Improved CASA-Based Net Ecosystem Productivity Estimation in China by Incorporating Developmental Factors into Autumn Phenology Model

Abstract

Accurately assessing the carbon sink intensity of China’s ecosystem is crucial for achieving carbon neutrality. However, existing ecosystem process models have significant uncertainties in net ecosystem productivity (NEP) estimates due to the lack of or insufficient description of phenological regulation. Although plant developmental factors have been proven to significantly influence autumn phenology, they have not been systematically incorporated into autumn phenology models. In this study, we modified the autumn phenology model (cold-degree-day, CDD) by incorporating the growing-season gross primary productivity (GPP) and the start of growing season (SOS) and used it as a constraint to improve the CASA model for quantifying NEP across China from 2003 to 2021. Validation results showed that the CDD model incorporating developmental factors significantly improved the simulation accuracy at the end of the growing season (EOS). More importantly, compared with flux tower observations, the NEP derived from the improved CASA model based on the above phenology model showed a 15.34% reduction in root mean square error and a 74% increase in the coefficient of determination relative to the original model. During the study period, China’s multiyear average total NEP was 489.67 ± 38.27 Tg C/yr, with the highest found in evergreen broadleaf forests and the lowest detected in shrublands. Temporally, China’s NEP demonstrated an overall increasing trend with an average rate of 1.75 g C/m²/yr². However, the growth rate of NEP remained far below concurrent carbon emissions from fossil fuel combustion totally, especially for eastern China, while the northeastern regions performed relatively better. The improved regional carbon flux estimation framework proposed in this study will provide important support for developing future climate change mitigation strategies.

1. Introduction

Terrestrial ecosystems have absorbed approximately 30% of anthropogenic carbon dioxide emissions over the past decade, effectively mitigating global climate warming [1]. China’s terrestrial ecosystems have been recognized as a significant component of the global carbon sink [2,3,4]. However, due to the diversity of China’s climate and ecosystems along with the significant spatial heterogeneity of terrestrial ecosystems, there are still considerable uncertainties in estimating the carbon sink capacity [3]. Reducing these uncertainties is crucial for refining the global carbon budget, supporting the formulation of climate policies, and predicting future climate change.

Net ecosystem productivity (NEP), calculated as the difference between net primary productivity (NPP) and soil heterotrophic respiration (HR), is a key indicator for assessing carbon fluxes in terrestrial ecosystems [5]. The estimation of regional terrestrial ecosystem carbon fluxes is typically categorized into ‘bottom-up’ and ‘top-down’ approaches [3]. Terrestrial ecosystem models, such as the widely used Carnegie–Ames–Stanford Approach (CASA) model, are common ‘bottom-up’ methods and are important tools for assessing carbon sinks on global and regional scales [6,7]. These models are usually driven by inputs, such as climate variables and land-use change data, and can predict future changes in ecosystem carbon sinks [3]. However, the factors driving carbon sinks in these models are often not fully comprehensive [3]. Recently, an increasing number of studies have highlighted the significant impact of changes in the growing season length of temperate plants on ecosystem productivity [8,9]. While warmer autumns may increase carbon loss by extending respiratory activity, the advance in the start of the growing season (SOS) and the delay in the end of the growing season (EOS) caused by climate warming have extended the carbon uptake period and increased the net carbon uptake of Northern Hemisphere forest ecosystems [8]. Therefore, incorporating vegetation phenology information into ecosystem models is crucial for accurately assessing vegetation productivity.

Traditional plant phenology models typically use accumulated temperature to determine EOS [10,11,12]. However, the drivers of autumn phenology are more complex, being influenced not only by climatic conditions but also by the developmental status of plants, including the growing season gross primary production (GPP) and SOS [9,10,12]. Developmental factors can influence plant senescence through direct mechanisms, such as carbon sink limitations from photosynthesis and programmed cell death, as well as indirect mechanisms, like intensified summer drought [13,14]. Studies have shown that warming-induced advances in SOS and increased growing season GPP may lead to earlier plant senescence, partially offsetting the delaying effect of warming on EOS [9,15]. Furthermore, if phenology models overlook plant developmental factors, they may overestimate future delays in autumn phenology, thereby inflating projections of future ecosystem productivity [9,12]. Integrating phenology algorithms into GPP models has been demonstrated to significantly enhance model accuracy and reduce biases compared to flux tower-measured GPP [11]. Therefore, improving phenology models and integrating their outputs with NPP models is essential for achieving more accurate regional estimates of NEP.

