Improving the Simulation Accuracy of the Net Ecosystem Productivity of Subtropical Forests in China: Sensitivity Analysis and Parameter Calibration Based on the BIOME-BGC Model
[ad_1]
1. Introduction
Forests are the largest carbon reservoirs in terrestrial ecosystems and help maintain global carbon balance and mitigate global climate change [1,2]. Net ecosystem productivity (NEP), an indicator of carbon sinks in terrestrial ecosystems, is the difference between the net primary productivity (NPP) of vegetation and heterotrophic respiration [3] and can quantitatively describe the net carbon exchange between terrestrial ecosystems and the atmosphere.
Models coupled with remote sensing capabilities merge the inherent advantages of remote sensing, such as real-time, dynamic, broad spatially synchronous monitoring, and multiresolution attributes with the quantitative simulation capabilities of models that address diverse aspects of the ecosystem’s carbon cycle. This fusion facilitates expansive simulations of the carbon cycle processes, establishing it as a paramount methodology for extensive forest carbon sink monitoring [4]. The process model is developed based on vegetation physiological ecology and ecosystem processes and functions [5]. As a paradigmatic ecosystem process model, BIOME-BGC simulates physiological and ecological processes, such as photosynthesis, respiration, and decomposition, across a range of scales [6]. Its intrinsic mechanistic rigor, high degree of integration, and excellent extrapolation capabilities [7] render it an invaluable asset for terrestrial ecosystem studies, with broad applications both within and beyond national borders [8,9,10]. However, many of the model’s input parameters cannot be derived from real measurements and errors in parameter estimation lead to large uncertainties in the simulation results. Therefore, the targeted optimization and calibration of key parameters can significantly reduce the workload of model calibration and improve model accuracy and predictive capability [11].
Global sensitivity analysis can simultaneously test the influence of multiple parameter changes and interactions between parameters on the model output and is necessary for model optimization and parameter calibration. Qualitative global sensitivity analyses provide a qualitative measure of the effect of model parameters on simulation outcomes and include the Fourier amplitude sensitivity test (FAST) [12] and the Morris method [13]. Quantitative global sensitivity analysis calculates the contribution rate of model parameters to the uncertainty of simulation outcomes and mainly includes the Sobol method [14] and the extended Fourier amplitude sensitivity test (EFAST) [15]. The Morris method can screen out sensitive parameters under a small sample size and has good applicability to models with many analysis parameters and large computational loads. It is a qualitative global sensitivity analysis method widely used at present. Wang et al. [16] employed the Morris method to identify the most critical parameters in a C-RIVE model under various nutrient conditions. DeJongea et al. [17] demonstrated that the computationally less costly Morris method can effectively screen for sensitive parameters of the CERES-maize crop model under different irrigation conditions. The EFAST combines the capabilities of the Sobol method, which calculates the parameter interactions, and the efficient sampling of the FAST. The EFAST analyzes not only the impact of each parameter’s variation on model output but also the influence of interactions between multiple parameters on model outcomes [18], finding widespread application in numerous nonlinear models. Li et al. [19] used the EFAST to assess the contributions of various input parameters in the DSSAT-CERES model to wheat yield and quality. Gilardelli et al. [20] employed the EFAST to evaluate the impact on the yield of parameter value variations in crop models simulating different crops under climate change. The assessment of sensitive parameters affecting the model using suitable methods is a prerequisite for model calibration and optimization, as well as the basis for localized application of the model.
The NEP of subtropical forest ecosystems in the East Asian monsoon region, which plays an important role in the global carbon cycle and carbon sink function, is 0.72 Pg C·a−1 [21]. China is an important distribution area for subtropical forests and evergreen broad-leaved forests and evergreen coniferous forests are widely distributed in subtropical China [21]. The bamboo forest is known as the second-largest forest in the world and its carbon stock accounts for ~0.94% of the global forest ecosystem carbon stock [22]. China is known as the “world’s bamboo kingdom”, with abundant bamboo resources widely distributed in subtropical areas, such as Zhejiang, Fujian, and Jiangxi.
