Increasing Ecosystem Fluxes Observed from Eddy Covariance and Solar-Induced Fluorescence Data

Abstract

Ecosystems modulate Earth’s climate through the exchange of carbon and water fluxes. However, long-term trends in these terrestrial fluxes remain unclear due to the lack of continuous measurements on the global scale. This study combined flux data from 197 eddy covariance sites with satellite-retrieved solar-induced chlorophyll fluorescence (SIF) to investigate spatiotemporal variations in gross primary productivity (GPP), evapotranspiration (ET), and their coupling via water use efficiency (WUE) from 2001 to 2020. We developed six global GPP and ET products at 0.05° spatial and 8-day temporal resolution, using two machine learning models and three SIF products, which integrate vegetation physiological parameters with data-driven approaches. These datasets provided mean estimates of 128 ± 2.3 Pg C yr⁻¹ for GPP, 522 ± 58.2 mm yr⁻¹ for ET, and 1.8 ± 0.21 g C kg⁻¹ H₂O yr⁻¹ for WUE, with upward trends of 0.22 ± 0.04 Pg C yr⁻² in GPP, 0.64 ± 0.14 mm yr⁻² in ET, and 0.0019 ± 0.0005 g C kg⁻¹ H₂O yr⁻² in WUE over the past two decades. These high-resolution datasets are valuable for exploring terrestrial carbon and water responses to climate change, as well as for benchmarking terrestrial biosphere models.

1. Introduction

Terrestrial ecosystems influence the land–atmosphere interactions of carbon and water fluxes through surface physical and biological processes [1,2,3]. Gross primary productivity (GPP) represents the total carbon gains by plant photosynthesis, while evapotranspiration (ET) accounts for the total water losses, including soil evaporation, canopy intercepted evaporation, and plant transpiration [4,5]. As the representation of terrestrial carbon and water interactions, water use efficiency (WUE) is calculated as the ratio of GPP and ET to indicate coupling relationships between carbon gains and water losses. The spatiotemporal variations in GPP, ET, and WUE substantially influence carbon and water cycles from regional to global scales [6]. Therefore, the accurate quantification of GPP, ET, and WUE globally is vital for advancing our understanding of terrestrial carbon and water dynamics and their interactions under the changing climate. Moreover, understanding long-term trends in WUE is also critical for predicting ecosystem responses to global change, improving Earth system models, and informing policies on climate mitigation, adaptation, and water sustainability.

Different approaches have been applied to quantify regional/global GPP and ET, including process-based models [7], semi-mechanistic models [8,9], and machine learning approaches [10,11,12]. Several machine learning (ML) methods for estimating GPP already exist, such as cost-effective and scalable ML frameworks for GPP modeling that combine remote sensing and flux tower data; hybrid models that merge machine learning techniques with process-based or light-use efficiency (LUE) models; and automated machine learning (AutoML) [13,14]. Each of these approaches has its advantages and disadvantages [2]. Process-based models, such as the Breathing Earth System Simulator [7] and Boreal Ecosystem Productivity Simulator [15], can simultaneously simulate regional GPP and ET based on land surface physical and chemical mechanisms [16]. However, their performance is highly dependent on parameterization schemes and model structures [17,18]. Semi-mechanistic models, such as the Light Use Efficiency (LUE) model and the Penman–Monteith equations, simulate regional GPP and ET based on satellite observations and empirical relationships [8]. However, establishing uniform constraint equations across different ecosystems is challenging, and these models are limited in their application to climate or vegetation change scenarios. Machine learning (ML) algorithms have also been applied to derive GPP and ET using ground- and satellite-based observations [19,20,21,22]. Validations have shown that these data-driven approaches often outperform traditional process-based and semi-mechanistic models [18,23]. However, most of these ML algorithms have not taken advantage of recent advances in remote sensing vegetation parameters, such as solar-induced chlorophyll fluorescence (SIF), which represents emitted electromagnetic radiation related to photosynthesis. Furthermore, only a few studies (e.g., FLUXCOM) have focused on the coupled relationships between carbon and water fluxes, such as water use efficiency (WUE), based on simultaneously derived GPP and ET.

SIF is a spectral signal (650–800 nm) emitted by the photosynthetic center of plants under solar illumination, reflecting the complex physiological functions of terrestrial plants [24]. SIF has been widely regarded as a direct indicator of plant photosynthesis [25,26,27,28] as it represents an energy flux emitted from plant chlorophyll molecules a few nanoseconds after light absorption [29]. Terrestrial GPP often shows strong correlations with SIF for almost all biomes in unstressed conditions [30], offering a more effective means to monitor photosynthesis compared to traditional vegetation indices like the Normalized Difference Vegetation Index and Enhanced Vegetation Index [28,29,31,32,33]. Given the physiological coupling between plant photosynthesis and transpiration, some studies have also shown that SIF could be used to estimate regional ET based on empirical SIF-ET relationships [34,35]. Zhou et al. [36] demonstrated that SIF-based ET models outperformed other satellite-based methods. However, the tight relationship between SIF and ET is influenced by environmental factors such as air temperature and vapor pressure deficit (VPD), which are linked to soil and vegetation water evaporation [37,38,39].

Considering the limitations of simultaneous carbon and water fluxes datasets, we developed global GPP, ET, and WUE datasets (referred to as ECSIF) by combining site-level Eddy Covariance measurements with satellite-retrieval SIF products from 2001 to 2020. We first explored the relationships between GPP (or ET) and three SIF products for different plant functional types (PFTs) at the global FLUXNET sites. We then constructed and validated sixty PFT-specific machine learning models for GPP and ET, respectively, using ground measurements, two ML algorithms, and three SIF products across ten PFTs. Thereafter, we generated six ECSIF datasets of global GPP, ET, and WUE at 8-day temporal resolution and 0.05° spatial resolutions, based on gridded SIF, meteorological, and land cover products. The main objectives of this study are to (a) validate the effectiveness of SIF products in deriving site-level GPP and ET, (b) generate SIF-based GPP, ET, and WUE products with high spatiotemporal resolution on the global scale, and (c) explore the spatiotemporal variations in GPP, ET and their coupling WUE over the past two decades.