In this study, we proposed an improved autumn phenology model and used it to constrain the output of the CASA model, aiming to enhance the prediction accuracy of NEP for China’s terrestrial ecosystems. The objectives of this study were: (1) to construct an autumn phenology model incorporating developmental factors; (2) to assess the NPP simulation performance of the CASA model modified by incorporating phenology information; and (3) to analyze the spatial patterns and temporal trends of NEP in China from 2003 to 2021 based on the phenology-modified CASA model. This study will contribute to accounting for ecosystem carbon sinks, thereby helping to formulate carbon reduction policies and achieve carbon neutrality goals.

2. Materials and Methods

2.1. Data Sources

To extract phenological metrics, the normalized difference vegetation index (NDVI) dataset from MODIS (MYD13C1) was used. The temperature (Tem) datasets used in the autumn phenology model and NEP simulation were also obtained from MODIS (MOD11C2). GPP data used to improve the phenology model came from the revised EC-LUE (Eddy Covariance Light Use Efficiency) model [16], and GPP estimates based on satellite NIRv (near-infrared reflectance) [17]. Due to the highly consistent interannual variations between the two datasets (Figure S1), we averaged them for model improvement.

The fraction of absorbed photosynthetically active radiation (FAPAR) and photosynthetically active radiation (PAR) required for the CASA model were obtained from Global Land Surface Satellite (GLASS) products. Monthly actual evapotranspiration (ET) was sourced from the ECMWF Reanalysis v5 (ERA5) dataset. Potential evapotranspiration (PET) and precipitation (Pre) were obtained from the China 1-km monthly potential evapotranspiration and precipitation datasets [18,19].

For the land cover data, the MODIS Land Cover Climate Modeling Grid (MCD12C1) Version 6 product was used. This study focused on nine major and widely distributed vegetation types in China, including evergreen needleleaf forest (ENF), evergreen broadleaf forest (EBF), deciduous needleleaf forest (DNF), deciduous broadleaf forest (DBF), mixed forests (MF), shrublands (SHB), savannas (SAV), grasslands (GRA), and croplands (CRO), with their distribution shown in Figure 1.

We used satellite data products from 2003 to 2021 to estimate NEP and test the feasibility of the new NEP simulation framework. All satellite data were standardized to a spatial resolution of 0.05° using the arithmetic average. The spatiotemporal resolution and data acquisition details of the satellite data used in this study are shown in

Table 1. The spatiotemporal resolution and data accessibility details of the satellite datasets.

Dataset	Spatial/Temporal Resolution	Data Access
MODIS NDVI (MYD13C1)	0.05°/16-day	https://ladsweb.modaps.eosdis.nasa.gov (accessed on 29 January 2025)
MODIS Tem (MOD11C2)	0.05°/8-day	https://ladsweb.modaps.eosdis.nasa.gov (accessed on 29 January 2025)
Revised EC-LUE GPP	0.05°/8-day	https://doi.org/10.6084/m9.figshare.8942336.v1 (accessed on 29 January 2025)
GPP based on NIRv	0.05°/monthly	https://doi.org/10.6084/m9.figshare.12981977.v2 (accessed on 29 January 2025)
GLASS FAPAR	250 m/8-day	http://glass.umd.edu/ (accessed on 29 January 2025)
GLASS PAR	0.05°/daily	http://glass.umd.edu/ (accessed on 29 January 2025)
ERA5 ET	0.1°/monthly	https://cds.climate.copernicus.eu/datasets (accessed on 29 January 2025)
China PET	1 km/monthly	http://loess.geodata.cn (accessed on 29 January 2025)
China Pre	1 km/monthly	https://doi.org/10.5281/zenodo.3114194 (accessed on 29 January 2025)
MODIS land use (MCD12C1)	0.05°/yearly	https://ladsweb.modaps.eosdis.nasa.gov (accessed on 29 January 2025)

.

