The FLUXNET2015 Tier 1 dataset (https://fluxnet.org/, last access: 14 July 2025), spanning approximately 1990–2014, integrates long-term EC flux observations from multiple regional networks worldwide and provides harmonized tower data after standardized flux partitioning, quality control, and gap-filling (Pastorello et al., 2020). To further expand the training samples for modelling, we additionally incorporated data from AmeriFlux (https://ameriflux.lbl.gov/, last access: 21 April 2025), covering ∼1990 to the present (data up to 2022 were used), that follow the same processing pipelines with FLUXNET (Chu et al., 2023). Given that the footprint of EC observations differs from the 0.05° grid of the upscaled product, we applied additional filtering to reduce scale mismatch (Chu et al., 2021). We first used the MODIS land cover product (MCD12Q1 Version 6.1 with 500 m resolution, covering 2001–2022) and reclassified its vegetation types following the scheme in Table S1 in the Supplement to ensure consistency with the flux-site classification. For each flux site, we identified the corresponding 0.05° grid cell containing the tower location and calculated the fractional coverage of each reclassified vegetation type from all 500 m MODIS pixels within that grid cell for each year. The vegetation type with the largest fractional coverage was defined as the dominant class of that 0.05° grid cell. We kept only those flux sites whose vegetation type consistently matched the dominant class of their corresponding 0.05° grid cell throughout all available years, resulting in a total of 190 representative sites selected from 316 available sites across diverse vegetation types and Whittaker biomes (Fig. 1, Table S1).

To ensure data quality for model training and evaluation, we applied the following procedures to the tower-based GPP (GPP_NT_VUT): (1) removed negative values that are physiologically unrealistic; (2) retained only records with quality-control flags equal to 0 or 1; and (3) converted all timestamps to local solar time (LST) to remove phase shifts related to solar position and ensure physical consistency of diurnal variations (see Sect. S1 for details).

To comprehensively evaluate the performance of our hourly GPP product, we compared it with two recently developed hourly GPP products that are also generated via machine-learning-based upscaling strategies. The first is FLUXCOM (Bodesheim et al., 2018), a pioneering upscaling product that provides global GPP at 0.5° resolution for 2001–2014; The second is X-BASE, a global 0.25° GPP product developed under the recent FLUXCOM-X framework covering 2001–2021 (Nelson et al., 2024). To our knowledge, these two products were upscaled from in-situ eddy covariance measurements and represent the efforts in this field. All products were evaluated against site-level GPP, focusing on overall accuracy, representation of diurnal cycles, and spatiotemporal patterns.

Prior to modeling, we performed feature selection to reduce the influence of redundant predictors (Fig. 2). Owing to its robust screening strategy, Boruta is particularly suitable for high-dimensional and multi-source modelling tasks (Kursa and Rudnicki, 2010). The algorithm operates by creating shadow features to establish a baseline. Through iterative testing, predictors with lower importance than their corresponding shadow features are removed. We implemented this process using the BorutaPy package in Scikit-learn with 100 iterations to ensure result stability. As shown in Fig. S2, the clumping index and LAI were discarded, indicating that their information was likely redundantly captured by EVI. The remaining predictors were then evaluated for multicollinearity using the variance inflation factor (VIF). The results showed that all retained predictors satisfied the collinearity criterion (VIF < 10) and were therefore used in subsequent modelling (Table S3). The final set of global gridded predictors used for modelling and upscaling is listed in Table S2.

To quantify the diurnal dynamics of vegetation photosynthesis, we used a set of metrics that describe the GPP diurnal curve from three complementary perspectives: absolute magnitude, diurnal asymmetry, and midday depression.

First, the absolute magnitude was defined as the mean GPP during the period from 06:00 to 18:00 LST, reflecting the average intensity of daytime photosynthetic activity.

Second, diurnal asymmetry describes the difference in photosynthetic activity between morning and afternoon (Liu et al., 2024). This behavior is commonly quantified using metrics such as ΔGPP, diurnal centroid, and peak time. In this study, ΔGPP was used as the primary asymmetry metric and was calculated as: 1ΔGPP=GPPAM-GPPPMGPPAM×100% where GPPAM and GPPPM denote the mean GPP during the morning (06:00–12:00 LST) and afternoon (12:00–18:00 LST), respectively. The diurnal centroid and peak time were used as supplementary indicators (Li et al., 2023; Wilson et al., 2003), a centroid or peak time occurring before noon indicates higher GPP in the morning than in the afternoon.