2. Materials and Methods

2.1. Site-Level Observations

We selected 197 sites from the FLUXNET2015 dataset for model development and validation, each with at least 1-year consistent observations of carbon and water fluxes, as well as meteorological measurements after 2001 (Figure S1 and Table S1). These sites were categorized into ten PFTs according to the International Geosphere-Biosphere Program (IGBP) land cover classification scheme. At each site, meteorological variables including total shortwave radiation (SW, W m⁻²), surface air pressure (Ps, hPa), air temperature (Tair, K), total precipitation (Pre, mm), wind speed (Wind, m s⁻¹), and vapor pressure difference (VPD, kPa) were observed at half-hour intervals. In addition, carbon and water fluxes, such as net ecosystem exchange (NEE) and latent heat (LH), were measured simultaneously. In this study, we used the carbon fluxes of GPP (μmol CO₂ m⁻² s⁻¹) derived from the mean values of daytime (GPP_DT_VUT_MEAN) and nighttime (GPP_NT_VUT_MEAN) fluxes based on the NEE-partitioning methods. For water fluxes, ET (mm hr⁻¹) was calculated from the observed LH using the following method [40]:

$$ET={\displaystyle \frac{LH}{a·{\rho }_{water}}}$$

(1)

Here, $a$ is the latent heat of the vaporization of water (2.45 KJ g⁻¹), and ${\rho }_{water}$ is the density of water (1 g cm⁻³). All datasets were aggregated into 8-day timescales to build ML models.

2.2. Gridded Observations

For the gridded data, we selected meteorological data, which include SW, Ps, Tair, Pre, Wind, and VPD, from the Modern-Era Retrospective analysis for Research and Applications version 2 (MERRA-2) reanalysis. In addition, we calculated direct (Direct) and diffuse (Diff) shortwave radiation by combining the shortwave radiation from MERRA-2 with the satellite-based diffuse radiative fraction (DF) provided by Ryu et al. [41]:

$$Direct={SW}_{MERRA-2}\times (1-DF)$$

(2)

$$Diff={SW}_{MERRA-2}\times DF$$

(3)

We also calculated site-level direct and diffuse SW using the same methods. Atmospheric carbon dioxide (CO₂) concentration data were obtained from measurements at the Mauna Loa Observatory. Global annual land cover data were obtained from the MCD12C1 product of the MODIS retrieval. All the gridded meteorological data were downscaled to a spatial resolution of 0.05° and an 8-day temporal resolution for the period of 2001–2020 on a global scale.

To evaluate the derived gridded GPP and ET datasets, we used the products from MODIS [42,43], PML-V2 [5], and X-BASE [44] as the basis for comparisons. Based on the radiation-use efficiency concept, MODIS GPP (MOD17A2H) was developed combining the Leaf Area Index (LAI) and Fraction of absorbed Photosynthetically Active Radiation (FPAR) with meteorological data. MODIS ET (MOD16A2) was estimated using an improved ET algorithm [45] based on the Penman–Monteith equation, which considered soil evaporation, canopy intercepted evaporation, and plant transpiration. PML-V2 GPP and ET were estimated using the PML-V2 model, which incorporates water–carbon coupled canopy conductance and CO₂ fertilization and were driven by meteorological and remote sensing data. X-BASE GPP and ET were upscaled from eddy covariance flux data using a combination of several ML algorithms and meteorological forcing data [44].

2.3. SIF Products

We selected three recently developed SIF products, including CSIF [46], GOSIF [47] and RTSIF [30], as model inputs. Specifically, the CSIF product was obtained by combining neural network models with MODIS-based surface reflectance and Orbiting Carbon Observatory-2 (OCO-2) SIF observations [46]. GOSIF was developed using a data-driven (i.e., ML) approach combining discrete OCO-2 SIF soundings, remote sensing data from MODIS, and meteorological reanalysis over the period 2001–2023 [47]. RTSIF was constructed using the XGB algorithm based on TROPOMI SIF [30], the nadir bidirectional reflectance distribution function adjusted reflectance [48], land surface temperature, photosynthetically active radiation, and land cover. The CSIF datasets were aggregated into a high spatial resolution of 0.05° and 8-day temporal resolution to match the GOSIF and RTSIF datasets.

2.4. Machine Learning Algorithms

We used two ML algorithms, eXtreme Gradient Boosting (XGB) and Random Forest (RF), to develop data-driven models. The XGB algorithm is an updated version of the Gradient Boosted Decision Tree (GBDT) algorithm [49] and can build augmented trees to handle complex nonlinear relationships between inputs and outputs. In addition, the XGB algorithm adds regularization terms to the loss function, effectively preventing overfitting during the training process [50]. The RF is an integrative algorithm that randomly samples features during the training of each tree, rather than using all features. This randomness increases model diversity and reduces over-reliance on the same features for each tree [51]. Both the XGB and RF algorithms are based on decision tree models and can offer excellent accuracy and efficient operation on large datasets without overfitting [52,53,54]. Due to the feature selection process, decision trees are inherently resistant to the covariance between input variables. Specifically, when two input features are highly correlated, the decision tree model will select only one of them as a splitting feature during the decision process [55]. Recently, the XGB model has been used to identify the main climatic drivers of GPP variability at global FLUXNET sites and to quantify reasonable responses of GPP to various climatic factors such as VPD [56]. Similarly, RF models have been used to estimate regional GPP for rice–wheat rotation croplands, showing better performance than traditional methods [57,58].

We constructed site-level ML models for each PFT using available observations (Figure 1), including nine inputs (SIF and climatic drivers) and two outputs (GPP or ET). For each output (GPP or ET), a total of sixty models were built using one of the two ML algorithms (XGB or RF) by combining one out of three SIF products (CSIF, GOSIF, or RTSIF) with site-level observations across ten PFTs. Observations from all sites within each PFT were randomly divided into training (80%) and validation (20%) datasets to develop PFT-dependent models. During the training process, we optimized model hyperparameters using cross-validation and hyperparameter tuning to identify the optimal parameter combinations. Specifically, the hyperparameters were set to several candidate values, and different combinations were sequentially tested using five-fold cross-validation. The optimal combination of hyperparameters was then selected based on the results of the cross-validation. For validation, we used the remaining 20% of site-level observations to evaluate the performance of our data-driven models in deriving GPP and ET. Thereafter, based on year-to-year dominant PFTs from MODIS for each grid cell, we extended data-driven models from site level to global scales using gridded meteorology from the MERRA-2 reanalysis and the three SIF products by applying the corresponding PFT-specific ML models at individual grids. Thus, we generated six GPP, ET, and WUE products (2 ML algorithms × 3 SIF datasets) at a spatial resolution of 0.05° and 8-day temporal resolution during 2001–2020 on the global scale.

