Estimating Global Gross Primary Production Using an Improved MODIS Leaf Area Index Dataset

Abstract

Remote sensing and process-coupled ecological models are widely used for the simulation of GPP, which plays a key role in estimating and monitoring terrestrial ecosystem productivity. However, most such models do not differentiate the C3 and C4 photosynthetic pathways and neglect the effect of nitrogen content on ${\mathrm{V}}_{\max }$ and ${\mathrm{J}}_{\max }$, leading to considerable bias in the estimation of gross primary productivity (GPP). Here, we developed a model driven by the leaf area index, climate, and atmospheric ${\mathrm{CO}}_2$ concentration to estimate global GPP with a spatial resolution of 0.1° and a temporal interval of 1 day from 2000 to 2022. We validated our model with ground-based GPP measurements at 128 flux tower sites, which yielded an accuracy of 72.3%. We found that the global GPP ranged from 116.4 $\mathrm{PgC}{\mathrm{year}}^{-1}$ to 133.94 $\mathrm{PgC}{\mathrm{year}}^{-1}$ from 2000 to 2022, with an average of 125.93 $\mathrm{PgC}{\mathrm{year}}^{-1}$. We also found that the global GPP showed an increasing trend of 0.548 $\mathrm{PgC}{\mathrm{year}}^{-1}$ during the study period. Further analyses using the structure equation model showed that atmospheric ${\mathrm{CO}}_2$ concentration and air temperature were the main drivers of the global GPP changes, total associations of 0.853 and 0.75, respectively, while precipitation represented a minor but negative contribution to global GPP.

1. Introduction

Gross primary productivity (GPP) refers to the cumulative organic carbon sequestered by plants through photosynthesis within a specific region, corresponding to a combination of the amount of photosynthetically produced products and the total amount of organic carbon sequestered by green plants through photosynthesis per unit of area and time [1,2,3,4]. GPP is one of the most fundamental indicators of ecosystem productivity and plays a central role in studying global carbon cycling and climate change [5]. It not only determines the initial influx of matter and energy into terrestrial ecosystems but also serves as an important metric for the characterization of plant activities and ecological functions.

The most widely used models for estimation of global GPP are divided into ecological process models (the Biome-BGC model and BEPS model) [6,7,8] and light-use efficiency (LUE) models (GLO-PEM, MODIS, CASA, C-FIX, VPM, and BEAMS) [2,9,10,11,12,13], each of which has its advantages and disadvantages. The light-use efficiency model calculates GPP mainly based on photosynthetically active radiation, temperature, and light-use efficiency parameters related to vegetation type [14,15,16]. However, the light-use efficiency model does not fully capture the physiological processes of plants and is only suitable for large-scale GPP estimates. Ecological process-based models based on remote sensing data were improved by the Farquhar model, which draws upon the principles of biochemistry and physics. These models have facilitated a deeper understanding and explanation of the biological mechanisms underlying photosynthesis. GPP product information can be found in Table 1.

The global distribution of C3 and C4 plants manifests distinct efficiencies in the assimilation of ${\mathrm{CO}}_2$ during photosynthesis, altering the regional atmospheric ${\mathrm{CO}}_2$ concentration and isotopic composition, thereby impacting the global carbon cycle [17]. Studies have revealed that C3 plants predominantly govern carbon absorption and GPP in temperate and high-latitude regions, whereas C4 plants assume significance in tropical and subtropical regions [18,19]. The distribution and influence of these two plant species might change considerably with climate change. C4 plants could expand their distribution and influence to more areas (including some traditional C3 plant areas), which could alter the carbon cycle and influence of these areas and, thus, affect the global carbon cycle and ecosystem productivity [17,20,21,22]. The nitrogen content in leaves influences the active amount and kinetic activity of RuBisCO, as well as the temperature-standardized rate of maximum RuBisCO carboxylation (${\mathrm{V}}_{\mathrm{cmax}25}$), thereby affecting the rate of plant photosynthesis [23].The terrestrial carbon cycle model is highly sensitive to the maximum carboxylation rate (${\mathrm{V}}_{\mathrm{cmax}}$) and the maximum electron transport rate (${\mathrm{J}}_{\max }$) derived from ${\mathrm{V}}_{\mathrm{cmax}25}$ during photosynthesis, which are influenced by leaf nitrogen content and plant nitrogen uptake [24]. However, current ecological models only couple the C3/C4 plant processes in the Community Land Model (CLM) and do not generate data products. Consequently, it is necessary to develop a model that can quickly and stably estimate global GPP on a large scale while coupling the C3/C4 processes and leaf nitrogen content with photosynthesis so as to reduce significant errors in the estimation of global GPP caused by current models.

The model we developed is driven by the leaf area index (LAI), distinguishes between C3 and C4 plants, optimizes the effects of nitrogen content on ${\mathrm{V}}_{\mathrm{cmax}}$ and ${\mathrm{J}}_{\max }$, and captures the physical and chemical information of the surface, improving the accuracy of GPP simulation. This study aims to (1) develop a moderate-resolution GPP dataset with a spatial resolution of 0.1 degree and a temporal resolution of one day, (2) use the above model to estimate the global GPP from 2000 to 2022, and (3) identify the key drivers of global annual GPP.

