The Role of Solar-Induced Chlorophyll Fluorescence (SIF) in the Mechanistic Simulation of Eco-Hydrological Processes

Highlights

What are the main findings?

• Meteorology-driven and remote sensing-driven mechanistic models show markedly different performance in simulating GPP and LE.
• Incorporating SIF substantially improves mechanistic model accuracy by reducing bias associated with uncertainties in parameters and forcing data.

What are the implications of the main findings?

• SIF provides critical physiological constraints for mechanistic eco-hydrological modeling, supporting more accurate assessment and management of terrestrial carbon-water processes.
• Quantifying the contributions of key drivers (e.g., SIF) enhances the interpretability of mechanistic eco-hydrological models and facilitates understanding of model mechanisms, supporting improved model development and application.

Abstract

Accurate quantification of ecohydrological processes is essential for effective water and carbon management in terrestrial ecosystems. Traditional simulations mainly rely on mechanistic models, yet their accuracy is often limited by inconsistencies in representing physical processes and uncertainties in parameterization. Integrating remote sensing signals offers a promising way to reduce these uncertainties and enhance model applicability. In this study, in-situ observations from a wheat cropland in the Guanzhong Plain were used to simulate gross primary productivity (GPP) and latent heat flux (LE) by comparing a forward model (STEMMUS-SCOPE) with a remote sensing-driven inverse model (STEMMUS-MLR). We further examined the role of solar-induced chlorophyll fluorescence (SIF), an emerging proxy for photosynthesis, as an input to improve mechanistic modeling of GPP and LE. Results show that STEMMUS-MLR outperformed STEMMUS-SCOPE in estimating water and carbon fluxes, demonstrating that incorporating SIF effectively reduces bias associated with uncertainties in parameters and forcing data. The contribution of SIF was quantified using Random Forest regression and Shapley additive explanations (SHAP), revealing that SIF markedly reduced the dependence of GPP and LE simulations on shortwave radiation (SW), air temperature (Ta), and leaf area index (LAI). These findings highlight the critical role of SIF in ecohydrological modeling of semi-arid cropland ecosystems and provide a scientific basis for advancing process understanding and improving the precision management of water and carbon budgets in terrestrial ecosystems.

1. Introduction

Strongly coupled water, carbon, and energy fluxes are fundamental components of the eco-hydrological cycle and the mass and energy exchange of ecosystems [1,2,3]. Among surface water fluxes, evapotranspiration (ET), comprising plant transpiration (T) and soil evaporation (Es), requires substantial energy absorption and influences the energy partitioning of land surface ecosystems [4,5,6,7,8]. Carbon fluxes mainly include gross primary productivity (GPP) and ecosystem respiration, which determine whether an ecosystem acts as a carbon sink or source [9,10,11,12,13]. Importantly, stomatal conductance simultaneously regulates water vapor and CO₂ exchange between plants and the atmosphere [14,15,16]. Energy fluxes, including net radiation (Rn), sensible heat flux (H), latent heat flux (LE), and soil heat flux (G), are the primary driving force of the water–carbon cycle [14,17,18,19]. These coupled exchanges of water, carbon, and energy, including ET and GPP, constitute core eco-hydrological processes [20]. Within the soil–plant–atmosphere continuum (SPAC) framework, these processes describe the transport and partitioning of mass and energy between soil and canopy, playing a critical role in regulating ecosystem functioning, resilience, and productivity [21,22].

Accurate quantification of these fluxes is critical for understanding ecohydrological processes in agroecosystems [23,24,25]. Current approaches mainly include in-situ observations and model simulations [26,27,28]. The eddy covariance (EC) technique provides direct and accurate measurements of water, carbon, and energy fluxes, enabling insights into ecosystem functioning [29,30,31]. However, its application is limited by sparse spatial coverage and high installation and maintenance costs, making it impractical for large-scale or long-term analyses [32,33,34,35,36]. Consequently, for regional to global applications, model simulations are indispensable [37,38]. Early models often relied on empirical or semi-empirical formulations of individual fluxes. Over time, mechanistic eco-hydrological models have evolved based on the SPAC framework, explicitly coupling water, carbon, and energy processes [21,22,39]. These mechanistic models are generally categorized as meteorology-driven or remote sensing-driven [40].