Simulations of the carbon and water cycles of various ecosystems found great uncertainty while monitoring subtropical forest carbon sinks, mainly for the following reasons. (1) When the BIOME-BGC model has been used to simulate the carbon cycle, the bamboo forest has not been treated as a separate forest, introducing uncertainty into the simulation. (2) Many studies have reported sensitivity analyses of BIOME-BGC parameters. Yan et al. [23] applied the EFAST to analyze and obtain sensitive parameters affecting the gross primary productivity (GPP) of the Changbaishan Forest. Tatarinov et al. [24] analyzed the effects of site and physiological–ecological parameters on the NPP of temperate forests in Europe. However, previous studies have predominantly focused on discerning the effects of parameter modifications on outputs, such as GPP and NPP. Notably, there is a paucity of evidence of the implications of model parameter fluctuations for NEP outputs, which amplifies the uncertainties in the simulation outcomes. (3) The calibration parameters of the BIOME-BGC model vary for different vegetation types under different environmental and site conditions. Raj et al. [25] investigated the sensitivity of the parameters in the BIOME-BGC model to the carbon fluxes of Douglas fir in the Speulderbos forest in the Netherlands and found that FLNR and FRC:LC (New fine root C: leaf C) were the key factors. Liu et al. [26] found that k was the most sensitive to estimate carbon fluxes in rubber forest ecosystems on Hainan Island. And FLRN was a low-sensitivity parameter but FRC:LC was not found to be the key control parameter. Kkumar and Raghubanshi [27] showed that FLNR and FRC:LC contributed less to the modeling of carbon fluxes in dry tropical forests. However, these studies suggest that the results of sensitivity analyses in other regions cannot be directly applied to simulate the carbon cycle in subtropical forests because of the climate, soil, and vegetation types.
In this study, we investigated the sensitivity of the physiological ecological parameters in the BIOME-BGC model to simulate the NEP of three forest ecosystems at five sites in subtropical China using the Morris method and EFAST. We aimed to screen key control parameters, improve the accuracy of model calibration, and provide scientific support for accurately assessing the carbon sink capacity of subtropical forests.
3. Results
3.1. Uncertainty Analysis
The uncertainty of the input data was calculated using the EFAST, as shown in Figure 4. The EBF (DHS site) had the highest NEP values, with a distribution ranging from 392 to 852 gC∙m−2∙a−1. The ENF (QYZ site) had a relatively scattered distribution of NEP, with a range of 150 to 690 gC∙m−2∙a−1. The NEP distribution of the BF (THY site) was relatively concentrated. Table 4 shows that there was no significant difference in the coefficient of variations (CV) between the two EBF sites, with a 13.86% CV for DHS and a 14.55% CV for TMS. The uncertainty in the ENF (CV, 23.22%) was significantly higher than that of the EBF and BF (CV = 14.24% and CV = 12.78%).
3.2. Morris Sensitivity Analysis
Figure 5 shows the sensitivity indices of the physiological and ecological parameters at each of the five sites (). For the EBF, k was the most influential parameter at both sites. In addition to k, the order of sensitive parameters at DHS was SLA, Gsmax, SC:LC, Q10Ract, and Gbl; for TMS, the sensitive parameters were in the order SLA, Gsmax, Gbl, FLNR, and Ract25. The order of the top three parameters was the same for both sites; however, the order of the less-sensitive parameters differed. The sensitive parameters affecting the ENF at QYZ were k, MRpern, Kc25, Q10Rac, Gbl, FLNR, and SC:LC. Additionally, k determines the amount of light energy that can be sequestered by the canopy, which directly affects the total photosynthetic primary productivity of the stand and shows high sensitivity in both the EBF and ENF. For the BF, AJ and THY were similar in terms of topography, soil, climate, and other conditions and their sensitivity parameters and rankings were also highly similar. The sensitivity parameters and rankings of AJ were Ract25 > Ko25 > Q10Rac > FLNR > FRC:LC > Kc25 > SCRages. Compared with AJ, THY had the fourth highest rank of Ko25 and the remaining sensitivities were consistent.
3.3. EFAST Sensitivity Analysis
The physiological and ecological parameters of Si and SiT for each of the five observation sites are shown in Figure 6.