Table 1. GPP product information. EPM: ecological process model; LUE: light-use efficiency; ML: machine learning; VOD: vegetation optical depth; SIF: solar-induced fluorescence; Region: country/continent; Pathway: photosynthetic pathway.

Dataset	Method	Pathway	Temporal Resolution	Spatial Resolution	Unit	Study Period	Spatial Coverage	Reference
BEPS	EPM	C3	day	0.07°	gC/${\mathrm{m}}^2$	1981–2019	Global	Jv et al. (2021) [25]
BESS	EPM	C3	day	0.05°	gC/${\mathrm{m}}^2$	1982–2019	Global	Li et al. (2023) [26]
EC-LUE	LUE	C3/C4	8 days	0.05°	kgC/${\mathrm{m}}^2$	1982–2018	Global	Zhang et al. (2020) [4]
MODIS	LUE	C3/C4	8 days	500 m	kgC/${\mathrm{m}}^2$	2003–2022	Global	Running et al. (2021) [27]
VPM	LUE	C3/C4	8 days	0.05°	kgC/${\mathrm{m}}^2$	2000–2016	Global	Zhang et al. (2017) [3]
Random Forest	ML	C3	10 days	0.10°	gC/${\mathrm{m}}^2$	1999–2020	Global	Zeng et al. (2020) [28]
VODCA2GPP	VOD	C3	8 days	0.25°	gC/${\mathrm{m}}^2$	1988–2020	Global	Benjamin et al. (2022) [29]
GLO–PEM	LUE	C3/C4	8 days	10 m	gC/${\mathrm{m}}^2$	1981–2023	China	Stephen et al. (1995) [10]
GLASS	LUE	C3	8 days	0.05°	gC/${\mathrm{m}}^2$	1981–2018	Global	Liang et al. (2021) [30]
FluxSat v2.0	ML	C3	day	0.05°	gC/${\mathrm{m}}^2$	2000–2020	Global	Joiner et al. (2019) [31]
SMUrF	SIF	C3	4 days	0.05°	gC/${\mathrm{m}}^2$	2010–2019	Region	Wu et al. (2021) [32]
SiB4	EPM	C3/C4	Monthly	0.50°	gC/${\mathrm{m}}^2$	2000–2018	Global	Haynes et al. (2021) [33]
SMAP L4	EPM	C3	day	0.09°	gC/${\mathrm{m}}^2$	2015–2024	Global	Kimball et al. (2021) [34]
CARDAMOM	EPM	C3	day	0.50°	gC/${\mathrm{m}}^2$	2001–2016	USA	Yang et al. (2021) [35]
NIRv-Index	SIF	C3	day	0.05°	gC/${\mathrm{m}}^2$	1982–2018	Global	Wang et al. (2020) [36]
PML-V2	LUE	C3/C4	Monthly	0.05°	gC/${\mathrm{m}}^2$	1982–2014	Global	Zhang et al. (2020) [37]
PML-V2	LUE	C3/C4	8 days	0.05°	gC/${\mathrm{m}}^2$	2002–2019	Global	Chen et al. (2019) [38]
BCC-ESM1	EPM	C3/C4	Monthly	2.81°	gC/${\mathrm{m}}^2$	1850–2014	Global	Wu et al. (2020) [39]
CNRM-CM6-1	EPM	C3/C4	Monthly	1.41°	gC/${\mathrm{m}}^2$	1850–2014	Global	Program et al. (2019) [40]
Neural Network	ML	C3	4 days	0.05°	gC/${\mathrm{m}}^2$	2000–2022	Global	Zhang et al. (2018)  [41]
MuSyQ	LUE	C3	8 days	0.05°	gC/${\mathrm{m}}^2$	1981–2018	Global	Wang et al. (2021) [42]
Blue Carbon	EPM	C3	16 days	250 m	gC/${\mathrm{m}}^2$	2000–2019	USA	Fergin et al. (2020)  [43]

Table 1. GPP product information. EPM: ecological process model; LUE: light-use efficiency; ML: machine learning; VOD: vegetation optical depth; SIF: solar-induced fluorescence; Region: country/continent; Pathway: photosynthetic pathway.