Solar-induced chlorophyll fluorescence (SIF) has recently emerged as a powerful remote sensing indicator for monitoring ecosystem functioning [41,42]. SIF is the faint red to far-red light (approximately 640–850 nm) emitted by chlorophyll molecules during photosynthesis [43]. It represents a small fraction (~0.5–2%) of absorbed photosynthetically active radiation (APAR) that is neither used for carbon fixation nor dissipated via non-photochemical quenching (NPQ) [44,45,46]. Unlike traditional reflectance-based vegetation indices (e.g., normalized difference vegetation index (NDVI) and enhanced vegetation index (EVI)), which primarily reflect canopy “greenness” or potential photosynthetic capacity, SIF is a direct byproduct of photochemical processes [47,48,49]. It directly captures actual photosynthetic activity and integrates real-time plant physiological responses to environmental conditions [42,50]. Numerous studies have demonstrated strong correlations between SIF and GPP in ecohydrological studies, highlighting its potential as a direct proxy for photosynthesis [50,51,52,53]. Moreover, because transpiration is closely coupled with carbon assimilation through stomatal conductance, SIF provides indirect information on canopy-level water fluxes [54,55]. By capturing rapid physiological responses to environmental stresses before structural changes occur, SIF provides a mechanistic pathway to reduce uncertainties in parameterizing complex SPAC models and to improve estimates of GPP and LE [20].

The STEMMUS-SCOPE model represents a meteorology-based mechanistic framework that integrates the canopy radiative transfer model (SCOPE) with the one-dimensional soil water and heat transport model (STEMMUS), incorporating root water uptake and growth processes. By coupling soil moisture (SM) dynamics with plant physiological responses, STEMMUS-SCOPE substantially improves simulation accuracy and performs robustly under different water stress conditions, including drought [56]. The model, initially validated in maize farmland, grassland, and shrubland ecosystems [57], has since been reliably confirmed across 170 global sites covering diverse climatic regimes and vegetation types, highlighting its ability to capture complex SPAC processes [58,59].

In contrast, the STEMMUS-MLR model is a remote sensing-driven mechanistic model that integrates STEMMUS with a mechanistic light response (MLR) framework. Unlike meteorology-driven models, STEMMUS-MLR uses SIF as its primary driver, complemented by meteorological variables [60]. By explicitly accounting for SM profiles and root distribution, STEMMUS-MLR represents ecohydrological processes well by leveraging the physiological information embedded in the SIF signal [20].

Despite these advancements, a key knowledge gap remains, and the relative performance of meteorology-driven and remote sensing-driven mechanistic models under identical conditions has not systematically been assessed yet. Addressing this gap is critical for clarifying the added value of SIF and improving the mechanistic quantification of ecohydrological processes. Therefore, this study uses in-situ observations from agroecosystems in the Guanzhong Plain to (1) compare the performance of the climate-driven STEMMUS-SCOPE model and the SIF-driven STEMMUS-MLR model in simulating GPP and LE and (2) reveal the role of SIF in ecohydrological simulations and quantify the contributions of different factors to GPP and LE simulations.

2. Materials and Methods

2.1. Site Description

The Yangling Site (34°17′ N, 108°04′ E; 521 m a.s.l.) is located in the Guanzhong Plain of Northwest China, within the Institute of Water Saving Agriculture in Arid Areas (IWSA) at Northwest Agriculture and Forestry University (NWAFU), Shanxi Province (Figure 1) [61]. The study area experiences an average of 2000 sunshine hours and an annual mean temperature of 12.9 °C, with a frost-free period of approximately 210 days, while climatic drought events frequently occur [62]. According to the United States Department of Agriculture Soil Classification System, the soil in the study area is classified as silty clay loam [63,64], with a bulk density of 1.35 g cm⁻³ and a field capacity of 0.235 m³ m⁻³ [65,66,67].