2.5. Interpretation of Factor Contributions in Machine Learning

We used the Shapley Additive Explanation (SHAP) method to interpret the output of machine learning models. SHAP values, based on cooperative game theory, quantify the contribution of each input variable to the predicted value [59]. The basic principle of SHAP is to calculate the marginal contribution of a feature, i.e., the effect of a feature on the prediction results when it is added to a combination of existing features. This process requires considering all possible combinations of features. For model f and feature set S with feature i, the Shapley value is calculated as follows:

$${\varphi }_i={\sum }_{S\subseteq N\backslash \left\{i\right\}}{\displaystyle \frac{\left|S\right|!\left(\left|N\right|-\left|S\right|-1\right)!}{\left|N\right|!}}\left[f_{S\cup \left\{i\right\}}\left(x_{S\cup \left\{i\right\}}\right)-f_S\left(x_S\right)\right]$$

(4)

where N is the set of all features, |S| is the size of S, $S\subseteq N\backslash \left\{i\right\}$ represents all subsets that do not contain feature i, and x_S represents the values of the input features in the set S. When calculating the Shapley value ${\varphi }_i$ for each feature, all possible permutations of the features are considered with ${\displaystyle \frac{\left|S\right|!\left(\left|N\right|-\left|S\right|-1\right)!}{\left|N\right|!}}$. For each permutation, the difference in model output before and after the addition of the particular feature, $f_{S\cup \left\{i\right\}}\left(x_{S\cup \left\{i\right\}}\right)-f_S\left(x_S\right)$, is calculated. The marginal contributions of the particular feature across all permutations are averaged to obtain the Shapley value, representing the average contribution of the feature to the prediction results.

The SHAP model has been widely used to interpret the factor contributions of XGB and RF models [60,61,62,63,64]. Specifically, Zhou et al. [56] combined the SHAP method with XGB models to quantify the contributions of individual climatic factors and their interactions to GPP variability at different time scales and across various PFTs at the global FLUXNET sites. In this study, the mean absolute Shapley value of each feature was normalized to quantify the relative contributions of individual features to GPP and ET.

3. Results

3.1. Relationships Between SIF and Ecosystem Fluxes

We examined the relationships between 8-day SIF and GPP, as well as between SIF and ET, at global FLUXNET sites (Figure S2 and

Table 1. The relationships of GPP, ET, and SIF across PFTs at FLUXNET sites.

PFTs			CRO N = 17	DBF N = 23	EBF N = 13	ENF N = 49	GRA N = 37	MF N = 9	SH N = 16	SAV N = 9	WET N = 18	WSA N = 6
CSIF	GPP	r	0.62	0.84	0.59	0.81	0.81	0.91	0.81	0.68	0.75	0.82
		Slope1	19.12	24.94	15.30	23.74	22.91	21.58	20.25	16.56	17.95	26.84
		Slope2	18.65	24.22	22.99	27.11	22.84	21.55	20.90	20.30	18.74	25.43
	ET	r	0.58	0.75	0.71	0.69	0.71	0.75	0.70	0.59	0.51	0.69
		Slope1	0.21	0.26	0.34	0.25	0.27	0.24	0.34	0.29	0.25	0.51
		Slope2	0.29	0.29	0.38	0.35	0.34	0.27	0.46	0.44	0.36	0.58
GOSIF	GPP	r	0.61	0.77	0.21	0.69	0.78	0.90	0.61	0.69	0.71	0.77
		Slope1	20.78	24.49	5.38	22.19	26.65	23.21	18.38	20.59	18.34	26.96
		Slope2	22.51	27.56	25.75	30.48	28.15	25.93	23.45	24.39	21.00	27.19
	ET	r	0.59	0.74	0.41	0.61	0.71	0.77	0.52	0.65	0.50	0.67
		Slope1	0.24	0.28	0.19	0.25	0.33	0.27	0.30	0.38	0.27	0.52
		Slope2	0.35	0.34	0.44	0.39	0.43	0.32	0.51	0.54	0.40	0.63
RTSIF	GPP	r	0.61	0.77	0.19	0.68	0.77	0.90	0.61	0.71	0.70	0.60
		Slope1	10.95	13.79	3.25	13.00	14.93	13.97	11.34	13.15	10.57	13.68
		Slope2	11.53	15.03	16.93	17.67	15.09	15.08	13.28	14.90	11.57	15.89
	ET	r	0.56	0.73	0.40	0.60	0.67	0.75	0.51	0.64	0.49	0.48
		Slope1	0.12	0.15	0.12	0.14	0.18	0.16	0.18	0.24	0.15	0.25
		Slope2	0.18	0.18	0.29	0.23	0.23	0.19	0.30	0.33	0.22	0.37

Notes: N represents the number of sites; the slope1 (slope2) indicates the linear regressions coefficient of GPP and SIF or ET and SIF with (without) intercepts, respectively.

). All SIF products showed positive correlations with the observed GPP, with correlation coefficients (r) of 0.78, 0.7, and 0.69 (p < 0.05) for CSIF, GOSIF, and RTSIF, respectively. The slopes (g C m² day⁻¹/m W m⁻² nm⁻¹ sr⁻¹) were 21.68 (23.02), 21.82 (26.63), and 12.35 (14.86) with (without) intercepts for the three SIF products (Figure S2a–c), respectively. Among various PFTs, the SIF-GPP relationships showed positive correlations from 0.59 to 0.91 for CSIF, 0.21 to 0.9 for GOSIF, and 0.19 to 0.9 for RTSIF products (Table 1), with the highest correlations observed in mixed forest (MF) and the lowest in evergreen broadleaf forest (EBF). For forests, excluding EBF due to its weaker relevance [28], the SIF-GPP correlations were higher (0.73 to 0.90) compared to non-forest ecosystems (0.61 to 0.79) for all SIF products. Additionally, the SIF-GPP slopes were higher for forests, ranging from 19.59 to 21.07 g C m² day⁻¹ per unit SIF, indicating the stronger reabsorption of SIF by more complex canopy structures [30,65].