Dataset	Method	Pathway	Temporal Resolution	Spatial Resolution	Unit	Study Period	Spatial Coverage	Reference
BEPS	EPM	C3	day	0.07°	gC/${\mathrm{m}}^2$	1981–2019	Global	Jv et al. (2021) [25]
BESS	EPM	C3	day	0.05°	gC/${\mathrm{m}}^2$	1982–2019	Global	Li et al. (2023) [26]
EC-LUE	LUE	C3/C4	8 days	0.05°	kgC/${\mathrm{m}}^2$	1982–2018	Global	Zhang et al. (2020) [4]
MODIS	LUE	C3/C4	8 days	500 m	kgC/${\mathrm{m}}^2$	2003–2022	Global	Running et al. (2021) [27]
VPM	LUE	C3/C4	8 days	0.05°	kgC/${\mathrm{m}}^2$	2000–2016	Global	Zhang et al. (2017) [3]
Random Forest	ML	C3	10 days	0.10°	gC/${\mathrm{m}}^2$	1999–2020	Global	Zeng et al. (2020) [28]
VODCA2GPP	VOD	C3	8 days	0.25°	gC/${\mathrm{m}}^2$	1988–2020	Global	Benjamin et al. (2022) [29]
GLO–PEM	LUE	C3/C4	8 days	10 m	gC/${\mathrm{m}}^2$	1981–2023	China	Stephen et al. (1995) [10]
GLASS	LUE	C3	8 days	0.05°	gC/${\mathrm{m}}^2$	1981–2018	Global	Liang et al. (2021) [30]
FluxSat v2.0	ML	C3	day	0.05°	gC/${\mathrm{m}}^2$	2000–2020	Global	Joiner et al. (2019) [31]
SMUrF	SIF	C3	4 days	0.05°	gC/${\mathrm{m}}^2$	2010–2019	Region	Wu et al. (2021) [32]
SiB4	EPM	C3/C4	Monthly	0.50°	gC/${\mathrm{m}}^2$	2000–2018	Global	Haynes et al. (2021) [33]
SMAP L4	EPM	C3	day	0.09°	gC/${\mathrm{m}}^2$	2015–2024	Global	Kimball et al. (2021) [34]
CARDAMOM	EPM	C3	day	0.50°	gC/${\mathrm{m}}^2$	2001–2016	USA	Yang et al. (2021) [35]
NIRv-Index	SIF	C3	day	0.05°	gC/${\mathrm{m}}^2$	1982–2018	Global	Wang et al. (2020) [36]
PML-V2	LUE	C3/C4	Monthly	0.05°	gC/${\mathrm{m}}^2$	1982–2014	Global	Zhang et al. (2020) [37]
PML-V2	LUE	C3/C4	8 days	0.05°	gC/${\mathrm{m}}^2$	2002–2019	Global	Chen et al. (2019) [38]
BCC-ESM1	EPM	C3/C4	Monthly	2.81°	gC/${\mathrm{m}}^2$	1850–2014	Global	Wu et al. (2020) [39]
CNRM-CM6-1	EPM	C3/C4	Monthly	1.41°	gC/${\mathrm{m}}^2$	1850–2014	Global	Program et al. (2019) [40]
Neural Network	ML	C3	4 days	0.05°	gC/${\mathrm{m}}^2$	2000–2022	Global	Zhang et al. (2018)  [41]
MuSyQ	LUE	C3	8 days	0.05°	gC/${\mathrm{m}}^2$	1981–2018	Global	Wang et al. (2021) [42]
Blue Carbon	EPM	C3	16 days	250 m	gC/${\mathrm{m}}^2$	2000–2019	USA	Fergin et al. (2020)  [43]

2. Methods and Data

2.1. Model Description

The direct and diffuse components of the total incident radiation were partitioned, with this partitioning process being determined by the proportion of cloud cover [44,45,46]. We applied two-stream radiative transfer approximation, which accounted for multiple scatterings within a finite canopy [44,46]. We implemented different scattering coefficients for direct and diffuse radiation, ensuring that the unique characteristics of the two radiation types were accurately represented [46]. The methods mentioned above allowed us to quantify the transmission and reflection factors for both direct and diffuse radiation, providing a detailed understanding of the interactions between canopy and radiation [46,47]. We integrated the two-leaf model with canopy radiation transfer to quantify the radiation absorption for both sunlit and shaded leaves [48,49]. The Ball–Berry–Leuning (BBL) stomatal conductance model was integrated into the photosynthesis process, optimizing the balance between ${\mathrm{CO}}_2$ uptake and water conservation [50]. Moreover, the calculation of the exact duration of global photosynthesis through sunrise, sunset, and local time computations facilitated a more accurate estimation of GPP. Finally, given the hourly time resolution of the data, accurately calculating the specific times for photosynthesis during sunrise and sunset in the respective time zone required precise determination of the daylight proportion in that zone (Appendix A).

We improved the corresponding variables of the Farquhar model for C3 plants [51] and the Collatz model for C4 [52] by integrating ${\mathrm{V}}_{\max }$, ${\mathrm{J}}_{\max }$, and Rd from the CLM model in the photosynthesis module [24,53]. The parameters of ${\mathrm{V}}_{\mathrm{cmax}25}$, ${\mathrm{J}}_{\max 25}$, and ${\mathrm{Rd}}_{25}$ were scaled over the canopy for sunlit and shaded leaves and were adjusted for leaf temperature. The maximum rate of carboxylation at 25 °C (${\mathrm{V}}_{\mathrm{cmax}25}$) varied with foliage nitrogen concentration, the fraction of leaf nitrogen in Rubisco, and the weight proportion of Rubisco relative to its nitrogen content [23]. A flow chart of the implementation of the proposed model by parallel computation with GPU is presented in Figure 1.

The proposed model incorporates seven input variables, but the forcing variables had to be resampled to align with the spatio-temporal resolution of the climate variables, as the spatio-temporal resolution for their metadata did not match (

Table 2. The input datasets of the proposed model. GPP: gross primary productivity simulated by the proposed model; C3/C4: the percentage of C4 plants in pixels; CI: clumping index; D2m: 2m dewpoint temperature; T2m: 2 m air temperature; SSRD: surface solar radiation downwards; LAI: leaf area index; Landcover: vegetation types.