2.2. In-Situ Data Collection and Processing

This study used an automated six-channel, tower-based SIF observation system installed approximately 2 m above the vegetation canopy to measure canopy radiance around the flux tower. The system was deployed on six tripods surrounding the flux tower to reduce potential footprint mismatch between SIF and flux observations. The instrument consisted of a cosine corrector for downwelling irradiance measurements, a spectroradiometer, and an optical multiplexer that sequentially switched between downwelling irradiance and multi-channel upwelling radiance measurements. Radiometric and spectral calibrations were performed prior to deployment to ensure measurement stability and accuracy. During data processing, dark current correction and integration time optimization were applied [68,69]. Subsequently, the “sandwich” sampling strategy (downwelling–upwelling–downwelling) was applied to further reduce the influence of rapid atmospheric variability on SIF retrieval [70]. SIF was retrieved from continuous spectral measurements using the singular value decomposition (SVD) method based on the O₂-A absorption band [71,72]. The SVD method assumes that SIF-free spectra can be represented as linear combinations of singular vectors derived from a SIF-free training dataset. In this study, the training dataset consisted of continuous downwelling solar spectral measurements [73,74,75]. The prescribed fluorescence spectral shape used for retrieval was derived from SCOPE simulations [69]. Top-of-canopy (TOC) SIF in the near-infrared band was retrieved using a spectral fitting window of 740–780 nm, and the resulting values were aggregated into 30-min averages [71]. Low-quality retrievals under low illumination (photosynthetically active radiation (PAR) < 30 μmol m⁻² s⁻¹) or high solar zenith angles (>65°) were excluded [69,76].

An EC system, consisting of a three-dimensional ultrasonic anemometer, an open-circuit infrared gas analyzer, and a data logger, was mounted on an adjustable-height tripod near the spectrometry system to measure CO₂ fluxes during the wheat growing season in this region. The sensor was positioned 2 m above the ground and oriented toward the prevailing wind direction. Wind speed (u) and direction, along with other boundary layer meteorological measurements, were recorded using a propeller anemometer. Air temperature (Ta) and relative humidity (RH) were recorded at 1.5 m above the soil surface using a combined temperature-humidity sensor, while Rn was measured with a four-component radiation sensor at 2 m above the soil surface. Rainfall was measured using a tipping bucket rain gauge, and G was measured with a self-regulating heat flux plate at 8 cm below the soil surface. All meteorological variables were recorded as 30-min averages using a data logger.

The 10 Hz raw data were converted into half-hour net ecosystem exchange (NEE) using EddyPro 7 software. Flux data affected by sensor malfunctions, low wind conditions (friction velocities below 0.1 m s⁻¹), or precipitation (Pre) were excluded [65]. Missing data compensation and decomposition of CO₂ into total ecosystem respiration (TER) and GPP were performed using the REddyProc online tool created by the Max Planck Institute for Biogeochemistry, with input data, including NEE, RH, Ta, soil temperature (Ts), and solar radiation (SR). The gap-filling algorithm in REddyProc efficiently fills in gaps for meteorological variables (Ta and vapor pressure deficit (VPD)) and flux data [61,77]. Thirty-minute GPP was calculated as the NEE–TER difference [56]. Based on 30-min Ta and RH data, VPD can be obtained following Campbell and Norman’s method [78].

2.3. STEMMUS-SCOPE Model

The STEMMUS-SCOPE model is a coupled model based on the soil–vegetation–atmosphere continuum. It incorporates radiative transfer, photochemical processes, and energy balance, simulating the dynamic growth of roots, SM, Ts, energy, water and carbon fluxes, and fluorescence (Figure 2) [56,79,80]. The model is primarily used to investigate vegetation responses under water-stress conditions [56,57]. Based on the initial SM profile simulated by STEMMUS, the STEMMUS-SCOPE model first calculates the water stress factor (WSF) as the limiting factor for the maximum carboxylation rate (V_cmax). Net photosynthesis (An), T, energy fluxes, GPP, and SIF were then calculated using SCOPEv1.73. The above-calculated T determines plant root water uptake (RWU) in STEMMUS. Root growth is then calculated from An. Based on RWU and dynamic root growth, STEMMUS calculates changes in Ts and SM [67,81,82]. Soil surface moisture is used to calculate the soil surface resistance, which is then used to compute Es. Overall, SM is subsequently used to calculate the WSF for the next time step. Meanwhile, SCOPE calculates the soil surface temperature using the energy balance, using it as the top boundary condition for STEMMUS. It also calculates the leaf water potential (LWP) through iteration [56]. The key functions used in STEMMUS-SCOPE are detailed in Appendix A.

