand.
4. Discussion
Stand total and component biomasses are frequently estimated using stand variables, which are usually easy to obtain via field investigations. In recent studies on stand biomass estimations, typical allometric equations based on the power-law model are often applied to increase the prediction accuracy of stand biomass [
6,
7,
8,
16]. In our M-1 model system, the stand basal area
was a significant and important predictor for the stand total and component biomasses of the six forest types, which confirmed its role in previous studies [
6,
7]. The estimated coefficients of the stand basal area were positive, indicating that the stand biomass increased as the stand basal area increased. However, for the same stand basal area, large variations occurred among the stand total and component biomasses. Therefore, to improve the prediction accuracy of the stand biomass models, other stand variables should be considered. Previous studies have verified that the dominant height of the stand was the second most important stand variable because it reflected the site quality [
7,
19]. Unfortunately, the stand dominant height was not available in our data from the permanent sample plots of the NFCI. Instead, we used the average stand height
Ha, which was statistically significant for most stand root, stem, branch, and foliage biomass models. As a result, we strongly suggest that future studies on stand biomass consider and investigate the effects of the stand average height or dominant height in the stand biomass models. Other stand variables, such as stand density and stand age, were also included in some stand component biomass equations, and they were commonly considered to represent the competition within a stand and the stage of stand development [
6]. The effects of stand density and stand age on some biomass components, such as stem or foliage, are statistically significant and widely recognized, which is likely due to their impacts on branching characteristics and biomass partitioning among the tree components [
6,
37]. However, the six forest types in this study were natural forests in the Eastern Da Xing’an Mountains. Therefore, they were not good choices for estimating stand biomass. In addition, information on silvicultural practices, such as thinning and tending, was not available, which may affect stand biomass accumulation and allocation and result in biased stand biomass estimations [
38].
Many studies indicate that there is a strong correlation between stand biomass and stand volume [
20,
21,
22]. For most forest types, a linear relationship between stand biomass and stand volume is obviously insufficient and controversial. Thus, many researchers have used nonlinear models to describe the relationship and improve model fitting [
7,
15,
16,
17,
24]. As expected,
was not constant over stand development in different forest types (
Table 2). Thus, using constant
values would provide biased estimations of stand biomass. In this study, Equations (5) and (6) were used to model the
variations, which is consistent with the methods used in other studies in the literature. In addition, some studies used stand age as a predictor for modeling BCEF [
15,
39,
40,
41]. Although the stand age was unavailable in this study, the stand variables (e.g.,
Ha and
Dq) used in our model systems were indirect surrogates for stand development to a certain extent.
Parresol [
28] noted that the aggregation approach is the standard method for ensuring the additivity of estimates of stand total and component biomasses. In Parresol’s method, a nonlinear model is assigned to each of the stand biomass components before aggregating the biomasses of these stand components into the stand total biomass. Notably, if the same predictor variables are used for modeling all stand biomass components, and the same weights are chosen for heteroscedasticity in the model residuals, NSUR will produce singular covariance matrices that do not guarantee a unique solution. Thus, when the same or different predictor variables are considered, the heteroscedasticity of the model residuals should be overcome by different sets of weights, and NSUR is a feasible parameter estimation method [
12,
42]. The aggregative models in this study were estimated using a weighted NSUR with different weight functions to explain the inherent correlations among the stand component biomasses in the same sample plot [
11,
12,
13,
27,
28]. Unfortunately, some nonadditive stand biomass models are still published because of the use of the ordinary least-squares regression (OLS) estimation method [
8,
10].
We evaluated different methods for quantifying stand biomass, and the results were nonconclusive. The results indicated that, except for deciduous broadleaf mixed forest, M-3 performed better than other models, especially for stand total and stem biomass. The reason for this finding may be that there is a close relationship between stand volume and stem biomass, and a higher proportion of stem biomass is included in the total biomass [
7,
43]. If the stand volume is not available, M-1 will be an effective method to accurately estimate the stand total and component biomasses. For the deciduous broadleaf mixed forest, the stand biomass models with stand variables (M-1) were generally less biased than those obtained from the stand biomass models with stand volume (M-2) or stand volume and BCEF (M-3). The poor performance of the latter two additive models may be the result of the biases after modeling
by using the stand variables. Many studies have shown that using a constant BCEF to quantify stand biomass introduces certain biases or errors in stand biomass or carbon estimations so that the development of
models is recommended when the data are available [
16]. Generally,
vary depending on the growth conditions and stand development stage expressed by the age and size of the stand [
10,
15,
16,
17,
41]. Thus, it is easy to understand that M-4 exhibited a worse prediction accuracy than M-2 and M-3. In general, both M-2 and M-3 are stand biomass estimations method based on stand volume, which is essentially different from M-1 [
44,
45,
46]. Overall, the use of M-1, M-2, and M-3 in conjunction with stand growth and yield models would be useful for the growth predictions of the six forest types in response to changes in stand conditions and is appropriate for sustainable forest ecosystem management studies.
To date, few studies have developed individual tree biomass equations for the major forest species in the Eastern Da Xing’an Mountains, Northeast China [
14]. However, the data required to apply these equations are not always available. In addition, some growth models for the six forest types in the region were previously developed at Northeast Forestry University, China, which mainly calculated the stand variables, such as
and
. For the stand biomass estimations, the use of stand variables can likely avoid the need to address complex error propagation procedures for stand biomass estimations at different spatial scales [
7]. Although estimating the stand biomass using stand variables may produce some differences in the tree-based biomass estimation, the differences seem acceptable at the forest or stand level for a large geographic region. Thus, the stand biomass models using stand variables (e.g.,
,
,
, and
) would be useful, convenient, and efficient.