To test the performance of NEP simulation, we obtained 52 site-year NEP flux records across 7 flux tower sites in China (Figure 1), downloaded from the National Science & Technology Infrastructure platform (https://www.nesdc.org.cn/ (accessed on 29 January 2025)). Finally, to evaluate the synergistic variation trend between fossil fuel carbon emissions and ecosystem carbon sequestration, we obtained apparent carbon emission data for China’s four economic regions (eastern, central, western, and northeastern) from 2003 to 2021 from the Carbon Emission Accounts and Datasets (CEADs) [20,21].

2.2. Methods

The framework for NEP improvement involved three steps: (1) improving the EOS simulation by introducing developmental factors (growing-season GPP and SOS) into the traditional autumn phenology model; (2) incorporating the improved EOS simulation into the CASA model to enhance the accuracy of NPP estimation; and finally, (3) quantifying the terrestrial NEP by subtracting the soil heterotrophic respiration from the modified NPP. The scheme of the work is illustrated in Figure 2.

2.2.1. Phenological Extraction Methods

Phenology metrics were retrieved based on MODIS NDVI data. NDVI observations during the non-growing season, defined as the multiyear daily average temperature below 0 °C, were first replaced with the average value in the non-growing season. The Savitzky–Golay filter was then applied to smooth and reconstruct the NDVI curve. A double logistic function was used to fit the reconstructed NDVI time series to extract SOS and EOS. Specific details of the phenology extraction can be found in Ren et al. (2020) [22].

2.2.2. Improvement of the Autumn Phenology Model

The traditional autumn phenology model is based on low temperature accumulation, known as the cold degree day (CDD) model [10]. Considering the potential impacts of SOS and growing season GPP on EOS [9,15], we incorporated SOS, GPP, and their combination into the CDD model, naming the models CDD_S, CDD_G, and CDD_SG, respectively. The CDD model hypothesizes that the cooling degree starts to accumulate when the daily temperature drops below the critical temperature and stops on the day of foliar senescence [10]. The formula was as follows:

$${CDD}_d=\sum _{d=d_0}^{d_y}\max (T_{th}-T_{\left(d\right)}, 0)$$

(1)

$$EOS=d_y, if {CDD}_d\ge {CDD}_{th}$$

(2)

where ${CDD}_d$ represents the accumulated cooling degree from day $d_0$ to day $d_y$; ${CDD}_{th}$ is the required threshold upon which EOS occurs; $T_{th}$ is the critical temperature; $T_{\left(d\right)}$ is the mean daily temperature on day d; and $d_0$ is set to 1 July [23]. The model has two free parameters, $T_{th}$ and ${CDD}_{th}$.

For the CDD_S and CDD_G models, ${CDD}_{th}$ depends on SOS and the growing season GPP (Equations (3) and (4)), while the ${CDD}_{th}$ for CDD_SG model is linearly linked to both SOS and the growing season GPP (Equation (5)). The ${CDD}_{th}$ of the modified models are expressed by the following equations:

$${CDD}_{th}=a+b\ast S_a$$

(3)

$${CDD}_{th}=a+b\ast G_a$$

(4)

$${CDD}_{th}=a+b\ast S_a+{c\ast G}_a$$

(5)

where $S_a$ is the SOS anomaly (difference from the multiyear average SOS); $G_a$ is the growing season GPP anomaly (difference from the multiyear average GPP); and $a$, $b$, and $c$ are free parameters. To optimize free parameters in models, the simulated annealing algorithm was employed in this study. The optimal parameter combination was determined with the smallest root mean square error (RMSE) between modelled and observed EOS.

2.2.3. Modification of the CASA Model

In the CASA model, NPP was determined by the absorbed photosynthetically active radiation (APAR) and the light use efficiency ($\epsilon$) [24]. Here, APAR was obtained from the PAR and FAPAR.

$$NPP\left(\chi ,t\right)=APAR\left(\chi ,t\right)\times \epsilon \left(\chi ,t\right)$$

(6)

$$APAR\left(\chi ,t\right)=PAR\left(\chi ,t\right)\times FPAR\left(\chi ,t\right)$$

(7)

where $\chi$ denotes the location; $t$ represents time; and $NPP\left(\chi ,t\right)$ is the NPP (g C/m²) at location $\chi$ and time $t$.