2.4. STEMMUS-MLR Model

Similar to STEMMUS-SCOPE, the STEMMUS-MLR model is an SIF-driven model which couples the MLR [41] with STEMMUS (Figure 2) [60,67,79,80,81,82]. It utilizes SIF and meteorological observations to simulate the terrestrial eco-hydrological processes. First, the MLR model accurately calculates GPP using SIF through the MLR framework. Then, canopy conductivity (Gc) can be calculated based on Ball–Woodrow–Berry (BWB) canopy conductance model, using GPP and WSF, which is calculated by the initial SM from the STEMMUS simulation. Then, T can be calculated using Gc and the dual-source Penman-Monteith (PM) equations. T is passed to STEMMUS as RWU. Meanwhile, STEMMUS calculates soil surface resistance based on surface SM, which is then used to calculate Es via the PM equation. The STEMMUS model simulates the updated SM for each layer using real-time updated RWU and Es, which then serves as the starting condition for the next time step [20]. The key functions used in STEMMUS-MLR can be found in Appendix A.

2.5. Random Forest Model and SHAP Method

In this study, a Random Forest (RF) model combined with the Shapley additive explanations (SHAP) method was used to quantitatively assess the importance of SIF in simulating eco-hydrological processes. Previous studies have demonstrated that the RF model outperforms other machine learning algorithms in simulation tasks [83,84,85]. Therefore, the RF model was employed to learn and approximate the outputs of the mechanistic model, constructing a data-driven surrogate model [86,87]. Based on this, SHAP was applied to evaluate the contribution of each input variable within the surrogate model, revealing the internal response mechanisms and key driving factors of the mechanistic model [88,89]. This can enhance model interpretability and transparency and further clarify the crucial role of SIF in eco-hydrological simulations. Specifically, input variables for SHAP analysis were selected based on the forcing variables and key drivers used in the STEMMUS-SCOPE and STEMMUS-MLR frameworks. Accordingly, the explanatory features included meteorological forcing variables (e.g., Ta and u), vegetation-related variables, and SIF observations. This design ensures that the SHAP interpretation focuses on quantifying the sensitivities and relative contributions of major drivers controlling mechanistic model outputs. For surrogate modeling, RF regression was implemented in Python 3.13 using scikit-learn, with mean squared error (MSE) as the splitting criterion. Because RF includes built-in mechanisms to reduce overfitting (e.g., bootstrap aggregation and random feature selection), only the random seed (random_state = 42) was specified, and all other hyperparameters were set to the default values in scikit-learn. Subsequently, SHAP values were computed using the TreeSHAP algorithm based on the trained RF surrogate model to quantify the contribution of each driver to the model outputs and improve interpretability.

2.6. Experiment Design

Figure 2 illustrates the workflow of this study, which consists of main components: datasets, simulations, verification, and interpretation. Datasets include meteorological observations and remote sensing data, and are used as model inputs after pre-processing, including missing value removal and variable selection. Simulations are conducted using the STEMMUS-SCOPE and STEMMUS-MLR models. The STEMMUS-SCOPE model uses meteorological data, soil parameters, and initial conditions as inputs. In contrast, the STEMMUS-MLR model simulates the ecohydrological processes by integrating remote sensing SIF with meteorological inputs. Both models use STEMMUS as the soil module, which incorporates root growth and mass balance equations. However, the two models differ in their canopy modules. The STEMMUS-SCOPE model integrates the SCOPE model, which simulates radiative transfer and energy balance processes. The STEMMUS-MLR model uses a canopy module based on the MLR framework combined with the PM equation. The simulated GPP and LE are then evaluated to systematically assess the applicability and performance of each model. Finally, the importance of SIF and meteorological inputs in eco-hydrological simulations is analyzed using the RF model combined with the SHAP analysis.

2.7. Statistical Analysis Methods

The accuracy of mechanistic models (STEMMUS-SCOPE and STEMMUS-MLR) is evaluated using the Coefficient of Determination (R²), Root Mean Square Error (RMSE) and Kling–Gupta Efficiency (KGE) [3,90,91]. The formulas for the metric are as follows:

$${\mathrm{R}}^2={\left({\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}\left({\mathrm{P}}_{\mathrm{i}}-\overline{\mathrm{P}}\right)\left({\mathrm{O}}_{\mathrm{i}}-\overline{\mathrm{O}}\right)}{\sqrt{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{P}}_{\mathrm{i}}-\overline{\mathrm{P}}\right)}^2{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{O}}_{\mathrm{i}}-\overline{\mathrm{O}}\right)}^2}}}\right)}^2$$