Product	Spatial Resolution	Temporal Resolution	Study Period	Reference
GPP	0.1° × 0.1°	1-Day	2000–2022	(This study)
C3/C4	0.1° × 0.1°	1-Year	Mean	Still et al., 2003 [18]
CI	0.1° × 0.1°	1-Year	2001–2019	Fang & Wei, 2021 [54]
D2m	0.1° × 0.1°	1-Hour	2000–2022	Balsamo et al., 2015 [55]
T2m	0.1° × 0.1°	1-Hour	2000–2022	Balsamo et al., 2015 [55]
SSRD	0.1° × 0.1°	1-Hour	2000–2022	Muñoz Sabater, 2019 [56]
LAI	0.1° × 0.1°	1-Day	2000–2022	Barnes et al., 2003 [57]
Landcover	0.1° × 0.1°	1-Year	2000–2022	Barnes et al., 2003 [57]

).

2.2. Data Source

2.2.1. Data from Flux Tower

The FLUXNET2015 dataset is a comprehensive and valuable collection of eddy covariance measurements obtained from various ecosystems around the world. FLUXNET2015 represents a collaborative effort involving multiple research institutions and networks. It encompasses measurements from a wide range of ecosystems, including forests, grasslands, croplands, wetlands, and more. The FLUXNET community comprises datasets on energy fluxes, carbon fluxes, and meteorological variables collected and processed at various sites—crucial for studying global ecosystem function and response [3]. More information is available at the FLUXNET website (https://fluxnet.org/data/fluxnet2015-dataset/ (accessed on 28 September 2024)).

The GPP ($\mathrm{GPP}\_\mathrm{NT}\_\mathrm{VUT}\_\mathrm{REF}$) of the FLUXNET2015 dataset’s eddy covariance flux towers was used in comparison with our model’s GPP in this study. We estimated the global GPP values using our model and extracted the model-estimated GPP value of each corresponding point of the flux tower sites by latitude and longitude [4]. Then, we selected 128 flux tower sites around the world (Figure 2) to validate the results of our model by removing outliers. Information about the flux tower sites used in this study is shown in Appendix B.

2.2.2. Data Driving the Model

ERA-Interim is a widely used reanalysis dataset that provides comprehensive and high-quality global atmospheric and surface parameters spanning several decades [55,58]. ERA-Interim, developed by the European Centre for Medium-Range Weather Forecasts (ECMWF), offers a wealth of information on atmospheric variables such as wind, temperature, humidity, and pressure, as well as surface parameters like sea ice, sea surface temperature, and soil moisture. The dataset contains 69 global variables with a temporal and spatial resolution of 1 h and 0.1 degree, respectively. We applied two variables from the reanalysis dataset as driving variables for the model, including 2 m temperature (T2m; K) and 2 m dewpoint temperature (D2m; K).

Google Earth Engine (GEE), developed by Google, is a cloud-based geospatial computing platform designed to allow users to leverage Google’s powerful computing power and computing resources to analyze and process large amounts of geospatial data [59,60]. It integrates satellite imagery, geospatial datasets, and other Earth observation data to support a wide range of applications in environmental monitoring, land use analysis, disaster response, and more. The global surface solar radiation downwards (SSRD) at a 0.1-degree spatial resolution and 1-h temporal resolution was extracted from the dataset [56].

MODIS, which stands for Moderate-Resolution Imaging Spectroradiometer, is a state-of-the-art Earth observation instrument aboard the NASA Terra and Aqua satellites [57,61]. Its broad spectral coverage allows it to monitor a variety of Earth processes, including cloud cover, sea surface temperature, vegetation status, land cover changes, and more. We retrieved two variables from the dataset, namely leaf area index (LAI, ${\mathrm{m}}^2{\mathrm{m}}^{-2}$), with a temporal and spatial resolution of 4 days and 500 m, respectively, and vegetation types (Landcover), with a temporal and spatial resolution of one year and 500 m, respectively. We obtained LAI data with a temporal resolution of 1 day by removing outliers [62] and fitting the double logistic equation [63] and data with a spatial resolution of 0.1 degree by resampling.

The accurate simulation of water, carbon, and energy exchanges between the atmosphere and biosphere relies heavily on understanding the global distribution of C3 and C4 plants. The distinctive physiological and functional characteristics of these plant types play a pivotal role. To achieve this understanding, we utilized the 1-degree spatial distribution map of global C3 and C4 vegetation [17]. This incorporation of the distribution data enabled us to account for the diverse physiological processes inherent to C3 and C4 vegetation, thereby enhancing the precision of our simulations.

The vegetation clumping index (CI) quantifies the degree of deviation from the random distribution of leaves and serves as an important structural parameter of the canopy that controls photosynthesis and evapotranspiration processes in terrestrial ecosystems [64,65]. The global clumping index (CI), with a temporal and spatial resolution of 8 days and 500 m, respectively, was download from National Ecological Science Data Center [54].