$\epsilon$ is constrained by both temperature and moisture and is calculated as follows:

$$\epsilon \left(\chi ,t\right)=T_{\epsilon 1}\left(\chi ,t\right)\times T_{\epsilon 2}\left(\chi ,t\right)\times W_{\epsilon }\left(\chi ,t\right)\times {\epsilon }_{max}$$

(8)

where $T_{\epsilon 1}\left(\chi ,t\right)$ and $T_{\epsilon 2}\left(\chi ,t\right)$ represent the effects of low and high temperatures on light use efficiency, respectively; $W_{\epsilon }\left(\chi ,t\right)$ is the moisture stress; and ${\epsilon }_{max}$ is the maximum light use efficiency under ideal conditions. Detailed calculation methods and values for $T_{\epsilon 1}$, $T_{\epsilon 2}$, and ${\epsilon }_{max}$ can be found in Liang et al. (2022) and Chen et al. (2024) [5,25].

The NPP calculated by the CASA model was interpolated to a 1-day resolution using a cubic spline function, and the CASA results were constrained by the actual vegetation phenological dates and temperature [11,26].

$${NPP}_{mod}=f\times {NPP}_{org}$$

(9)

$$f=\left\{\begin{array}{c}0, if T\le 0\\ {\displaystyle \frac{1}{\left(\frac{5-T}{5}\right)+1}}\times {\displaystyle \frac{1}{\left(\frac{SOS-d}{SOS}\right)+1}}, if T\le 5 and d\le SOS\\ {\displaystyle \frac{1}{\left(\frac{SOS-d}{SOS}\right)+1}}, if T>5 and d\le SOS\\ {\displaystyle \frac{1}{\left(\frac{d-EOS}{EOS}\right)+1}}, if T>5 and d\ge EOS\\ {\displaystyle \frac{1}{\left(\frac{d-EOS}{EOS}\right)+1}}\times {\displaystyle \frac{1}{\left(\frac{5-T}{5}\right)+1}}, if T\le 5 and d\ge EOS\end{array}\right.$$

(10)

where ${NPP}_{org}$ is the monthly NPP simulated by the original algorithm in the CASA model, and ${NPP}_{mod}$ is the result after being constrained with $f$. $d$ is the day of year, and $T$ is daily average temperature. SOS was satellite-derived, while EOS was derived from the CDD_SG model.

2.2.4. NEP Estimation

NEP was obtained by subtracting soil heterotrophic respiration (HR) from NPP.

$$\mathrm{N}\mathrm{E}\mathrm{P}=\mathrm{N}\mathrm{P}\mathrm{P}-\mathrm{H}\mathrm{R}$$

(11)

The geostatistical model of soil respiration (GSMSR), driven by temperature, precipitation, and soil organic carbon density [27], was employed to obtain the monthly soil respiration ($R_S$).

$$R_S=(R_{D_{S=0}}+M{D}_S)e^{\ln \alpha e^{\beta T}T/10}(P+P_0)/(K+P)$$

(12)

where $R_{D_{S=0}}$ is set as 0.588, $M$ to 0.118, $\alpha$ to 1.83, $\beta$ to −0.006, $P_0$ to 2.972, and $K$ to 5.657 [27]. $D_S$ is the soil organic carbon density at 20 cm depth, varies across different vegetation types; for specific values, refer to Chen et al. (2024) [25]. $T$ and $P$ represent the monthly average temperature and precipitation, respectively.

HR was calculated based on the empirical relationship between $R_S$ and $\mathrm{H}\mathrm{R}$ [5].

$$\mathrm{H}\mathrm{R}=-0.0009R_S^2+0.6011R_S+4.8874$$

(13)

where $R_S$ is the monthly soil respiration.