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{1}{\mathrm{n}}}\sum _{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{P}}_{\mathrm{i}}-{\mathrm{O}}_{\mathrm{i}}\right)}^2}$$

$$\mathrm{K}\mathrm{G}\mathrm{E}=1-\sqrt{{\left(\mathrm{R}-1\right)}^2+{\left(\mathsf{\alpha }-1\right)}^2+{\left(\mathsf{\beta }-1\right)}^2}$$

where ${\mathrm{P}}_{\mathrm{i}}$ represents the i-th predicted value, $\overline{\mathrm{P}}$ is the mean of the predicted values, ${\mathrm{O}}_{\mathrm{i}}$ is the i-th observed value, $\overline{\mathrm{O}}$ is the mean of the observed values, n is the sample size, R is the Pearson correlation coefficient, α is the ratio of the standard deviations of the simulated and observed values, and β is the ratio of the mean simulated to observed values.

3. Results

3.1. In-Situ Meteorology, Fluxes, and SIF Observations

Data were collected from January to June 2021, corresponding to the winter wheat growing season. Figure 3 presents observed data aggregated at hourly and daily scales. The dataset consists of three groups: (1) meteorological variables as inputs for mechanistic models (Ta, u, Pre, shortwave radiation (SW), longwave radiation (LW), VPD, air pressure (P), leaf area index (LAI), and CO₂ concentration) and (2) observed fluxes as target outputs (Rn, LE, H, GPP); and (3) remotely sensed SIF.

During the observation period, Ta ranged from −13.3 to 47.7 °C, with a mean of 11.4 °C, while P remained stable (938.4–984.6 hPa). Daily average wind was 1.09 m s⁻¹, fluctuating between 0 and 4.66 m s⁻¹. SW and LW showed overall increasing trends, with hourly peaks of 1068.1 W m⁻² and 446.3 W m⁻², respectively. Precipitation occurred mainly in May–June (totaling 149.3 mm), increasing air humidity and reducing evaporative demand. VPD decreased during Precipitation events but exhibited an overall increasing trend, with higher values in May–June than in January–April. LAI rose sharply in early February during regreening, peaked at 5.88–6.23 m² m⁻² in late April–early May, and then declined until harvest. Daily GPP peaked in April–May during vigorous wheat growth, while GPP was lower in other months due to reduced LAI and radiation. Daily mean LE peaked at ~400–450 W m⁻² in early May with peak LAI.

3.2. STEMMUS-SCOPE and STEMMUS-MLR Model Simulations

Figure 4 illustrates the simulation results of the STEMMUS-SCOPE model for both GPP and LE at hourly and daily scales. On the daily scale, the simulated trends generally match the observations, although significant deviations appear in January and June. The model performs similarly well for GPP (KGE = 0.67, slope = 0.62) and LE (KGE = 0.88, slope = 0.85), with slightly higher accuracy for LE. At the hourly scale, performance declines for GPP (KGE = 0.72, slope = 0.64) but remains robust for LE (KGE = 0.89, slope = 0.85).

As shown in Figure 5, the STEMMUS-MLR model achieves consistently high accuracy for both fluxes. For LE, KGE reaches 0.89 at the hourly scale and 0.91 at the daily scale, with a slope of 0.97 and 1.00; while for GPP, the corresponding values are 0.92 and 0.90, with a slope of 0.96 and 1.04. Minor discrepancies occur toward the end of the study period, but simulations generally align well with observations during the main growing season, demonstrating robust overall performance.

Overall, by requiring fewer parameters, the STEMMUS-MLR model achieves improved simulations of both GPP and LE across temporal scales compared with STEMMUS-SCOPE. These findings suggest that although STEMMUS-SCOPE effectively represents eco-hydrological processes, incorporating SIF into mechanistic models can further improve simulation accuracy, both for GPP and LE.

3.3. SHAP-Based Feature Importance and Contributions of SIF and Other Factors

To quantify the contributions of input drivers to GPP and LE, we applied the RF model and the SHAP method for feature importance analysis. As shown in Figure 6, we compared feature importance with and without SIF to reveal the information redundancy between SIF and environmental factors. Without SIF, SW, LAI, Ta, and VPD emerged as the dominant features for GPP. However, after introducing SIF, the contributions of SW, LAI, and Ta decreased markedly, with SIF becoming the most influential predictor. This reflects the strong physiological link between SIF and GPP, as simulations using SIF alone already achieved high accuracy (KGE = 0.91, Figure 7).