Finally, we introduced the relative midday depression (RMD) to quantify the reduction in photosynthesis around midday (10:00–14:00 LST), which reflects the physiological response of vegetation to stress. Based on the approach developed by (Zhu and Zhu, 2025), RMD quantified the normalized difference between the observed GPP curve and a hypothetical ideal GPP curve. Specifically, this ideal curve was constructed by fitting a parabolic curve to non-midday hour GPP values, thereby representing the expected diurnal trajectory in the absence of midday depression. The deviation between the observed GPP curve and this ideal curve was then used as a proxy for midday photosynthetic depression. RMD is calculated as: 2RMD=∑t=10t=14GPPit-GPPat∑t=10t=14GPPat×100% where GPPi(t) and GPPa(t) are the actual and ideal GPP values at time t, respectively. A larger RMD indicates a greater deviation of actual GPP from the ideal curve and thus a stronger midday depression. To validate the sufficiency of these three metrics in capturing diurnal pattern, we conducted additional experiments, the details of which are presented in Sect. S6.

Because these diurnal metrics are standardized, they are directly comparable across ecosystems and provide a consistent basis for global-scale analysis. Accordingly, we produced global maps of these diurnal indicators to visualize their spatial patterns. To assess how well different hourly GPP products capture diurnal dynamics, we constructed a three-dimensional metric space with mean GPP, ΔGPP, and RMD as the three axes. The distance between each product's data points and the site-based reference points in this metric space was used to quantify its ability to reproduce the observed diurnal variation patterns.

To assess how well EGO GPP captures diurnal vegetation photosynthesis under extreme climate conditions across ecosystems, we analyzed two well-documented extremes: the severe drought that occurred in the central North America in June 2021 (Jiang et al., 2025) and the European heatwave in August 2003 (Garcia-Leon et al., 2021). For each case, we used flux tower observations with valid records for the corresponding months, covering the major ecosystem types in the central North America and Europe, respectively.

The evaluation consisted of two parts. To ensure a rigorous and independent assessment, we utilized the aggregated estimated GPP derived from the 10-fold cross-validation (i.e., the compilation of the 10 test sets), rather than the final upscaled product generated from the full dataset. Firstly, we quantified the accuracy of the estimated GPP during the extreme event. The estimated hourly GPP were paired with corresponding tower-based GPP, and GPP anomalies during the event periods were calculated for both datasets. Specifically, the anomaly was calculated as (GPPextreme- GPP_reference) / GPP_reference, where GPP_extreme represents the value during the specific event month, and GPP_reference denotes the multi-year average for the same month excluding the event year. We then compared the R2, RMSE, and bias to evaluate the consistency between EGO GPP and the observation. Second, we examined the ability of EGO GPP to reproduce diurnal photosynthetic dynamics under extreme events. For each period, we extracted diurnal curves from both the flux tower data and EGO GPP at each site. Using the metrics defined in Sect. 3.4, we derived diurnal metrics for every site and grouped them into dryland and non-dryland classes. The capacity of EGO GPP to represent stress-induced changes was evaluated by calculating the absolute difference in metrics between the event and reference periods (e.g., RMD_difference = RMDextreme- RMD_reference). These differences were then compared with tower-based results to verify whether the EGO accurately captured the shifts in diurnal patterns.