In comparison, the SIF-ET relationships showed lower correlations, with an r of 0.59–0.67 (p < 0.05) across the three SIF products (Figure S2d–f). The linear slopes of SIF-ET (mm hr⁻¹/m W m⁻² nm⁻¹ sr⁻¹) were 0.26, 0.27, and 0.15 for CSIF, GOSIF, and RTSIF, respectively. Among the various PFTs, positive SIF-ET correlations were found, ranging from 0.51 to 0.75 for CSIF, 0.41 to 0.77 for GOSIF, and 0.4 to 0.75 for RTSIF. The lowest correlations were found for EBF and wetland (WET) sites (Table 1). The linear slopes of SIF-ET ranged from 0.21 to 0.51 for CSIF, 0.19 to 0.52 for GOSIF, and 0.12–0.25 for RTSIF across various PFTs, with the highest slopes observed in woody savanna (WSA, Table 1). Forest ecosystems showed higher SIF-ET correlations (0.63–0.76) and slopes (0.21–0.23) compared to non-forest ecosystems (r: 0.5–0.7; slope: 0.19–0.43), though these differences were less pronounced than those observed for the SIF-GPP relationships.

3.2. Site-Level Validations of ML Models Across Various PFTs

For each PFT, we constructed six data-driven models using 80% of the total data samples for all FLUXNET sites with the same PFT classification (see Section 2). We validated the derived GPP and ET averaged for the six (2 ML × 3 SIF) models using the remaining 20% of observations (Figure 2 and Figure 3). The GPP derived by six data-driven models showed a Nash–Sutcliffe Efficiency (NSE) of 0.82 ± 0.11 and relative mean biases (RMB) of 0.34 ± 1.91% against observations average for various PFTs (Figure 2). The highest NSE value of 0.91 was found for mixed forest (MF), while the lowest value of 0.54 was found at cropland (CRO), potentially due to missing human disturbances in ML models such as irrigation and fertilizer management [66,67]. The largest bias of 4.55% was found for savanna (SAV), and the lowest bias was −0.09% at EBF sites. For ET, the data-driven models also exhibited a good performance, with an average NSE of 0.8 ± 0.07 and an RMB of −0.33 ± 1.68% across all PFTs (Figure 3). Among different PFTs, the highest NSE value of 0.87 was found for EBF, while the lowest NSE of 0.69 was found at GRA sites. The largest bias was 3.14% at SAV sites, while the smallest bias was 0.33% for EBF.

Data-driven models showed slightly varied performances in estimating GPP and ET across two algorithms and three SIF products (

Table 2. The independent validation of derived GPP and ET across PFTs.

PFTs				CRO N = 959	DBF N = 1214	EBF N = 622	ENF N = 2995	GRA N = 1795	MF N = 637	SH N = 701	SAV N = 427	WET N = 613	WSA N = 411
CSIF	XGB	GPP	NSE	0.52	0.90	0.86	0.71	0.86	0.91	0.92	0.81	0.81	0.85
			RMB	1.84%	0.05%	0.21%	1.64%	0.24%	0.09%	−0.08%	3.49%	−1.34%	−3.02%
		ET	NSE	0.75	0.86	0.86	0.80	0.81	0.83	0.72	0.69	0.86	0.88
			RMB	0.65%	−1.38%	0.25%	−0.71%	−1.49%	0.99%	0.12%	4.40%	−2.25%	−1.82%
	RF	GPP	NSE	0.57	0.90	0.85	0.80	0.87	0.91	0.91	0.80	0.81	0.87
			RMB	2.47%	0.01%	0.32%	1.00%	0.34%	0.34%	0.43%	5.06%	−1.67%	−2.23%
		ET	NSE	0.74	0.87	0.89	0.75	0.81	0.82	0.79	0.78	0.88	0.86
			RMB	1.17%	−2.20%	0.36%	−1.11%	−1.45%	1.13%	0.26%	1.67%	−2.73%	−2.71%
GSIF	XGB	GPP	NSE	0.51	0.90	0.86	0.70	0.85	0.91	0.90	0.82	0.80	0.85
			RMB	1.87%	−0.73%	0.07%	1.78%	−1.26%	−0.13%	1.24%	4.04%	−3.68%	−1.65%
		ET	NSE	0.74	0.85	0.86	0.75	0.81	0.85	0.66	0.69	0.85	0.87
			RMB	0.74%	−1.01%	0.43%	−0.56%	−1.25%	0.86%	0.79%	4.44%	−2.52%	−1.50%
	RF	GPP	NSE	0.55	0.90	0.86	0.80	0.87	0.92	0.89	0.78	0.81	0.87
			RMB	2.25%	−0.06%	0.02%	1.06%	0.10%	0.05%	0.87%	6.61%	−2.76%	−1.28%
		ET	NSE	0.74	0.85	0.88	0.75	0.81	0.82	0.78	0.77	0.88	0.85
			RMB	0.79%	−2.47%	0.20%	−1.02%	−1.46%	0.83%	0.47%	3.29%	−2.99%	−2.15%
RSIF	XGB	GPP	NSE	0.52	0.90	0.86	0.69	0.85	0.91	0.92	0.82	0.81	0.84
			RMB	1.54%	−0.36%	−0.34%	1.62%	−1.29%	−0.91%	0.29%	2.96%	−2.01%	−2.87%
		ET	NSE	0.74	0.85	0.86	0.76	0.1	0.84	0.69	0.70	0.86	0.87
			RMB	0.46%	−1.83%	0.39%	−0.58%	−0.94%	0.80%	−0.14%	3.05%	−2.8%	−1.52%
	RF	GPP	NSE	0.56	0.89	0.86	0.80	0.86	0.92	0.90	0.80	0.82	0.86
			RMB	1.94%	−0.19%	0.26%	1.50%	−0.29%	−0.16%	0.89%	5.14%	−2.58%	−2.24%
		ET	NSE	0.74	0.85	0.88	0.75	0.80	0.82	0.78	0.78	0.88	0.85
			RMB	0.82%	−2.58%	0.36%	−0.76%	−1.17%	0.89%	0.52%	1.98%	−2.91%	−2.93%

).

