1. Introduction
To obtain information on the potential future productivity of plantation forests, forest managers rely on various types of growth models [
1]. The stand level growth model can be used to predict the development of even-aged forests, but for mixed forests or uneven-aged forests the individual tree growth model is required [
2]. The individual tree growth model is a bottom-up modeling method, starting from individual trees in the system and ending at the stand level, aiming to reveal and predict the growth mechanism of individual trees [
3]. Compared with pure even-aged forests, the complexity, depth, and breadth of the forest growth simulations for mixed uneven-aged forests increased, and the scale gradually shifted from stands to individual trees [
4]. And further compared with the whole stand model, the individual tree model is more applicable [
5], promoting the widespread use of individual tree growth models in forest management [
6].
According to the concept of mathematical statistics, the model can be divided into parametric model, non-parametric model, and semi-parametric model. A parametric model mainly has the form of an algebraic equation, differential equation, and transfer function [
7]. It is widely used in forestry because of its fast modeling speed and small amount of data required. With the development of computers, non-parametric models have expanded the application range of forestry models with their strong applicability and lax requirements on model assumptions, such as non-parametric additive model [
8], the classification and regression tree (CART) [
9], support vector machines (SVM) [
10], random forest (RF), and neural network (NN). In recent years, semi-parametric models have been gradually applied to the study of tree mortality and inside boundary timber, such as the Cox proportional hazards model [
11] and semi-parametric geographically weighted Poisson regression (SGWPR). The semi-parametric model contains a parametric part and non-parametric part, which overcomes the deficiency of the parametric model and non-parametric model and solves the “dimensional disaster” of the non-parametric model and other problems. However, due to the relative complexity of the semi-parametric model, there are many problems from theoretical research to practical application. At present, the construction of tree growth models is still dominated by parametric models and non-parametric models, and semi-parametric models are rarely studied.
Regression analysis is the most commonly used method for the modeling of individual tree growth models [
12,
13,
14], but the traditional regression method is usually based on certain statistical assumptions, such as data independence, normal distribution, and equal variance. Forest growth data have the characteristics of continuous observation and hierarchy, which makes it difficult to satisfy the assumptions in general. As a result, it is difficult for traditional regression models to achieve higher prediction accuracy; so, it is necessary to try new modeling methods [
15]. With the development of data processing technology, artificial intelligence has provided more advanced technical ideas to overcome such problems. An artificial neural network was used earlier in individual tree growth modeling [
16], which can approximate any nonlinear trend to the maximum extent [
17]. In particular, the BP neural network (BP) is the most commonly used artificial neural network in model research due to its strong generalization ability, high accuracy, and small error [
18]. In addition, random forest (RF) in ensemble learning has also been gradually applied in forestry research [
19]. RF can be highly parallelized in training and can provide the feature importance of each factor on the output. RF has a strong model generalization ability and is not sensitive to partial feature loss [
20]. Weiskitte [
21] studied the climate-driven site index with RF. Kilham [
22] applied RF to the selection of felled trees and the prediction of stand accumulation. Currently, RF is widely applied in forestry remote sensing and other fields; however, it is rarely applied to the prediction of forest growth and harvest.
Tree growth is affected by many factors, among which genotype, climate, and site are the main driving factors [
23,
24]. Therefore, environmental factors should be included in the growth model to explain the growth pattern and improve the accuracy of tree growth prediction [
25,
26]. The individual tree growth model usually includes a competition index to quantify the degree of competition pressure of trees in the stand [
27]. Generally speaking, adding competition effects to the growth model can improve the model performance [
28]. Many current studies add a competition index dependent on distance or independent distance to the tree growth model to quantify competition [
29]. It is used to predict tree diameter at breast height (DBH) growth or tree height (H) growth [
30,
31,
32], which improves the accuracy of the model. Although some studies have shown that the growth of trees is obviously correlated with the distance of surrounding trees [
33,
34], some scholars have found that the competition index of independent distance is more effective in fitting the growth of trees [
35], which may be related to the difference in research objects and the calculation of the competition index.
Site is an important environmental factor affecting tree growth, which determines the water, heat, and soil fertility and is the core factor for the growth and distribution evolution of trees under natural conditions [
36,
37]. For example, slope gradient leads to light heterogeneity [
38]. Slope aspect can affect light, temperature, and soil physical and chemical properties and then affect tree growth and stand structure [
39,
40]. The influence of soil on plants is also well known; nitrogen, phosphorus, potassium and other nutrients, and humidity in soil vary with soil thickness and soil type [
41]. Adding site factors into the growth model is of great significance for selecting reasonable afforestation tree species, increasing ecological stability, and supporting forest productivity [
42,
43]. Climate is an important external environment that determines the dynamic change of tree distribution and forest function at the regional scale [
44,
45]. Under the same site conditions, different climates will lead to different tree growth. Therefore, it is crucial to add climate factors into the growth model to improve its universality [
46,
47].
Close-to-nature forest management is a feasible theory and technology to improve the stability of forest ecosystems [
48]. Building mixed uneven-aged forest is one of the main management measures of close-to-nature forest management. In order to build a mixed uneven-aged forest, it is necessary to conduct management operations on the forest, such as tending and thinning, adjusting the stand structure, etc. One of the important problems to be solved is to estimate the results of management operations, such as the harvest volume and biomass, etc. The solution is to build an individual tree growth model. There are many factors affecting forest growth, and it is very important to study the co-driving mechanism of competition and the environment to predict tree growth. However, there are much research on grassland and aquatic ecosystem, but there is little research on forest ecosystems [
49,
50] and most of them focus on the unilateral effects of stand density, climate factors, and tree species composition on tree growth. There are few individual tree growth models that include competition, site, and climate factors.