2.2.5. Evaluation of Model Performances

We assessed the accuracy of the phenology model based on two metrics: RMSE and Pearson correlation coefficient (r) [10,28]. The original CASA output ${NPP}_{org}$ and the phenology-modified CASA output ${NPP}_{mod}$ were compared with the MODIS gap-filled yearly NPP dataset (MOD17A3HGF) using the coefficient of determination (R²) and RMSE. Additionally, the performance of NEP estimation was also evaluated using R² and RMSE by comparing NEP flux data with NEP data produced by either the original CASA or the phenology-modified CASA [11].

2.2.6. Analysis

We calculated the trend of terrestrial NEP across China for each pixel using the Theil-Sen trend estimator and assessed the significance of the trends at the 0.05 level with the non-parametric Mann–Kendall test [29,30]. One-way analysis of variance (ANOVA) with Tukey’s HSD test was used to examine differences in the performance of phenology models as well as differences in NEP and its trends across different vegetation types and regions. Finally, we assessed the interannual trend of the difference between fossil fuel carbon emissions and ecosystem carbon sequestration in China’s four economic regions. All analyses were performed in Python.

3. Results

3.1. Performances Assessment of Models

Overall, the models that incorporated developmental factors outperformed the original phenology model (Figure 3a,b). The CDD_SG model achieved the highest average r values (0.10), significantly surpassing the other models, followed by the CDD_S (0.09), CDD_G (0.03), and CDD models (0.01). In terms of the prediction accuracy, the EOS simulated by the CDD_G model exhibited a significantly lower RMSE (14.27 days) compared to other models, followed by CDD_SG (14.95 days), CDD_S (15.32 days), and CDD (18.28 days). Based on these results, we introduced the EOS predicted by the CDD_SG model into the CASA model.

The NPP produced by the phenology-modified CASA model was more consistent with the MODIS NPP (Figure 3c,d). Specifically, the NPP simulated by the phenology-modified CASA model displayed a relatively higher R² (0.58 vs. 0.53) and lower RMSE (3615.5 vs. 3826.7 g C/m²/yr) compared to the counterparts of the original CASA model. Similarly, the following comparison of NEP based on the original CASA model and the phenology-modified CASA model against flux tower NEP observations revealed that the RMSE of the phenology-modified model decreased by 15.34% compared to the original model, while the R² increased from 0.27 to 0.47, representing an improvement of approximately 74% (Figure 3e,f).

3.2. Spatial Patterns of NEP Estimates

From 2003 to 2021, the total annual average NEP over China was 489.67 ± 38.27 Tg C/yr (mean ± standard error), with an average of 68.78 g C/m²/yr per unit area (Figure 4a,b). Of the pixels, 52.0% showed a multiyear average NEP greater than 0, serving as carbon sinks (Figure 4c). Spatially, stronger carbon sink capabilities were observed in the southeast, southwest, and the three northeastern provinces, generally exceeding 300 g C/m²/yr, while most of the Bohai Sea Rim area, Inner Mongolia, and the central region of the Qinghai–Tibet Plateau exhibited weak carbon sources (Figure 4a).

Across diverse vegetation types, the multiyear average NEP of EBF (600.53 g C/m²/yr) was significantly higher than that of other vegetation types, followed by DBF (348.64 g C/m²/yr) and SAV (236.48 g C/m²/yr) (Figure 4d). In contrast, SHB, GRA, and CRO functioned as carbon sources, displaying negative multiyear average NEP values of −113.34, −46.16, and −45.43 g C/m²/yr, respectively. From a regional perspective, all four economic regions in China played a role as carbon sinks from 2003 to 2021 (Figure 4e). The multiyear average NEP in the eastern region (174.17 g C/m²/yr) was the greatest, likely due to the highest forest coverage in China, followed by the northeastern, central, and western regions, with NEP values of 119.52, 101.59, and 32.49 g C/m²/yr, respectively.