2.3. Computation Platform

The proposed model was run a desktop PC with an Ubuntu 22.04 system equipped with 12 Intel i7-12700K cores running at 3.6 GHz with 128 GB of memory. The NAS delivers network storage solutions that excel in performance, fault tolerance, and data security, powered by an AMD Ryzen V1500B 2.2 GHz processor and 16 GB of memory. It offers a storage capacity of 128 TB, utilizing eight 16 TB hard disks (WUH721816ALE6L4-16TB) configured in a RAID10 array. Our model leverages OpenMP acceleration technology, enabling high-precision, large-scale parallel computations of various driving factors for GPP modeling.

2.4. Data Analysis

We analyzed the relevance of different models for the estimation of GPP and the sensitivity of GPP to environmental variables. The analyses were performed using the R version 4.3.3 programming language, along with the appropriate packages [66]. Pearson correlation analysis and the Shapiro–Wilk normality test were carried out using functions from the stats package [66], specifically utilizing the “cor.test” and “shapiro.test” functions. The structural equation model was built by employing the “psem” function, which is available in the piecewiseSEM package [67].

Structural Equation Modeling (SEM) is a statistical method that is used to analyze the relationships between several variables [67,68]. SEM combines elements of factor analysis and path analysis and enables the simultaneous investigation of relationships between observed variables and latent variables [67,68,69]. This approach allows for the modeling of complex multivariable relationships [70].

3. Results

3.1. Model Validation

We compared the mean annual GPP estimated from flux towers with the mean annual GPP simulated by the proposed model (Figure 3). We found that the proposed model accounted for 72.3% of the spatial variability in GPP across all validation sites. The proposed model performed very well at most sites, with a statistically significant p value < 0.01. The mean relative RMSE (RRMSE) and mean bias over all the sites were 36.3% and 9.23% $\mathrm{gC}{\mathrm{m}}^{-2}{\mathrm{year}}^{-1}$, respectively. The relative root mean square error (RMSE) values comparing the observed data and the simulated data for each site are shown in Appendix B.

3.2. The Dynamic of Global GPP from 2000 to 2022

The long-term trend of annual global summed GPP simulated by the proposed model, BEPS [71,72,73], MODIS [27], and VPM [3] shows an increase over the time series (Figure 4). The proposed model quantified the annual global GPP between 116.4 $\mathrm{PgC}{\mathrm{year}}^{-1}$ and 133.94 $\mathrm{PgC}{\mathrm{year}}^{-1}$, averaging 125.93 $\mathrm{PgC}{\mathrm{year}}^{-1}$ from 2000 to 2022. In contrast, the BEPS model produced annual global summed GPP estimates ranging from 117.04 $\mathrm{PgC}{\mathrm{year}}^{-1}$ to 129.92 $\mathrm{PgC}{\mathrm{year}}^{-1}$, averaging 122.98 $\mathrm{PgC}$${\mathrm{year}}^{-1}$ from 2000 to 2019. Additionally, the MODIS model simulated annual global summed GPP ranging from 95.28 $\mathrm{PgC}$${\mathrm{year}}^{-1}$ to 103.92 $\mathrm{PgC}{\mathrm{year}}^{-1}$, averaging 99.1 $\mathrm{PgC}{\mathrm{year}}^{-1}$ from 2003 to 2022. Lastly, another VPM model simulation of annual global summed GPP spanned 116.67 $\mathrm{PgC}{\mathrm{year}}^{-1}$ to 126.29 $\mathrm{PgC}{\mathrm{year}}^{-1}$, averaging 121.09 $\mathrm{PgC}$${\mathrm{year}}^{-1}$ from 2000 to 2017. We found a significant increasing trend for the proposed model, BEPS, MODIS, and VPM, with average increase rates of 0.548 $\mathrm{PgC}{\mathrm{year}}^{-1}$, 0.685 $\mathrm{PgC}{\mathrm{year}}^{-1}$, 0.34 $\mathrm{PgC}{\mathrm{year}}^{-1}$, and 0.53 $\mathrm{PgC}{\mathrm{year}}^{-1}$ globally, respectively. A significance test of the increasing trend showed that all p values were below 0.01. The GPP values estimated by the proposed model showed correlations of 0.767, 0.765, and 0.657 with those of the BEPS, VPM, and MODIS models, respectively. All significance test p values were below 0.01 (

Table 3. Statistical analysis of GPP for the proposed model, BEPS, VPM, and MODIS.

Model	${\mathbf{R}}^2$ (Compared to Our Study)	p Value (Compared to Our Study)	Linear Regressions	${\mathbf{R}}^2$	p Value
Our study	1	<0.001	$y=0.548x-976.06$	0.8633	<0.001
BEPS	0.767	<0.001	$y=0.685x-1253.23$	0.9758	<0.001
MODIS	0.657	<0.001	$y=0.34x-585.61$	0.8077	<0.001
VPM	0.765	<0.001	$y=0.536x-935.68$	0.9636	<0.001

).