For GPP, the highest NSE value was 0.83 ± 0.1 for the CSIF-RF model, while the lowest NSE value was 0.81 ± 0.12 for the GOSIF-XGB model. The lowest RMB was −0.14 ± 1.69% for the RTSIF-XGB model, and the highest bias was 0.69 ± 2.35% for the GOSIF-RF model. Across various PFTs, all six data-driven models showed the lowest NSE values of 0.51–0.57 for CRO sites and 0.69–0.8 for ENF sites. In contrast, high NSE values were found for MF (0.91–0.92) and SH (0.89–0.92) sites, with low biases of −0.34%–0.32% for EBF and −0.91%–0.34% for MF. For ET, the highest NSE was 0.82 ± 0.05 for the CSIF-RF model, while the lowest NSE value was 0.73 ± 0.2 for the RTSIF-XGB model. The lowest RMB was 0.04 ± 1.82% for the GOSIF-XGB model, and the highest RMB was −0.58 ± 1.68% for the RTSIF-RF model.

Based on the SHAP values, we further explored the contributions of SIF and other factors to the variability of site-level GPP and ET using either XGB (Figure 4) or RF (Figure S3) models. For both algorithms, SIF showed the highest contributions to GPP variability, averaging 24%–49% across most PFTs excluding EBF (16%), where the correlations between SIF and GPP (or ET) were usually low (Table 1). However, Tair acted as the dominant contributor, with 30%–31% for DBF, ENF, and WET. Diffuse radiation and VPD contributed to GPP variability by 2–12% and 5–11% across various PFTs, respectively, while other factors had no significant impacts. For ET variability, SIF showed lower mean absolute SHAP values, up to 10–42%, compared to GPP variability (Figure 4 and Figure S3). Following SIF contributions, VPD and Tair influenced ET variability by 10–38% and 6–16%, respectively, among various PFTs, indicating the significant impacts of water and heat conditions on soil evaporation and plant transpiration. Other factors, such as diffuse radiation and CO₂, showed limited impacts on ET variability across the three SIF products and two algorithms. Overall, SIF showed tight relationships and dominant roles in GPP and ET variability at site-level scales. This indicates a high potential and advantage in deriving GPP and ET simultaneously by taking into account SIF and associated meteorology in the ML models.

3.3. Spatiotemporal Variations in ECSIF-Based GPP and ET

We further applied PFT-dependent models to construct the ECSIF_GPP and ECSIF_ET datasets on a global scale and compared the spatiotemporal variations in average GPP and ET from six ECSIF products with the ensemble of satellite-based products (Figure 5 and Figure 6). On the global scale, high GPP (>2000 gC m⁻² yr⁻¹) and ET (>1000 mm yr⁻¹) values were found over the tropical regions such as the Amazon, Central Africa, and Southeastern Asia (Figure 5a,d). In contrast, low GPP (<1000 gC m⁻² yr⁻¹) and ET (500 < mm yr⁻¹) values appeared at mid–high latitudes (e.g., boreal Asia), consistent with the observed spatial pattern (Figure 5b,e). Compared to the ensemble products, ECSIF_GPP showed a correlation of 0.93 and bias of 10.33% globally, with positive differences over the Amazon and Central Africa (Figure 5c). Similarly, ECSIF_ET exhibited a correlation of 0.94 and bias of 2.79% against observations (Figure 5f). The highest r for GPP (ET) was achieved up to 0.94 (0.92) with the CSIF-RF model (

Table 3. Summary of GPP and ET from the ECSIF datasets.

SIFs	Algorithms	Variables	Global	r	RMB	Trends
CSIF	XGB	GPP	124.8 ± 1.6 Pg C yr−1	0.92	8.02%	0.25 Pg C yr−2
		ET	582.4 ± 5.0 mm yr−1	0.88	15.88%	0.72 mm yr−2
	RF	GPP	128.1 ± 1.4 Pg C yr−1	0.94	11.13%	0.22 Pg C yr−2
		ET	460.4 ± 3.6 mm yr−1	0.95	−9.04%	0.52 mm yr−2
GSIF	XGB	GPP	126.1 ± 1.7 Pg C yr−1	0.89	9.20%	0.26 Pg C yr−2
		ET	583.8 ± 6.2 mm yr−1	0.91	14.99%	0.87 mm yr−2
	RF	GPP	130.8 ± 1.5 Pg C yr−1	0.94	12.48%	0.24 Pg C yr−2
		ET	468.2 ± 4.0 mm yr−1	0.95	−8.28%	0.58 mm yr−2
RSIF	XGB	GPP	125.3 ± 1.2 Pg C yr−1	0.89	8.66%	0.16 Pg C yr−2
		ET	574.5 ± 5.4 mm yr−1	0.89	12.06%	0.71 mm yr−2
	RF	GPP	130.1 ± 1.2 Pg C yr−1	0.93	12.49%	0.19 Pg C yr−2
		ET	463.4 ± 3.3 mm yr−1	0.94	−8.89%	0.45 mm yr−2

), and RF models showed better performance in both r and RMB than XGB-based models.

On the global scale, ECSIF products yielded a total GPP of 128 ± 2.3 Pg C yr⁻¹ (mean ± standard deviation) and area-weighed ET of 522 ± 58.2 mm yr⁻¹ during 2001–2020. These values are within the range of 106–145.8 Pg C yr⁻¹ for GPP and 417–560 mm yr⁻¹ for ET as estimated by different studies based on varied methods and data sources (

Table 4. Summary of global GPP and ET products.