3.3. Temporal Trends of NEP Estimates

Overall, this study found an increasing tendency of NEP in China with an average rate of 1.75 g C/m²/yr² from 2003 to 2021 (Figure 5a,b). Specifically, this increasing trend was identified in 73.30% of the pixels, with 20.32% passing the significance test (Figure 5c). Spatially, the increase in NEP was primarily concentrated in central regions, such as Guizhou, Chongqing, and Hunan, while most eastern coastal areas exhibited a declining trend. Further investigation revealed an increment of NEP across all vegetation types except for EBF and DNF. Among them, SAV exhibited the highest NEP growth rate of 3.40 g C/m²/yr², followed by MF (1.83 g C/m²/yr²) and DBF (1.81 g C/m²/yr²). Subregionally, all areas showed positive NEP trends during the study period, except the economically developed eastern region (−0.01 g C/m²/yr²), with the central region exhibiting the most pronounced increase at 2.21 g C/m²/yr².

By analyzing the trends in the difference between total apparent carbon emissions and ecosystem NEP, we identified significant regional disparities in carbon neutrality progress across China’s regions (Figure 6). The central region exhibited the fastest net carbon emission growth, reaching 40.8 Tg C/yr, suggesting that carbon emissions from fossil fuel consumption substantially outpaced the enhancement in ecosystem carbon sequestration. The western and eastern regions showed comparable net carbon emission growth rates of 36.9 and 36.1 Tg C/yr, respectively. In contrast, the northeastern region performed the most favorably, exhibiting both the lowest total net carbon emissions throughout the study period and the most modest growth rate at 6.8 Tg C/yr.

4. Discussion

Accurate phenological models are crucial tools for predicting vegetation phenological responses to future climate scenarios and their subsequent impacts on ecosystem functions [31,32]. Given the significant influence of developmental factors on autumn phenology, this study enhanced the traditional autumn phenology model by incorporating growing season GPP and SOS. The results showed that whether considering GPP or SOS alone or incorporating both factors, the modified autumn phenology models outperformed the original temperature-based model, significantly enhancing the simulation capability of autumn phenology (Figure 3). Traditional temperature-based models predict that under future warming conditions, vegetation productivity will increase due to extended growing seasons [8,33]. However, in temperate forests of the Northern Hemisphere, increased growing season productivity and earlier spring phenology triggered by rising CO₂ concentrations, higher temperatures, or changes in light levels may accelerate plant senescence, potentially offsetting the promotional effects of warming on growing season extension [9,15]. Therefore, current phenological models that do not consider developmental factors may seriously overestimate future growing season length, leading to significant biases in global carbon cycle estimates [8,9]. Future research should fully integrate developmental factors into predictions of plant phenology, which may revise our expectations of plant carbon sequestration capacity and avoid overly optimistic estimates of ecosystem carbon sink potential under future climate change.

In recent years, numerous studies have employed various methods to estimate the carbon sink capacity of China’s terrestrial ecosystems [4,5,25]. While estimates vary across methodologies, there is consensus that China’s terrestrial ecosystems act as a significant carbon sink [2,3]. Piao et al. (2022) [3] systematically compared estimates from different periods and methods, revealing that China’s terrestrial ecosystems exhibit a carbon sink capacity ranging from 0.12 to 1.11 Pg C/yr. Our study estimated an average NEP of 0.49 Pg C/yr for the period 2003–2021, which falls within the credible range but is slightly higher than estimates from ecosystem process models. This discrepancy likely stems from variations in temporal coverage, as China’s NEP has shown significant increases since the 21st century [34], while other studies include NEP data from before 2000, lowering the overall average in those estimates. On the other hand, the spatial distribution of NEP estimated in our study aligns with previous studies [2,4,25], with strong carbon sink areas primarily located in the southeastern and southwestern parts of China. Forests serve as the primary carbon sinks, while shrubs, grasslands, and croplands act as weak carbon sources (Figure 4). Forest ecosystems have high productivity and carbon storage capacity, and studies have shown that existing global forests alone can account for the land carbon sink [35]. In contrast, shrubs and grasslands have low biomass and limited carbon storage capacity, with carbon absorption and respiration being generally balanced, resulting in a weak carbon sink or even a carbon source [36,37].