3.3. The Spatial Pattern of Global GPP

The spatial patterns of the mean annual GPP simulated by the proposed model, BEPS, MODIS, and VPM are generally consistent (Figure 5). The proposed model estimated the mean annual GPP to range from 0 $\mathrm{KgC}{\mathrm{m}}^{-2}{\mathrm{year}}^{-1}$ to 4.0 $\mathrm{KgC}{\mathrm{m}}^{-2}{\mathrm{year}}^{-1}$, with an average of 0.837 $\mathrm{KgC}{\mathrm{m}}^{-2}{\mathrm{year}}^{-1}$ during the study period of 2000 to 2022 (Figure 5A). In contrast, the BEPS model produced an annual GPP of 0.002 $\mathrm{KgC}{\mathrm{m}}^{-2}{\mathrm{year}}^{-1}$ to 3.89 $\mathrm{KgC}{\mathrm{m}}^{-2}{\mathrm{year}}^{-1}$, with an average of 1.02 $\mathrm{KgC}{\mathrm{m}}^{-2}{\mathrm{year}}^{-1}$ from 2000 to 2019 (Figure 5B). Additionally, the MODIS model simulated the mean annual GPP to span from 0 $\mathrm{KgC}{\mathrm{m}}^{-2}{\mathrm{year}}^{-1}$ to 3.5 $\mathrm{KgC}{\mathrm{m}}^{-2}{\mathrm{year}}^{-1}$, with an average of 0.711 $\mathrm{KgC}{\mathrm{m}}^{-2}{\mathrm{year}}^{-1}$ from 2003 to 2022 (Figure 5C). Lastly, the VPM model simulated the mean annual GPP to range from 0 $\mathrm{KgC}{\mathrm{m}}^{-2}{\mathrm{year}}^{-1}$ to 4.29 $\mathrm{KgC}{\mathrm{m}}^{-2}{\mathrm{year}}^{-1}$, with an average of 0.88 $\mathrm{KgC}{\mathrm{m}}^{-2}{\mathrm{year}}^{-1}$ from 2000 to 2017 (Figure 5D). We observed that the peak annual GPP values occurred predominantly in tropical regions, particularly within the evergreen broadleaf forests of the Amazon and Southeast Asia, while the lowest GPP values were primarily situated in cold and arid areas.

The long-term trend of GPP was analyzed by the linear regression method (Figure 6). The global GPP showed an increasing trend of 72.7%, 85.7%, 68.1%, and 78.3% for the proposed model, BEPS, MODIS, and VPM, respectively (Figure 6). The statistically significant trends at a 95% confidence level were 46.03%, 62.03%, 30%, and 56.7% for the current model, BEPS, MODIS, and VPM, respectively. These significant increases were predominantly found in regions across Europe, Southeast Asia, and Africa. The significantly decreasing trends of global terrestrial production were 17.23%, 15.1%, 6.5%, and 21.7% for the proposed model, BEPS, MODIS, and VPM, respectively. These trends were distributed across various locations and particularly notable in tropical rainforests such as the Congo Basin and the Amazon. These spatial patterns in annual mean GPP trends, along with the updated estimates extending to 2022, are consistent with findings from previous studies [74].

The apparent discrepancy between the positive GPP trend from 2000 to 2022 (Figure 4) and the negative values in the global spatial distribution of GPP (Figure 6) could be clarified by distinguishing between temporal trends and spatial patterns. Figure 4 shows the overall global GPP trend, indicating an increase in average GPP over time. In contrast, Figure 6 shows the spatial distribution of GPP at 0.1-degree resolution, with individual pixels reflecting localized changes in structural, climatic, and physiological parameters of C3 and C4 vegetation. While the global average GPP trend is positive, localized declines occurred in certain regions due to factors such as climate change or land use change. Even though the global GPP trend is upward, the higher-resolution spatial data showed more local fluctuations, which do not contradict the overall positive trend.

4. Discussion

4.1. Environmental Characteristics of Flux Towers Suitable for Validating GPP

The flux sites we selected for comparison are characterized by relatively homogeneous land surface types, with minimal heterogeneity or topographical differences in the surrounding areas [75]. This uniformity helped to minimize potential errors caused by surface complexity when comparing lower-resolution global data with higher-resolution flux tower data [76]. In addition, the uniformity of terrain, vegetation types, and climatic conditions around the flux sites improved the feasibility and reliability of the comparison [77]. This analysis allowed us to more effectively assess the applicability and accuracy of global datasets in specific regions, providing a solid foundation for data model refinement and optimization [78].

In this study, we compared the environmental variables of the flux sites with those of the surrounding areas within a radius of 5 km, focusing on the distribution of elevation, leaf area index (LAI), and vegetation types. We compared the elevation of the flux sites with the average elevation in the surrounding 5 km and found that a correlation coefficient of 98.5% and statistical significance of p < 0.01 indicate a high degree of consistency (Figure 7A). We also compared the average annual LAI of the site with the average annual LAI of the area within 5 km, which resulted in a correlation coefficient of 98.5% and statistical significance of p < 0.01, also indicating a strong similarity (Figure 7B). We calculated the ratio of vegetation types at the site to vegetation types within a 5 km radius of the site (Figure 7C). The results showed that the average proportions for CRO, DBF, DNF, EBF, ENF, GRA, MF, OSH, SAV, WET, and WSA were 0.92, 0.81, 0.80, 0.90, 0.78, 0.92, 0.89, 0.90, 0.97, 0.90, and 0.86, respectively. This indicates that the flux site data adequately represent their spatial environment [79,80]. These results confirm the reliability of the flux site data and provide important evidence for its application in large-scale environmental studies [81]. These consistent environmental characteristics indicate that the climatic and ecological conditions in the region are relatively stable, which helps to minimize the fluctuations in meteorological data caused by terrain variations and ensure more reliable data collection at the flux towers.