For LE, SW, LAI, and Ta were the primary predictors in the absence of SIF. After incorporating SIF, their importance declined, while SIF ranked as the second most important variable. Simulations using SW and SIF yielded high accuracy (KGE = 0.88, Figure 7), whereas SIF alone resulted in substantially lower accuracy (KGE = 0.77, Figure A1). This demonstrates that SIF alone is insufficient for LE, where radiation-driven Es remains critical. Collectively, these results indicate redundancy between SIF and variables such as SW and LAI, and suggest that SIF plays a central role in GPP modeling, while accurate LE simulations require the joint influence of radiation and soil water availability.

Beyond relative importance, SHAP dependence plots further revealed nonlinear relationships between drivers and fluxes (Figure 8 and Figure 9). For GPP, SIF exhibited the strongest positive association, forming a tightly clustered distribution, thereby reflecting its ability to capture dynamic changes in photosynthesis (Figure 8, right panel a). LAI showed a dual effect: negative at low values (LAI < ~4 m² m⁻²) due to limited canopy light interception, but positive at higher values as canopy efficiency improved (Figure 8, right panel b). SW and LW both had generally positive contributions, highlighting the pivotal role of radiation in photosynthesis (Figure 8, right panel b).

In contrast, LE simulations were primarily controlled by SW (Figure 9). Although the SHAP value of SW decreased slightly after SIF was introduced, it remained the dominant driver, confirming the importance of radiation as the main energy source for ET. Meanwhile, SIF ranked second in importance, reinforcing its complementary role but also demonstrating redundancy with other radiation- and vegetation-related variables.

Together, these results emphasize that SIF could improve the attribution and simulation of GPP and LE. Its strong explanatory power highlights its critical role in GPP estimation, while also revealing the continued importance of Es and root-soil water interactions in LE simulation.

3.4. The “Equivalent Modeling” Approach for Interpreting Mechanistic Models

To investigate how input variables shape the outputs of mechanistic models (STEMMUS-SCOPE and STEMMUS-MLR), we employed the concept of an “equivalent model”. Specifically, an RF model was trained on the simulations and forcing of the mechanistic models to approximate their behavior, and SHAP was then applied to evaluate feature importance. This approach reveals the internal response mechanisms and dominant drivers of the mechanistic models, while maintaining their physical consistency and improving interpretability.

For STEMMUS-SCOPE, the SHAP analysis highlighted SW and LAI as major drivers of GPP (Figure 10 and Figure A2), with influence patterns closely matching those from the RF model trained on observational data. This consistency indicates that STEMMUS-SCOPE captures the fundamental processes of radiation input, vegetation structure, and air humidity. However, the importance of Ta was reduced, suggesting that the model does not fully capture the regulatory effects of temperature on photosynthesis and instead overemphasizes the impact of VPD. Refining temperature-related parameterization or introducing physiologically based response functions could improve its accuracy under varying thermal conditions. For LE, STEMMUS-SCOPE overemphasized the role of VPD while underestimating the contribution of LAI. This arises from the fact that, within STEMMUS-SCOPE, VPD directly enters the LE calculation and governs stomatal conductance, thereby indirectly affecting GPP estimation (Figure 10 and Figure A3).

For STEMMUS-MLR, the contribution of SIF to GPP was generally consistent with that of the observationally trained RF model (Figure 10 and Figure A4). However, the importance of Ta increased markedly, as temperature is required for calculating the photosystem II operating efficiency (q_L) within the model. In contrast, the importance of SW declined substantially. This may be attributed to the dominant role of SIF as the primary input of STEMMUS-MLR, whose feature importance far exceeded that of SW (Figure 10), thereby diminishing the latter’s influence in the model structure. These findings reveal the limitations of linear models in handling strong multicollinearity and nonlinear ecological responses. For the simulation of LE with STEMMUS-MLR, the model consistently identified SW as the dominant controlling factor (Figure 10 and Figure A5). However, the role of VPD was again overestimated, likely because the PM formulation depends strongly on VPD. At the same time, the contribution of SIF was underestimated, suggesting that the model’s capacity to fully utilize SIF data requires further improvement (Figure 10 and Figure A5). The overestimation of VPD may also stem from the fact that both the BWB canopy conductance model and the PM formulation explicitly require VPD as an input. In contrast, STEMMUS-MLR showed a weaker influence of VPD on GPP, mainly because GPP is derived directly from SIF and then used to calculate canopy conductance and T, without VPD being involved in the computation of GPP itself.