Consistent with other studies, China’s terrestrial carbon sink showed a significant upward trend from 2003 to 2021 (Figure 5), likely driven by multiple factors. First, the fertilization effect of increased atmospheric CO₂ concentrations significantly enhanced plant photosynthesis efficiency, thereby promoting carbon uptake [17,38]. Second, in both the northern and southern forest regions, climate warming extended the growing season, which may have further strengthened carbon sequestration capacity [8]. Additionally, in southeastern China, increased anthropogenic nitrogen deposition has been shown to have a significant positive impact on plant growth [39], potentially providing crucial support for the enhancement of carbon sinks [40]. Finally, human interventions played a key role, such as large-scale afforestation projects contributing to over 70% of forest area expansion, natural forest protection, and ecological restoration programs, like the conversion of grazing land to grassland, which effectively promoted vegetation growth and enhanced ecosystem carbon sequestration capacity [41,42]. The combined effects of CO₂ fertilization, climate warming, increased nitrogen deposition, and human interventions collectively drove the continuous growth of China’s terrestrial carbon sink.

Our analysis of the residual emissions after offsetting industrial carbon emissions with the carbon sink of China’s terrestrial ecosystems revealed that, across all economic regions, the growth in ecosystem carbon sinks was smaller than the increase in carbon emissions from fossil fuel combustion during the same period (Figure 6). There were also regional differences in carbon neutrality levels. The economically developed eastern regions, including the Bohai Sea Rim and the Huang–Huai–Hai region, were far from achieving carbon neutrality. These areas were predominantly cropland with inherently weak carbon sink capacity, combined with high levels of industrial development and dense populations, leading to significantly higher apparent carbon emissions and residual emissions compared to other regions. In contrast, the northeastern region, with its relatively slower economic growth, showed a slower increase in residual emissions and exhibited better carbon neutrality levels than other regions. The western and central regions had relatively similar levels of carbon neutrality. Studies have suggested that as China’s forests gradually mature and age, the carbon sink capacity might decline, further reducing the ability of terrestrial ecosystems to offset industrial carbon emissions [3,43]. Therefore, achieving carbon neutrality requires the coordinated implementation of multiple strategies.

The NEP simulation framework we developed also has some uncertainties. First, discrepancies may exist between the actual start and end dates of plant photosynthesis and the SOS and EOS derived from the phenology model. Studies have shown that photosynthesis can continue until chlorophyll depletion in the leaves [44]. Second, ecosystem process models, such as CASA, do not account for human factors, like irrigation, fertilization, and forest management, all of which can significantly influence plant growth [3]. This adds uncertainty to the NEP estimates, and caution is needed when applying these models to agricultural ecosystems. Furthermore, ecosystem process models rely on remote sensing data, and the accuracy of these estimates is highly dependent on the quality of the remote sensing data [11]. The relatively coarse resolution of remote sensing data may also obscure NEP variations in specific plant types [22]. In the future, when extending our NEP accounting framework to larger study areas, special attention should be given to data quality, and multi-source data should be used for cross-validation to improve the reliability of the estimates. Additionally, more factors influencing plant growth should be incorporated into carbon sink accounting models to improve their comprehensiveness and accuracy.

5. Conclusions

This study enhanced NEP estimation by incorporating developmental factors (growing-season GPP and SOS) into autumn phenology model and using the modified phenology outputs to constrain the CASA model. The integration of developmental factors significantly improved the autumn phenology model’s performance, resulting in lower RMSE and higher correlation. The CASA model, constrained by phenological information, exhibited markedly better alignment with MODIS NPP products. Moreover, compared to the original model, the NEP produced by the phenology-modified CASA model achieved a 15.34% reduction in RMSE and a 74% improvement in R² against flux tower observations. From 2003 to 2021, the annual average total NEP of China’s terrestrial ecosystems was 489.67 ± 38.27 Tg C/yr and significantly increased with an average growth rate of 1.75 g C/m²/yr². Among vegetation types, evergreen broadleaf forests maintained the highest NEP, while shrublands exhibited the lowest values. However, the NEP growth rates across all four economic regions remained insufficient to offset the concurrent increases in industrial carbon emissions, with eastern China showing the highest net carbon emissions and facing significant challenges in achieving carbon neutrality. This study proposes a new and more accurate framework for NEP estimation, which provides valuable insights for understanding and predicting China’s carbon sink strength and offers scientific support for the formulation of carbon neutrality policies.