4.2. Drivers of Global GPP Changes from 2000 to 2022

GPP quantifies the carbon uptake of plants through photosynthesis, which is influenced by factors such as temperature, available water, and ${\mathrm{CO}}_2$ concentration and is crucial in determining the overall productivity of plant ecosystems (Figure 8). Globally, the sensitivity of environmental variables to GPP was assessed and compared with 10 remote sensing models. It was found that the main cause of the increase in GPP was the positive response of plants to increases in atmospheric ${\mathrm{CO}}_2$ concentration, air temperature, and precipitation, with mean values of 23.9 ± 12.62 $\mathrm{PgC}{\mathrm{year}}^{-1}{100\mathrm{ppm}}^{-1}$, 6.655.3 $\mathrm{PgC}{\mathrm{year}}^{-1}{°\mathrm{C}}^{-1}$, and 2.62.03 $\mathrm{PgC}{\mathrm{year}}^{-1}{100\mathrm{mm}}^{-1}$, respectively [82]. The C-Fix model evaluated the impact of climatic factors on GPP trends, spanning from 1982 to 2015, revealing that the global average contributions of atmospheric ${\mathrm{CO}}_2$ concentration, air temperature, and water to the GPP trend were 65.37%, 13.07%, and 11.74%, respectively [83]. Atmospheric ${\mathrm{CO}}_2$ concentration was a key environmental factor that affected global GPP and made the largest positive contribution to its increase [84].

In this study, we investigated the correlation of global summed GPP with the climate variables of air temperature, atmospheric ${\mathrm{CO}}_2$ concentration, and precipitation to quantify the total effect of environmental variables on long-term changes in GPP. Over the last two decades, global trends have shown continuous increases in atmospheric ${\mathrm{CO}}_2$ concentration, air temperature, and precipitation [85], with increases of 2.19 ppm per year [86,87], 0.0405 °C per year [88], and 0.377 mm per year, respectively. The correlation between GPP and air temperature, atmospheric ${\mathrm{CO}}_2$ concentration, and precipitation was 0.866, 0.927, and 0.131, respectively (Figure 8). The emission of greenhouse gases such as ${\mathrm{CO}}_2$ contributed to global warming through the greenhouse effect and led to a prolonged growing season, especially at high and mid-latitudes, resulting in increased GPP [89]. The analysis of global process model results suggested that part of the increase in terrestrial carbon uptake could be attributed to increased vegetation productivity caused by the fertilization effect of elevated atmospheric ${\mathrm{CO}}_2$ concentrations [90,91,92]. Adequate precipitation effectively boosted GPP by enhancing soil moisture and plant growth, while excessive precipitation posed the risk of causing flooding, damaging plants, and inhibiting GPP due to altered soil properties.

Structural equation modeling (SEM) is a significant method for the analysis of ecological data that is capable of quantifying both direct and indirect causal relationships among multiple variables (Figure 9). SEM not only determines the individual path coefficient for each relationship but also assesses the overall model fit to provide a comprehensive understanding of natural systems [67,68,69]. A structural equation model was developed to explore the impacts of storm frequency on both the structure of kelp forest communities and the complexity of their food webs with a p value = 0.115 [67,68]. The SEM model was used to construct the direct response of species richness to abiotic stress and disturbance and the intervention effect on community biomass [70].

We were able to quantify both the direct and indirect paths of atmospheric ${\mathrm{CO}}_2$ concentration, air temperature, and precipitation for increasing GPP based on the SEM model (Figure 9). The total effects on GPP change, including the effects of atmospheric ${\mathrm{CO}}_2$ concentration, air temperature, and precipitation change, were 0.853, 0.75, and −0.144, respectively, when the direct and indirect effects of the latent variables were aggregated (Figure 9A). The increasing concentration of ${\mathrm{CO}}_2$ altered the surface energy balance and led to climate change, which subsequently affected atmospheric circulation and the global water cycle, changing the temporal and spatial distribution of precipitation [93,94]. The change in ${\mathrm{CO}}_2$ had not only a positive and direct effect, with an influence coefficient of 0.261, but also a positive and indirect effect, with an influence coefficient of 0.62, on the change in GPP through its positive effect on temperature, as well as a positive and indirect effect, with an influence coefficient of -0.028, on the GPP change through its negative effect on precipitation (Figure 9B).