Variables	Methods	Global	Temporal Cover	Spatial Resolution	Temporal Resolution	References
GPP and ET	BESS	122 ± 25 Pg C yr−1	2000–2015	1 km	8 day	Jiang and Ryu [16]
		501 mm yr−1
	BEPS	124 ± 4 Pg C yr−1	1982–2016	0.5° × 0.6°	hourly	He, et al. [68]
		485 ± 8 mm yr−1
	PML-V2	145.8 Pg C yr−1	2002–2017	500 m	8 day	Zhang et al. [5]
		560 mm yr−1
	X-BASE	124.7 ± 2.1 Pg C yr−1	2001–2020	0.05° × 0.05°	hourly	Nelson et al. [44]
		574 ± 7 mm yr−1
	SIF-ML	128 ± 2.3 Pg C yr−1	2001–2020	0.05° × 0.05°	8 day	This study
		522 ± 58.2 mm yr−1
GPP	EC-LUE	111 ± 21 Pg C yr−1	2000–2003	0.5° × 0.6°	8 day	Yuan, et al. [8]
	BEPS	132 ± 22 Pg C yr−1	2003	1° × 1°	hourly	Chen, et al. [69]
	LUE	122–130 Pg C yr−1	2000–2016	500 m	8 day	Zhang et al. [70]
	LUE	108–119 Pg C yr−1	2004–2012	1 km	8 day	Yu, et al. [71]
	ML&BESS	131–163 Pg C yr−1	2005–2015	0.5° × 0.5°	monthly	Badgley, et al. [72]
	GOSIF	136 ± 9 Pg C yr−1	2000–2023	0.05° × 0.05°	8 day	Li and Xiao [73]
	EC-LUE	106 ± 3 Pg C yr−1	1982–2017	0.05° × 0.05°	5, 6, 8 day	Zheng, et al. [74]
	ML	117 ± 1.5 Pg C yr−1	1999–2019	0.05° × 0.05°	monthly	Guo et al. [58]
ET	RS-PM	417 ± 38 mm yr−1	2000–2003	0.5° × 0.6°	8 day	Yuan, et al. [8]
	ML	500 ± 23 mm yr−1	1982–2008	0.5° × 0.5°	monthly	Jung, et al. [20]
	PT	522 mm yr−1	1980–2016	0.25° × 0.25°	daily	Miralles, et al. [75]
	WB	558–650 mm yr−1	1982–2009	0.5° × 0.5°	annual	Zeng, et al. [76]
	PML-V1	538 ± 57 mm yr−1	1998–2012	0.5° × 0.5°	monthly	Zhang et al. [77]

Notes: the original units of km³ were converted to mm year⁻¹ through dividing by the total land area excluding Antarctica and Greenland.

). Among the six sets of ECSIF products, the highest (lowest) GPP estimates were 130.8 (124.8) Pg C yr⁻¹ for the GOSIF-RF (CSIF-XGB) model (Table 3). The highest and lowest ET estimates were 582.4 mm yr⁻¹ for CSIF-XGB and 460.4 mm yr⁻¹ for CSIF-RF models, respectively. Regionally, the total GPP and area-weighed ET were 34.3 Pg C yr⁻¹ and 931 mm yr⁻¹ in South America, 34.4 Pg C yr⁻¹ and 548 mm yr⁻¹ in Asia, 25.7 Pg C yr⁻¹ and 745 mm yr⁻¹ in Africa, 17.4 Pg C yr⁻¹ and 492 mm yr⁻¹ in North America, 9.9 Pg C yr⁻¹ and 465 mm yr⁻¹ in Europe, and 6.1 Pg C yr⁻¹ and 481 mm yr⁻¹ in Oceania.

For trends, ECSIF_GPP showed annual trends of 0.22 ± 0.04 Pg C yr⁻², mainly resulting from positive trends over eastern China, India, and Europe (Figure 6a). This positive trend was similar to the 0.35 ± 0.11 Pg C yr⁻² observed in ensemble GPP products (Figure 6b), though the ECSIF_GPP slightly underestimated the positive GPP trends in boreal regions (Figure 6c). Meanwhile, ECSIF_ET exhibited a positive trend of 0.64 ± 0.14 mm yr⁻², with significant spatial heterogeneity, primarily driven by positive changes over eastern China and negative trends over Central Africa (Figure 6d). This trend was lower than that of 1.73 ± 0.64 mm yr⁻² in the ensemble ET products (Figure 6e). The largest differences were found over South America and Central Africa (Figure 6f), potentially due to limitations in training sites in the southern hemisphere. For different ML algorithms, ECSIF showed higher trends of 0.16–0.26 Pg C yr⁻² for GPP and 0.71–0.87 mm yr⁻² for ET with XGB, as compared to 0.19–0.24 Pg C yr⁻² for GPP and 0.45–0.58 mm yr⁻² for ET with the RF algorithms (Table 3).

3.4. Spatiotemporal Variations in Global WUE

We further explored spatiotemporal variations in WUE (=GPP/ET) for ECSIF products on the global scale (Figure 7). High WUE (>2 g C kg⁻¹ H₂O yr⁻¹) was found mainly in forest regions, including tropical rainforest (e.g., Amazon and Central Africa) and temperate forest (e.g., Europe and Eastern US). The low WUE appeared in the dry (e.g., Middle East, Western U.S., and Australia) and/or cold regions (e.g., boreal North America and Eurasia) (Figure 7a). This pattern of WUE was similar to satellite-based observations (Figure 7b), though the ECSIF product estimated lower WUE at the high latitudes (Figure 7c). On the global scale, ECSIF products yielded an area-weighted WUE of 1.8 ± 0.207 g C kg⁻¹ H₂O yr⁻¹ (mean ± standard deviation) during 2001–2020, close to 1.63 ± 0.118 g C kg⁻¹ H₂O yr⁻¹ from ensemble products. Meanwhile, ECSIF products showed positive trends of 0.0019 ± 0.00048 g C kg⁻¹ H₂O yr⁻² in WUE, with regional hotspots over Eastern China, India, Europe, Sahel, and northwestern US (Figure 7d), where cropland is usually located [78]. However, the ensemble satellite-based products revealed positive trends close to 0 (0± 0.00206 g C kg⁻¹ H₂O yr⁻²) for WUE (Figure 7e), which differed from the trends of 0.0008 g C kg⁻¹ H₂O yr⁻² in PML-V2, −0.0028 g C kg⁻¹ H₂O yr⁻² in MODIS, and 0.002 g C kg⁻¹ H₂O yr⁻² in X-BASE. The trends in WUE for ECSIF and X-BASE demonstrate a high degree of spatial coherence, particularly in the Northern Hemisphere, South America, and Australia (Figure S4).

Semi-mechanistic products showed larger differences globally, likely due to the varying models used in the two products. MODIS WUE displayed significant decreasing trends in Africa, Australia, and the Middle East, primarily because the positive trends in ET exceeded the positive trends in GPP in these regions. This imbalance resulted in a negative trend in global WUE (Figure S5a–c). In contrast, PML WUE showed a broader decline across the tropics, driven by higher ET trends relative to GPP at low latitudes, along with severe GPP declines in parts of Africa and South America. These declines were offset by significant increases in WUE in much of the Northern Hemisphere, including Europe, Central Asia, and parts of North America, ultimately resulting in a slight positive trend in global WUE for PML (Figure S5d–f). The most pronounced differences were observed in tropical regions, attributed to the strong positive ET trends derived from ensemble products (Figure 6e). Therefore, the accurate estimate of GPP and ET is of great importance to better understand ecosystem carbon and water fluxes and their coupling.

