1. Introduction
Net primary productivity (NPP) is the remaining portion of the total organic matter produced by photosynthesis per unit area of green vegetation per unit time, excluding its autotrophic respiration [
1]. It straightforwardly indicates the productivity and quality of an ecosystem in the context of natural ambient conditions, and is one of the significant elements in deciding the carbon source/sink of an ecosystem [
2,
3]. In the current context of global climate change, the accurate and timely estimation of the NPP in China’s land ecosystems is critical to implementing the “Double Carbon Strategy” and the advancement of ecological civilization in the country [
4].
The integration of remotely sensed data and ecological models has emerged as the primary method for large-scale vegetation NPP estimation owing to its broad coverage, low acquisition cost, and short revisit period. The models currently used for NPP estimation include statistical models, process models, and light energy efficiency models [
5]. Statistical models typically rely on the vegetation index (VI) and leaf area index (LAI) for estimation, and the model data are relatively accessible. However, the physiological and ecological mechanisms underlying these models are not always well understood, which can lead to errors in the estimation results [
6]. While process-based models, for example the Biome BioGeoChemical Cycles model (BIOME-BGC), Boreal Ecosystem Productivity Simulator (BEPS), Terrestrial Ecosystem Model (TEM), and Integrated Terrestrial Ecosystem C-budget model (InTEC), have clear mechanisms and produce more accurate NPP estimates, their model structures are relatively complexity and require numerous parameters, which can be challenging to scale up to the regional level [
7,
8,
9,
10]. Light energy efficiency models, such as the Carnegie–Ames–Stanford Approach (CASA) and Global Production Efficiency Model (GLO-PEM), have been developed [
11,
12,
13], of which the CASA model is based on the theory of plant photosynthesis mechanism and has minimal dependence on ground measurement data. Additionally, the CASA model is relatively straightforward, with few input parameters and high ease of use and applicability, making it one of the most commonly utilized models for NPP estimation across various spatial scales [
14].
Despite these advantages, the CASA model has deficiencies in simulating vegetation NPP, such as uncertainties in the estimation of vegetation light energy utilization. To address this, researchers have attempted to compensate for these deficiencies by constructing new water stress factors or correcting the maximum light energy utilization for ideal contexts, thereby improving the accuracy of vegetation NPP estimation [
15,
16,
17]. Additionally, the nonlinear relationship between NDVI and vegetation NPP can be a limitation for accurate estimation of NPP using the CASA model [
18,
19]. However, the use of NDVI to estimate FPAR has two main issues. First, the relationship between NDVI and vegetation NPP is nonlinear and easily reaches saturation, which may affect the accuracy of NPP estimation. Second, NDVI is mainly reflective of vegetation leaf conditions rather than actual photosynthesis. Several studies have proposed using alternative vegetation indices that are less affected by saturation, such as the NDVI
2 [
20], red-edge normalized vegetation index [
21], and enhanced vegetation index (EVI) [
22]. While these methods may improve the estimation accuracy of vegetation NPP to an extent, they cannot fundamentally address the nonlinear relationship and saturation issues between vegetation indices and NPP. Solar-induced chlorophyll fluorescence (SIF) is a promising alternative to traditional reflectance-based vegetation indices, as it directly measures the fluorescence emitted by plants during photosynthesis [
23]. Recently, SIF has been increasingly used as a tool for estimating vegetation carbon stocks due to its ability to directly measure plant photosynthetic activity. However, there remains much debate among scholars regarding the relationship between SIF and vegetation green biomass. Additionally, the spatial resolution of SIF is limited, with values ranging from 40 × 80 km
2 to 0.05°, which makes it difficult to obtain detailed information on green vegetation biomass [
24]. Therefore, the construction of a new remotely sensed vegetation index to improve the accuracy of NPP estimation by solving the problem of linearity and saturation is a pressing issue.
Previous research has attempted to improve the performance of vegetation indices using empirical or parameter transformations. However, these methods only construct new vegetation indices that try to utilize information from bands other than the NIR and red bands, and it is not clear whether all available information is being used. Moreover, these methods fail to address the saturation problem that occurs in vegetation indices. In contrast, the kernel function approach takes a machine learning perspective and can derive nonlinear approaches from linear algorithms while remaining reliant on linear arithmetic for operations [
25]. This approach is universal, meaning that any index of kernel vegetation embedded in a Gaussian kernel can optimize the use of spectral information, leading to lower uncertainty in the propagation of kernelized vegetation indices compared to non-kernelized ones.
Therefore, in this study, a new vegetation index is used to address the saturation phenomenon in estimating the vegetation NPP of Chinese terrestrial ecosystems over the past two decades, then the Theil–Sen median and Mann–Kendall trend are used to explore the trends of vegetation NPP in Chinese terrestrial ecosystems. Second-order partial correlation, multiple correlation analysis, the differential equation method, and the PLS-SEM method are used to analyze the role of each influencing factor on the change in vegetation NPP in Chinese terrestrial ecosystems. This study can provide a valuable scientific reference for ecological environmental conservation, carbon cycle, and resource allocation research in China.
5. Discussion
5.1. Relationship between Drivers and NPP Changes at Each Stage
The driving factors affecting vegetation NPP changes may differ at different stages, and the main contributors may differ for each time period. In this study, the relationships between each driver and vegetation NPP changes were analyzed using a structural equation PLS-SEM with a five-year cycle by integrating soil, climate, socioeconomic, and topographical factors.