The EC-LUE model assessed the impact of atmospheric ${\mathrm{CO}}_2$ concentration on GPP trends, discovering that a 100 ${\mathrm{ppm}}^{-1}$ rise in atmospheric ${\mathrm{CO}}_2$ concentration significantly boosted the global GPP by 12.31 ± 0.61 PgC [4]. While precipitation made a minor but negative contribution to global GPP, the VPM, VI (Vegetation Indices), TG (Temperature–Greenness), and GR (Greenness–Radiation) models had respective mean values of −1.4949 $\mathrm{PgC}{\mathrm{year}}^{-1}{100\mathrm{mm}}^{-1}$, −0.85859 $\mathrm{PgC}{\mathrm{year}}^{-1}{100\mathrm{mm}}^{-1}$, −0.85859 $\mathrm{PgC}{\mathrm{year}}^{-1}{100\mathrm{mm}}^{-1}$, and −0.36364 $\mathrm{PgC}{\mathrm{year}}^{-1}$${100\mathrm{mm}}^{-1}$ [82]. The VPM model focuses on radiation and temperature in GPP estimation and neglects precipitation, while the VI and TG models focus on vegetation indices and temperature and indirectly reflect the effects of precipitation. The GR model, which relies heavily on radiation, associates increased precipitation with lower radiation and GPP. The global GPP has a negative impact at the global level but not at the regional level, where the increase in precipitation depends on the climatic characteristics of the region, the water requirements of the ecosystems, and interaction with other relevant environmental variables. At the global scale, an increase in precipitation could lead to cloud thickening and reduced solar radiation, as well as waterlogged soils that limit the oxygen supply to roots, thereby inhibiting plant photosynthesis. At a regional scale, the effects of precipitation on GPP vary depending on the type of vegetation, topography, and soil conditions. GPP changes are influenced primarily by an increase in radiation and temperature at mid- and high latitudes, by changes in moisture conditions at low latitudes, and by fluctuations in both temperature and moisture conditions at high altitudes [95]. Changes in photosynthetically active radiation, relative humidity, and temperature were the primary drivers of the interannual variability in GPP, accounting for 80% in deciduous broadleaf forests and evergreen coniferous forests, 65% in crop plants, and 58% in shrubs [96]. Different plants have different light saturation points and radiation utilization rates, which means that the radiation intensity required to achieve maximum photosynthetic efficiency varies from plant to plant. Temperature affects the activity of key enzymes such as Rubisco, influencing the rate of photosynthesis. Humidity affects the transpiration and stomatal conductance of plants, which, in turn, affect ${\mathrm{CO}}_2$ uptake and water use efficiency.

5. Conclusions

We estimated global GPP from 2000 to 2022 based on GPU by integrating the Farquhar model for C3 plants and the Collatz model for C4. The proposed model showed strong performance in simulating both spatial and interannual variations in GPP globally. It explained 72.3% of the spatial variation at all validation sites and showed interannual changes consistent with other models. We assessed the contributions of climate factors to GPP trends and found that these environmental variables exhibit substantial long-term changes and contribute significantly to vegetation production on an interannual scale. Atmospheric ${\mathrm{CO}}_2$ concentration is an important environmental factor that increases global annual GPP according to the SEM model. The model we developed was able to effectively capture temporal and spatial changes in GPP and provides a reliable alternative for long-term estimates of global GPP.

While the Light-Use Efficiency (LUE) model has been widely used to estimate vegetation productivity, it has encountered several limitations, including the failure to account for environmental stressors, inconsistencies in spatial and temporal scales, region-specific parameterization, and the inability to fully capture the dynamic interactions within complex ecosystems [97,98,99,100]. Although vegetation index models have been widely used to monitor vegetation condition and productivity, they are limited by their sensitivity to atmospheric conditions, soil background effects, and saturation in densely vegetated areas, which reduces their accuracy in estimating biophysical parameters [2,101]. AI models for estimating gross primary production (GPP) often encounter limitations due to their dependence on the quality and quantity of input data, the complexity of ecosystem processes, and difficulties in generalizing across different biomes, leading to uncertainties in model predictions [102,103]. Ecosystem process-based models for estimating GPP have the advantage of integrating detailed physiological and ecological mechanisms, which enable more accurate predictions of carbon fluxes at different temporal and spatial scales [81,104,105]. Our dataset provides a higher temporal resolution than the Blue Carbon, SiB4, BCC-ESM1, and CNRM-CM6-1 datasets; higher spatial accuracy than CARDAMOM, BCC-ESM1, and CNRM-CM6-1; and a longer temporal span than BEPS, BESS, Blue Carbon, SiB4, SMAP L4, and CARDAMOM.

The proposed model generated a global GPP dataset with a resolution of 10 km, but this moderate spatial resolution led to the loss of important fine-scale details, resulting in an oversimplified representation of heterogeneous landscapes and lower accuracy in detecting small-scale ecological processes and land use changes. We developed the model using parallel computing techniques, including GPUs and multithreading, which significantly enhanced the simulation speed. Users applying this model to simulate global GPP are required to set up the development environment and possess proficiency in computer skills, such as C++ programming standards. These technical demands pose considerable challenges for beginners. Furthermore, since the model was designed for a global scale, recalibration of parameters is necessary to improve simulation accuracy when applying it to smaller-scale GPP simulations.

For future studies, we plan to improve the spatial accuracy of GPP simulation and more accurately reflect the characteristics of surface changes. We plan to further develop the net primary production (NPP) and net ecosystem production (NEP) modules to enable a more comprehensive simulation of the carbon cycle. These improvements are crucial for understanding the dynamics of the carbon cycle in ecosystems and the impact of climate change on vegetation productivity. We also plan to integrate additional variables, optimize computational algorithms, and improve spatial resolution to increase accuracy and efficiency. In addition, developing modular features, creating a user-friendly interface, increasing calibration efforts, and open-sourcing the model would expand its applicability and ensure reliability in different ecosystems and under different climate conditions. We anticipate that these improvements will enable the model to play an important role in broader ecological research and climate policy development.