For different PFTs, forests exhibited a larger WUE (2.73 g C kg⁻¹ H₂O yr⁻¹) than non-forests (1.58 g C kg⁻¹ H₂O yr⁻¹). The largest WUE of 3.14 g C kg⁻¹ H₂O yr⁻¹ was observed in DBF, while the lowest was 1.27 g C kg⁻¹ H₂O yr⁻¹ in GRA. Croplands showed an intermediate WUE of 2.19 g C kg⁻¹ H₂O yr⁻¹. Similarly, the trend for forests (0.0018 g C kg⁻¹ H₂O yr⁻²) was higher than that for non-forests (0.0008 g C kg⁻¹ H₂O yr⁻²). Within forests, DNF, ENF, and MF all showed significant positive trends ranging from 0.0017 to 0.0048 g C kg⁻¹ H₂O yr⁻². In contrast, DBF and EBF showed insignificant trends of 0.0002 g C kg⁻¹ H₂O yr⁻² and −0.0004 g C kg⁻¹ H₂O yr⁻², respectively. For non-forests, CRO and GRA showed positive trends of 0.0042 g C kg⁻¹ H₂O yr⁻² and 0.0037 g C kg⁻¹ H₂O yr⁻², respectively, while the others showed no significant trends (Table S2).

4. Discussion

Our results indicate that site-level SIF showed a strong agreement with flux tower GPP at global sites (Figure S2a–c). However, the SIF-ET correlations were consistently lower than the SIF-GPP relationships (Figure S2d–f). This difference was likely attributed to the closer physiological link between SIF and GPP compared to SIF and ET at the canopy/leaf scale [79,80]. Theoretically, changes in SIF directly reflect plant photosynthesis due to the immediate energy flux emitted during the light reaction [81,82]. Thus, increases (or decreases) in SIF correspond to positive (or negative) changes in plant photosynthesis at the canopy scale [83]. Linked by the electron transport rate (J), the SIF-GPP relationship can be broken into two components: SIF–J and J–GPP. The SIF-J relationship, governed by energy allocation in the light reaction, can be modeled using the Mechanistic Light Response (MLR) model [84]. Meanwhile, the J-GPP relationship is linked to carbon assimilation during the Calvin cycle and aligns with the Eco-Evolutionary Optimality (EEO) theory [85], which states that stomata maximize carbon gain while minimizing water loss [86]. This highlights SIF’s key role in providing physiological insights for estimating both GPP and ET. The SHAP analysis further revealed that SIF accounted for 24–49% of GPP variability across all PFTs (excluding EBF), a significantly higher contribution than any climatic factors (Figure 4 and Figure S3). The high contribution of SIF is likely because it is the only vegetation state variable included in the predictor set, capturing all effects associated with canopy structure and APARchl variability. In contrast, SIF had a weaker influence on terrestrial ET, primarily through its role in linking photosynthesis and stomatal transpiration [39]. Increases in SIF enhance transpirations via stomatal opening, which is driven by photosynthesis [35,87,88]. However, soil and canopy evaporation, which are substantial components of ET, are more strongly controlled by environmental factors like temperature and VPD rather than photosynthesis [89]. For example, Lu et al. [35] showed that correlations between ET and SIF could be influenced by VPD, air temperature, and radiation at the site level. Consistent with this, our SHAP results indicate that temperature, rather than SIF, was the dominant driver of ET variability across most PFTs (Figure 4 and Figure S3). These findings highlight the importance of incorporating both SIF and meteorological factors when estimating ET.

The derived ECSIF datasets presented spatial patterns and temporal trends in global GPP, ET, and WUE consistent with observations (Figure 5, Figure 6, Figure 7, Figures S4, S6 and S7). Previous studies reported global GPP estimates of 106–145.8 Pg C yr⁻¹ and ET estimates of 417–574 mm yr⁻¹ across various periods and resolutions. These values align with the 124.8–130.8 Pg C yr⁻¹ for GPP and 460.4–583.8 mm yr⁻¹ for ET obtained from six ECSIF datasets (Table 4). Meanwhile, ECSIF products showed 1.55–2.03 g C kg⁻¹ H₂O yr⁻¹ of global WUE during 2001–2020, falling within the range of 1.53–1.96 g C kg⁻¹ H₂O yr⁻¹ from previous estimates [22,44,90,91]. In terms of temporal trends, ECSIF products showed significant trends of 0.16–0.26 Pg C yr⁻² for global GPP during 2001–2020. These values are close to the 0.2–0.46 Pg C yr⁻² reported Zhang et al. [5], Running et al. [42], Guo et al. [58], and [44] for the similar periods. However, global trends in ET showed large differences among various studies [92]. ECSIF products showed positive trends of 0.45–0.87 mm yr⁻² of global ET during 2001–2020, which were similar to 1.09 mm yr⁻² caused by Nelson et al. [44] for the period 2001–2020. These positive values were much lower than the value of 2 mm yr⁻² from PML-V2 datasets for similar periods [5]. Additionally, Jiang and Ryu [16] found negative global ET trends up to −0.65 mm yr⁻² during 2001–2011, while other studies indicated no significant trends [76,93]. For WUE, ECSIF products showed positive trends of 0.0001–0.0024 g C kg⁻¹ H₂O yr⁻² during 2001–2020, similar to 0.0025 g C kg⁻¹ H₂O yr⁻² and 0.002 g C kg⁻¹ H₂O yr⁻² as estimated by Xue et al. [90] and Nelson, et al. [44]. Furthermore, previous studies revealed substantial increases in WUE from regional [94,95] and global [96] scales, confirming the reasonableness of our derived positive trends of WUE. Meanwhile, we also compared the performances of GPP and ET among ECSIF and FLUXCOM X-BASE with global FLUXNET sites (Figure S6), and we found ECSIF showed NSE (bias) by 0.51 (−16.12%) and 0.62 (3.2%) with site-level GPP and ET, close to FLUXCOM X-BASE products. Overall, ECSIF products showed good performance in capturing the magnitudes and trends in GPP, ET, and WUE compared to ground and satellite-based observations and other studies. Considering uncertainties in various ECSIF products, we suggested that the average of six ECSIF products be used as benchmark GPP and ET with lowest uncertainties.