4. Discussion

4.1. SIF Better Represents GPP than LE

SIF has been widely recognized as a reliable remote sensing proxy for photosynthetic activity [44,50,92]. Physiologically, SIF is emitted as a by-product of photosynthesis, representing a small fraction of APAR. It is closely linked to photosystem II activity, reflecting light-use efficiency and electron transport rates in leaves [44,50,93]. Because of this mechanistic link, numerous studies have consistently demonstrated a strong correlation between SIF and GPP across ecosystems and climate regimes [50,52,53]. Our findings further confirm that SIF plays a more dominant role in simulating GPP compared to traditional meteorological variables, providing a direct, observation-based constraint on photosynthetic dynamics that climate-driven mechanistic models often struggle to reproduce.

In contrast, the role of SIF in representing LE or ET is far less robust. Unlike GPP, which is directly tied to photochemistry, LE is influenced by multiple physiological and hydrological processes, including stomatal conductance, VPD, SM availability, and aerodynamic conditions [94,95,96,97]. SIF primarily reflects the physiological status of leaves and their photosynthetic activity but does not account for non-photosynthetic components of water fluxes, such as Es and canopy interception [51,98]. This limitation is particularly evident in water-limited or semi-arid ecosystems, where Es can contribute substantially to total LE [99,100].

Moreover, under drought stress, stomatal closure can strongly reduce transpiration without necessarily inhibiting the photochemical activity that generates SIF, leading to a decoupling of the SIF–LE relationship [101,102]. These physiological mechanisms explain why empirical or machine learning models relying solely on SIF often produce biased ET estimates, particularly under water stress or heterogeneous soil conditions. This highlights the importance of coupling SIF observations with process-based models to better represent both photosynthetic and non-photosynthetic components of water fluxes, rather than relying on SIF alone for ET or LE estimation.

4.2. The Role of SIF in Mechanistic Models

Mechanistic models are extensively employed to simulate eco-hydrological processes based on fundamental principles, such as mass and energy conservation [103,104]. These models rely on detailed physical parameters [105,106]. Their explicit parameterization of complex processes and adherence to energy and mass conservation generally yield more reliable simulations when driven by regional datasets [107,108]. Mechanistically integrating SIF with environmental variables, such as SM, VPD, and canopy conductance, has been shown to improve ET estimation [97,109]. In this study, the climate-driven mechanistic model (STEMMUS-SCOPE) exhibited lower accuracy compared with the SIF-driven model (STEMMUS-MLR). This finding further proves that relying solely on meteorological variables and LAI cannot fully represent the underlying physiological processes of vegetation. In contrast, STEMMUS-MLR, which integrates SIF as the primary input, outperformed STEMMUS-SCOPE in simulating both GPP and LE.

Structurally, STEMMUS-SCOPE and STEMMUS-MLR share the same STEMMUS soil module but differ mainly in their canopy and photosynthesis representations. Specifically, STEMMUS-SCOPE uses a multilayer radiative transfer and energy balance scheme to simulate canopy radiation transfer, vertical profiles of temperature and humidity, and layer-specific photosynthesis and transpiration [56]. In contrast, STEMMUS-MLR, a big-leaf model, uses a simplified MLR framework and dual-source PM model. It lacks the complex multilayer canopy structure and requires far fewer physiological parameters [20,60]. Consequently, STEMMUS-MLR significantly reduces computational cost compared with STEMMUS-SCOPE, with a runtime approximately one-tenth as long [20]. Although structurally simpler, the photosynthesis formulation of STEMMUS-MLR remains consistent with the Farquhar–von Caemmerer–Berry (FvCB) biochemical model implemented in SCOPE. The key difference is in how the electron transport rate is estimated. In STEMMUS-SCOPE, within the FvCB model, the maximum electron transport rate is mainly driven by PAR via radiative transfer simulations. In STEMMUS-MLR, the actual electron transport rate is constrained by SIF observations. Although a relatively simple structure in STEMMUS-MLR, this study highlights that SIF effectively represents photosynthetic processes, compensating for the limitations of climate-driven models that rely on precise parameterization of photosynthetic capacity (e.g., V_cmax) and stomatal regulation, thereby improving overall model performance.