This research first validated the reliability and validity of the PLS-SEM model (
Table 5). VIF was used to detect multicollinearity in the variables and provided values between 1 and 5, indicating that there was no significant covariance between the factors. A composite reliability (CR) greater than 0.7 and not lower than 0.6 indicated that the model was reliable, and an average variance extracted (AVE) greater than 0.5 demonstrated that the model was valid. The data in this study satisfied the above conditions well and prove that the model was reasonable and reliable [
42].
The PLS-SEM model of the relationship between the latent and explicit variables for the changes in NPP in 2001–2020 is shown in
Figure 20. The results show that changes in human activities always had a negative impact on NPP changes, with path coefficients of −0.017, −0.075, −0.041, and −0.006 for 2001–2005, 2005–2010, 2010–2015, and 2015–2020, respectively. For the periods 2001–2005 and 2015–2020, topographical change had a positive impact, with path coefficients of 0.096 and 0.234, respectively. Soil change had a positive impact in 2001–2005, 2010–2015, and 2015–2020, with path coefficients of 0.078, 0.463, and 0.220, respectively. In addition, the path coefficients and main impact factors of the climate changed. The impact of climate (evaporation and precipitation) on NPP changes in 2001–2005 was positive (
β = 0.375). The impact of climate (evaporation and temperature) was positive (
β = 0.092) in 2005–2010 and became negative in 2010–2015. The impact of climate on NPP changes (evaporation and solar radiation) was positive (
β = 0.151) for the period of 2015–2020.
The amount of fertilizer used has an effect on vegetation NPP changes; unfortunately, raster data on fertilizer inputs such as N and P were lacking in this study. In a forthcoming follow-up study, we intend to obtain raster data on various types of fertilizer inputs to enhance the accuracy and reliability of the influence factors of vegetation NPP.
5.2. Comparison with Other NPP Reports
The data sources used in NPP studies can be broadly divided into two categories: existing NPP products [
43,
44,
45,
46,
47] (
Table 6), which have a fixed spatial resolution, and estimations of NPP using NDVI and FPAR. To analyze the relationship between NPP changes and influencing factors, studies have mainly explored the overall response mechanism of NPP to climate change and human activities; in addition, they have considered mainly the effects of temperature and precipitation [
36,
43,
46,
48,
49]. Other studies have considered additional influencing factors, such as relative humidity, sunshine hours, chlorophyll-a, photosynthetic availability, wind stress curl, salinity, nitrates, phosphates, and CO
2 concentration [
44,
47,
50]. Compared with other studies, this paper conducted a relatively comprehensive study of the climatic and anthropogenic influences on vegetation NPP; however, other influencing factors, such as CO
2 concentrations, nitrates, and phosphates, could be considered in subsequent studies.
5.3. Limitations and Prospects
Although the accuracy of kNDVI estimates at the site scale was higher than that of NDVI, unfortunately, ChinaFLUX shared flux observations only for certain stations. This makes it difficult to rely on observational data alone when assessing the quality of NPP data at the national scale. In the future, more measured data will be sought as much as possible to improve the quality of the estimated NPP data.
For accuracy validation, direct real NPP data were not available in this research, and were obtained by converting flux data. It is worth noting that the value of the NPP/GPP ratio α used in conversion varies with geography, meteorology, and vegetation type, while in this study we used the regional average NPP/GPP ratio (α = 0.64), which could potentially have an impact on the NPP conversion value.
6. Conclusions
Based on an improved CASA model driven by the new kNDVI vegetation index, this paper estimated the NPP of China’s terrestrial ecosystem vegetation from 2001 to 2020 and used the Theil–Sen median, Mann–Kendall trend, second-order partial correlation, multiple correlation analysis, differential equation method, and PLS-SEM method to analyze the temporal and spatial evolution of vegetation NPP in China’s terrestrial ecosystems. The results show that kNDVI can effectively improve the saturation problem, cope with complex phenological changes, and improve the precision of NPP estimation. The NPP of Chinese vegetation presented a fluctuating upward trend on the time scale from 2001 to 2020. It increased by a mean of 0.7235 g C/m2 each year, showing a low northwest to high southeast pattern in the spatial distribution, with the highest average multi-year NPP value by province in Taiwan Province and the lowest in Xinjiang Uygur Autonomous Region. The changes in NPP were dominated by an increasing trend (72.96%), and the areas showing significant and very significant increases were concentrated in the Northeast Plain, Loess Plateau, and Sichuan Basin. In Northeast and North China, vegetation NPP was mainly influenced by precipitation, while in Southwest and South China it was mainly influenced by temperature and in Central China by solar radiation. With the passage of time, the dominant factors influencing the changes in NPP change and have different impacts. The main influencing factors of the climate in 2001–2005 were evaporation and precipitation, and the main influencing factors of the climate in 2005–2010 were evaporation and temperature; moreover, the influences of both phases were positive. However, in 2010–2015 the impact of the climate was negative. Climate change and human activities have a major positive effect on NPP change. The areas where the NPP increased due to climate change were mainly located in Northeast and North China, and the areas where the NPP increased due to human activities were mainly located in Central and South China. The results of this research can be used in China and around the world to better prepare for and respond to climate change.