Previous studies have found that rising atmospheric CO₂ concentrations could enhance plant photosynthesis, driving the trend in global GPP upward through the CO₂ fertilization effect [97,98,99]. The coupling effects of leaf stomata and plant transpiration further amplify this process, as increased greening leads to higher global ET due to the expanded leaf area [92,100]. However, increases in plant photosynthesis typically outpace those in ET [101], resulting in an overall increase in WUE attributed to the CO₂ fertilization effect. Meanwhile, elevated CO₂ concentrations can partially inhibit leaf stomata, further enhancing WUE [95]. Variations in GPP were often influenced by surface radiation and temperature, particularly at the mid–high latitudes [98], while ET variations were mainly controlled by water exchange processes, such as precipitation [92,102]. Our results show that WUE trends were positively correlated with GPP enhancements driven by rising temperatures and negatively correlated with ET responses to precipitation variations (Figure 7d and Figure S8c,d). Similar findings have been reported by Xue et al. [90]. As a result, changes in WUE were determined by the competitions between regional GPP and ET, with GPP increases dominating WUE enhancements in most areas.

Nevertheless, there are some limitations and uncertainties in this study. First, the selection of SIF datasets may influence GPP and ET estimates of ECSIF products. For example, GOSIF-based datasets showed higher GPP and ET magnitudes and trends than CSIF- and RTSIF-based datasets (Table 3), due to differences among various SIF products. In addition, the time spans of SIF datasets limited the temporal extensions of ECSIF products, such as CSIF (2001–2020) and RTSIF (2001–2020). Based on that, we plan to update the annual ECSIF_GPP and ECSIF_ET datasets in the future, considering the continuously updated GOSIF dataset. Second, we selected only XGB and RF algorithms as the data-driven models to derive ECSIF products. Our analyses showed that the RF model caused larger biases than the XGB algorithm (Table 3). Other machine (or deep) learning algorithms may potentially perform better than the current ECSIF datasets. Third, the time spans and spatial unevenness of global FLUXNET sites may influence the accuracy of the upscaled GPP and ET estimates [21], as well as the effectiveness of ECSIF datasets. For example, the data used for training in this paper are only available up to 2014, and over 86% of the sites analyzed in this study were located in the Northern Hemisphere, with 64% in America and Europe. Therefore, more evenly distributed flux tower sites are needed to improve the accuracy of ML models in future studies. Fourth, uncertainties in input datasets (e.g., site-level GPP, CO₂ products, and land use data) may introduce biases into the derived GPP and ET estimates from the XGB and RF models. We examined the effects of these datasets on global-scale GPP and ET (Table S3–S5). Using site-level GPP data derived from daytime and nighttime methods, we found similar magnitudes and trends for global GPP, ET and WUE as the original estimates (Table S3). In addition, we used heterogeneous CO₂ datasets [103] to estimate global GPP, ET, and WUE, yielding values of 129 Pg C yr⁻¹, 471 mm yr⁻¹, and 1.95 g C kg⁻¹ H₂O yr⁻¹, respectively, which are close to our estimates using unified CO₂ datasets (Table S4). Finally, we assessed the impact of land use datasets on derived GPP, ET, and WUE, finding small biases of −1.28%, −1.61%, and −0.3%, respectively, compared to our estimates derived from dominant land cover datasets (Table S5).

Despite these limitations, our derived ECSIF products showed several advantages compared to previous GPP or/and ET products and provide a reliable tool to explore WUE: (1) Comprehensive considerations of plant physiology and environmental constraints. The ECSIF products fully accounted for the effects of plant physiology (e.g., SIF) and environmental changes (e.g., diffuse radiation), which were usually ignored in previous studies especially for ET products [104,105]. Zhou et al. [56] found that diffuse radiation dominated 30% of site-level GPP variability at the sub-daily scales. Wang et al. [106] also found that diffuse radiation significantly contributed to ET variability due to changes in transpiration. Our results confirmed that diffuse radiation can contribute significantly, both up to 8%to GPP and ET variability (Figure 4 and Figure S3), comparable to the contributions of VPD (8% and 10%). (2) Simultaneous derivation of GPP and ET. Previous terrestrial products typically included only GPP or ET [5], with very few studies capable of deriving both GPP and ET on a global scale. In this study, ECSIF products included six combined datasets of global GPP and ET using three SIF datasets and two ML algorithms to minimize the uncertainties from single methods or datasets. These datasets are expected to be useful for exploring water and carbon fluxes and their interactions such as WUE at regional and global scales in the future. (3) High spatiotemporal resolution. ECSIF estimated global GPP, ET, and WUE at a spatial resolution of ~5km and an 8-day temporal resolution. As a result, ECSIF products could serve as benchmarks for model validations and data explorations at a very fine resolution.

5. Conclusions

In this study, we developed the ECSIF-based product with six datasets of GPP, ET, and WUE at 0.05° spatial resolution and 8-day temporal resolution during 2001–2020. These datasets were based on three SIF products and two machine learning algorithms. First, we examined the relationships of SIF and GPP, as well as SIF and ET, at global FLUXNET sites. In general, positive correlations were found between SIF and these flux variables across most PFTs. We then created six ML models for each of ten PFTs, based on the cross-combination of two ML algorithms and three SIF datasets. Validations of these models showed that the derived GPP and ET values had high correlations and low biases when compared to site-level observations. Using SHAP methods, we confirmed that SIF played a dominant role in predicting site-level GPP and ET. Finally, we generated global ECSIF_GPP and ECSIF_ET datasets by applying the ML models to gridded measurements and found significant increases in global GPP, ET, and WUE during 2001–2020. The causes and consequences of these profound changes in global carbon and water fluxes deserve further investigations in future studies. Moreover, we consider using more spatially distributed site datasets with longer temporal coverage, SIF data observed from satellite directly, and different machine learning algorithms such as deep learning to derive more accurate, further spatiotemporally spanning, and higher-resolution global GPP and ET.