From a modeling perspective, two key implications can be drawn: First, integrating SIF into ecohydrological models provides a powerful constraint on photosynthesis-related processes, significantly improving GPP simulations. Second, accurate representation of LE requires explicit consideration of additional hydrometeorological variables (e.g., SM, VPD, and aerodynamic resistance) and process-based representations of Es. Therefore, while SIF-driven models hold promise for advancing carbon cycle modeling, their application to water fluxes must be complemented by other data sources to avoid systematic biases. In summary, this study reinforces the role of SIF as a robust indicator of photosynthetic carbon uptake while cautioning against its direct application to LE or ET modeling.

4.3. Limitations and Future Perspectives

This study also has several limitations: First, the SHAP analysis was performed using an RF surrogate model rather than being applied directly to the mechanistic models. Although the surrogate model accurately reproduces the outputs of STEMMUS-MLR (e.g., GPP: KGE = 0.98; LE: KGE = 0.98) and STEMMUS-SCOPE (e.g., GPP: KGE = 0.96; LE: KGE = 0.99; Figure A6), SHAP explanations primarily reflect statistical relationships learned by the RF model rather than strictly mechanistic processes. However, because the surrogate model inputs were selected from the same forcing variables used in the mechanistic models, the high agreement between surrogate predictions and mechanistic outputs suggests that the machine learning model can reliably capture the underlying relationships embedded within the mechanistic frameworks. Second, the study focuses solely on the Yangling Site and C3 cropland, which limits the generalizability of the results across different agroecosystems, climatic conditions, or crop types. Third, although the study area is prone to climatic droughts, the dataset only covers hourly observations from January to June 2021. The relatively short time span may not allow for a fully comprehensive assessment of drought impacts on ecosystem functioning, which could influence the robustness of the results. Future research should expand to multiple sites and climatic zones to improve representativeness and to better evaluate the influence of extreme events, such as droughts, on model performance.

In the future, integrating machine learning, mechanistic models, and remote sensing signals offers promising opportunities. While machine learning approaches provide high computational efficiency, particularly at regional scales, mechanistic models offer greater robustness in evaluating ecosystem responses to extreme climate events and could serve as valuable training data for machine learning models. In addition, SIF provides direct physiological information of vegetation and mitigates the influence of parameter uncertainties on model simulations. Future work will utilize multi-year, multi-site datasets, including TROPOMI SIF satellite products and FLUXNET long-term flux observations, to evaluate ecosystem functioning across broader spatial and temporal scales. In addition, C3 and C4 vegetation types in specific vegetation types (e.g., cropland and grassland) will be identified and classified to support more robust and accurate cross-ecosystem and cross-climate assessments. By leveraging the complementary strengths of these approaches, future studies can achieve more accurate simulations of ecosystem functioning across broader spatial and temporal scales using SIF, thereby advancing predictions of ecosystem responses to climatic stressors. Moreover, these enhanced simulations can inform practical applications, including optimizing irrigation schedules, guiding water resource management, and supporting carbon accounting at farm or regional scales.

5. Conclusions

This study evaluated the performance of two eco-hydrological models in simulating LE and GPP under different levels of data availability. Results show that incorporating SIF substantially improved the accuracy of both hourly and daily simulations in STEMMUS-MLR (GPP: KGE = 0.90; LE: KGE = 0.91), outperforming STEMMUS-SCOPE (GPP: KGE = 0.67; LE: KGE = 0.88), which relied solely on climatic inputs. This indicates that SIF serves as a key driver that enhances the predictive capability of process-based eco-hydrological models. In STEMMUS-MLR, SIF directly improves GPP estimation and, consequently, strengthens the simulation of T and SM dynamics, leading to a more realistic representation of eco-hydrological processes. The SHAP analysis further reveals that the improved accuracy stems from the integration of physiological information carried by SIF—capturing vegetation responses to radiation, Ta, and structural information from LAI. However, the representation of certain driving processes in the model still has room for improvement. Overall, these findings highlight the critical role of SIF in advancing eco-hydrological modeling of semi-arid cropland ecosystems and its potential application in Earth System Models. Practically, SIF-based models can support water management decisions, guide irrigation planning, and improve carbon accounting, thereby bridging the gap between mechanistic modeling and real-world ecosystem management. Building on this, future research should explore the development of digital twin systems by incorporating advanced remote sensing signals into mechanistic models.