For the EBF, the highly sensitive parameters affecting the NEP simulation results at DHS were k, SLA, Gsmax, and SC:LC and the moderately sensitive parameters were Gbl, Q10Ract, and FRC:LC. Among them, the SiT of k was much larger than that of Si; the SiT of Gsmax, Gbl, and Q10Ract was >0.1 and that of Si was <0.1, suggesting that these four parameters mainly influence the NEP simulated values through interaction with other parameters. The SiT of FRC:LC was <0; whereas, that of Si was >0, indicating that FRC:LC directly affects the NEP simulation results. Additionally, k was shown to be a highly sensitive parameter that affected the NEP simulation results at TMS and the moderately sensitive parameters were FLNR, Gbl, Gsmax, Ract25, C:Nleaf, and SLA. In addition, k, SLA, FLNR, and Gbl were common sensitive parameters at DHS and TMS. The EBF physiological and ecological parameters of different sites showed similar combinations of sensitive parameters.
The high-sensitivity parameters affecting the NEP simulation results at QYZ were k and MRpern and the medium-sensitivity parameters were Q10Ract, Gbl, FLNR, SC:LC, and Kc25, where the SiT of SC:LC was >0.1 and that of Si was <0.1. This suggests that SC:LC does not directly affect the results but rather varies synergistically with the other parameters, affecting the simulated NEP.
For the BF, the high-sensitivity parameters affecting the NEP simulation results at AJ were Ko25, SCRages, Ract25, and FRC:LC and the medium-sensitivity parameters included Fer, Kc25, Q10Ract, FLNR, and k. The high-sensitivity parameters affecting the NEP simulation results at THY were Ract25, Q10Ract, and SCRages and the medium-sensitivity parameters included FRC:LC, Ko25, FLNR, and Fer. The differences in the Si and the SiT of the sensitive parameters at both sites were small and they mainly affected the NEP of the BF through direct effects. Among these, SCRages and Fer are bamboo-forest-specific parameters that significantly affected the NEP simulation results for the BF.
3.4. Comparative Analysis of Morris and EFAST Results
As shown in Figure 7, the combination of the two sensitivity analysis methods led to ten key sensitive parameters affecting NEP: k, SC:LC, SLA, FRC:LC, Gsmax, MRpern, Ko25, Ract25, Q10Ract, and SCRages (for the BF only). The most sensitive parameter of the EBF and ENF calculated using the Morris method was k. The key control parameter of the BF (THY) was Ract25, which is consistent with the results of the EFAST. At DHS, the EFAST added FRC:LC over the Morris method. At TMS, C:Nleaf was added; at QYZ, the types of sensitive parameters were the same but the ordering was different. At AJ and THY, the EFAST added the bamboo-forest-specific parameter Fer over the Morris method. The types of sensitive parameters screened by the two methods that have a large impact on the NEP were highly similar; however, the order of importance of sensitive parameters was different. The EFAST generally calculates more sensitive parameters than the Morris method, partly because of the number of samples and partly due to the effects caused by the coupling of the parameters. The key control parameters vary from site to site and the key control parameters of each site show the same sensitivity at other sites with different degrees of sensitivity. Additionally, k is the common sensitive parameter in QYZ, DHS, TMS, and AJ and showed high sensitivity in the EBF (QYZ) and ENF (DHS and TMS), with a SiT of 0.4 or more, the first in the Morris method. In addition to k, MRpern was the most sensitive parameter for the ENF (QYZ) (SiT = 0.34 and = 0.54), SLA was the most sensitive parameter for the EBF (DHS) (SiT = 0.39 and ranged from 0.65 to 0.75), and Ko25 and Ract25 were the most influential parameters for the BF (AJ and THY) (SiT = 0.48, SiT = 0.56 and ranged from 0.5 to 1).
3.5. BIOME-BGC Model Validation
The carbon flux values simulated by the parameter localized model are closer to the flux observations than the original model. The original model had unrealistic peaks while the accuracy was significantly improved by the parameter-localized model (Figure 8). The R of NEP simulated using the original model ranged from 0.45 to 0.66 and NRMSE ranged from 15.88 to 28.43% for the five sites; the R of NEP simulated using the parameter-optimized model ranged from 0.61 to 0.75 and NRMSE ranged from 14.83 to 21.09%. Compared with the original model, the average R of the localized model improved by 25.19% and the average NRMSE reduced by 21.74%. The localized model can better simulate the changing characteristics of carbon fluxes in subtropical forests.
[ad_2]