Chinese fir (
Cunninghamia Lanceolata (Lamb.) Hook.) is one of the most important timber species in the south of China. The results of the ninth national continuous forest inventory (CFI) show that the area of Chinese fir plantations accounts for 1/4 of the total area of China’s artificial arbor forest, and the planted area ranks the first among afforestation tree species [
51]. As one of the six large forest regions in China, Fujian Province is located in the subtropical climate zone along the southeast coast of China, with complex and changeable climatic conditions and frequent extreme weather and climate events, making it a highly sensitive area to climate change [
52]. Therefore, we took Chinese fir as the study tree species and Fujian Province as the study area to reveal the mechanism of competition factors (Comp), site factors (Site), and climate factors (Clim) effects on individual trees. The aims of this research study are: (1) to quantify the influence degree of factors, in order to select the relatively important factors in the individual tree growth model; (2) to explore the use of the re-parameterized (RP) method, BP, and RF algorithm to construct an individual tree growth model, which compares the accuracy and adaptability of different methods; and (3) to add the screened factors into the model and construct a regionally compatible individual tree growth model under the combined influence of environment and competition. The results of this study can provide references for the construction of individual tree growth models for different tree species in different regions, provide model support for growth harvest prediction of mixed uneven-aged forests, and also have indirect support significance for predicting forest carbon stocks and rational response to climate change.
3. Discussion
Existing studies on individual tree growth models are constructed using different methods for Chinese fir. Using multiple stepwise regression methods to build the linear mixed-effects model with sample plots as random effects, the R
2 of the model is 0.676 [
55]. Considering climate and competition, an individual tree growth model at DBH was constructed using the Bayesian model with an R
2 of 0.6121 [
56]. The results of our study clearly demonstrated the advantages and potential of machine learning algorithms in the individual growth modeling of Chinese fir. Compared with the formula model, the machine learning model can model the complex nonlinear relationship without being restricted by statistical assumptions. Especially when the model contains independent variables and a large amount of data, it is more convenient to use machine learning in the modeling process and has obvious advantages.
The influence of environmental factors on the growth of an individual tree was analyzed through the results of 10-fold cross-validation.
Supplementary Figure S7 shows that, at the level of individual trees, the age of individual trees is the main factor affecting the growth of trees. The relative importance of individual trees was extremely high, exceeding 60%, and the relative importance of competition factors to the model was about 8%.
In Site, PW had the most significant effect. In the BP and RF models, the model accuracy was increased by about 4%. Followed by HB and TRHD, the model accuracy increased by about 3%, slope aspect (PX) and DM had the least effect, and the contribution rate was about 1%. This result was not entirely consistent with those of previous studies [
57]. The main reason for this is the difference in the study area. Fujian Province has more hills and mountains, and the vertical variation in heat and precipitation is more pronounced, which determines the growth difference of the study subjects; therefore, PW and HB become important factors.
PW represents the soil erosion and accumulation capacity. Generally speaking, soil temperature gradually decreases from top to bottom, while moisture gradually increases. Chinese fir is an acidic positive tree species that likes deep, fertile, and moist soil and has good drainage conditions, and a down slope position is more conducive. HB represents temperature, elevation increase, and temperature decrease, which is not conducive to the growth of Chinese fir at DBH [
58]. TRHD represents soil fertility and water holding capacity. The thicker the soil, the higher the comprehensive fertility and water holding capacity, promoting the growth of DBH. This conclusion is consistent with the findings of Monserud et al. [
59]. PX represents light, and light is more sufficient on sunny slope than on shady slopes, and the photosynthesis of Chinese fir is stronger, promoting DBH growth; this conclusion is consistent with the findings of LU [
60]. DM represents the soil nutrient distribution and temperature and humidity conditions in the woodland. Compared with hills, Chinese fir has a higher material accumulation capacity in mountains with abundant rainfall and high air humidity [
61].
We also found that the partial dependence of a variable on DBH growth is highly correlated with the relative importance of the variable (by comparing the ranking results of characteristic importance). Specifically, when the relative importance of a variable is greater, the average annual diameter growth of individual trees changes more sharply with the change in the variable. When the relative importance of a variable is less, the average annual diameter growth of individual trees changes more smoothly with the change in the variable. At the same time, competition factors, site factors, and climate factors have interactive effects on the growth of individual trees.
5. Conclusions
The re-parameterized model and machine learning model were used to construct the DBH model and H model of Chinese fir. The input variables included T, Comp, Site, and Clim. The results showed that: (1) The addition of competition and environmental factors could improve the prediction accuracy of the individual tree growth model. In terms of site factors, PW had the most significant effect, followed by HB and PX. In terms of climate factors, the DD18 contribution rate was the highest, followed by MAP, and Eref. (2) Among the three models, the RF model fits best. The model constructed in this study could well reflect the growth process of individual DBH and H of Chinese fir in Fujian Province and has reference significance for the growth models of other provinces and other tree species.
Climate factors were added into the individual tree growth model, which fully reflected the spatial differences of climate factors and solved the key problem of applicability of the individual tree growth model in different regions. The model was applied to the forest management decision support system. The results showed that the individual tree growth model, including T, Comp, Site, and Clim, could be used to predict the stand harvest in different sites and climates and provide decision support for harvest prediction and management method selection of mixed uneven-aged forests.
Since the area of Chinese fir plantation forest is the largest in terms of plantation forests in China, this paper only took Chinese fir as an example to construct the individual tree growth model. In future research, the modeling approach in this paper could be applied to the construction of individual tree growth models for other tree species, such as Masson pine (Pinus massoniana Lamb.) and broadleaf species. We also recommend using longer study years to investigate the variables to provide more reliable results.