Comparative experiments among the three baseline models (Fig. S3) showed that XGBoost achieved the best performance (test set: R2 = 0.71, RMSE = 4.48 µmol CO₂ m⁻² s⁻¹), followed by Random Forest, whereas LSTM performed relatively poor. The proposed CKML-GPP model further improved prediction accuracy over XGBoost, increasing R2 by 0.03 for the training set and by 0.05 for the testing set. These improvements indicate that the PCMCI-guided model not only enhances predictive skill to capture complex driving relationships, but also improved its generalization to unseen test samples. We further added a sensitivity experiment in which CKML-GPP was applied to daytime-partitioned GPP (GPP_DT) and compared with GPP_NT (Fig. S4). The results show that the model performance based on GPP_DT was generally consistent with that based on GPP_NT, suggesting the robustness of CKML-GPP across different partitioning methods. Based on the established CKML-GPP model, we generated the global hourly upscaling product, EGO GPP, and then compared with FLUXCOM and X-BASE against EC flux GPP (Fig. 3). To ensure a fair comparison, all products were compared under identical conditions using only the test set samples for EGO (Sect. 3.2.2). Under this setup, EGO achieved an R2 of 0.76 and an RMSE of 4.17 µmol CO₂ m⁻² s⁻¹, outperforming the other products which show R2 values around 0.60 and RMSE value around 5.5 µmol CO₂ m⁻² s⁻¹. Regarding bias, both FLUXCOM and X-BASE exhibited systematically underestimation of GPP, with substantial negative biases of -2.35 and -2.33 µmol CO₂ m⁻² s⁻¹, respectively. In contrast, EGO effectively mitigated this underestimation, reducing the bias to merely -0.36 µmol CO₂ m⁻² s⁻¹. This improvement was also reflected in the regression slope (Fig. 3a–c), where FLUXCOM and X-BASE remained around 0.60, whereas EGO reached 0.78. As shown in Fig. 3d, EGO GPP aligned well with tower GPP across all quantiles, suggested excellent performance for both low and high GPP ranges. Moreover, we conducted a sensitivity experiment by resampling EGO to 0.25° and 0.5°, matching the spatial resolutions of X-BASE and FLUXCOM, respectively (Fig. S6). Although its performance declined slightly after resampling (R2 = 0.73 at 0.25° and R2 = 0.71 at 0.5°), EGO still outperformed both products, indicating robustness across spatial scales.

We then applied the SHAP method to interpret feature contributions in CKML-GPP. The results (Fig. 4) showed that SW was the dominant contribution (34.04 %), consistent with its role as the primary energy source for vegetation photosynthesis, EVI (19.54 %), Ta (10.03 %), LSWI (9.32 %), VPD (5.64 %), and Aridity (5.33 %) were identified as the next most influential drivers, together explaining over 80 % of the model's output variance. Among these, SW, EVI, Ta, LSWI and Aridity exerted positive effects on GPP, whereas VPD showed a strong inhibitory effect. Regions dominated by forests and croplands tended to exhibit higher GPP, while areas with larger proportions of grasslands or shrublands showed lower GPP. SHAP interaction analysis (Fig. S7) revealed pronounced synergistic effects among SW, EVI, and LSWI, at a given level of EVI or LSWI, GPP increased substantially with increasing SW, indicating a synergistic coupling among energy supply, canopy structure, and surface moisture in regulating diurnal carbon uptake. In contrast, a strong negative interaction was observed between VPD and Aridity: the suppressive effect of VPD on GPP was more pronounced in drier environments (lower aridity index), implying that ecosystems in drylands are more vulnerable to atmospheric drought.

We further compared product accuracy across vegetation types and along the aridity gradient. As shown in Fig. 5, EGO GPP consistently outperformed the other products for nearly all vegetation types. For EBF, ENF, and GRA, the R2 values of EGO GPP all exceeded 0.75. From Fig. 5b, the accuracy of FLUXCOM and X-BASE generally declined markedly with decreasing aridity index. In extremely arid zones (aridity index < 0.2), their R2 values dropped sharply, indicating that these products struggle to represent photosynthesis in very dry environments. In contrast, EGO GPP maintained relatively high accuracy in these zones (R2> 0.6), and its performance showed weaker sensitivity to the aridity gradient. To explicitly assess the role of aridity index, we trained additional models with and without aridity index as a predictor (Fig. S8). Incorporating aridity index improved model performance for both the training and test sets, with particularly pronounced gains in drylands. In extremely arid zones, including aridity index increased R2 by 0.13 for the training set and by 0.08 for the test set (Fig. S9). The largest improvements were found in GRA, SAV, and SHR biomes, which are predominantly distributed in arid and semi-arid regions. These results indicate that aridity index provides complementary information on water availability that is not fully captured by other meteorological or vegetation variables, especially in water-limited ecosystems.

To evaluate the capability of EGO in capturing diurnal photosynthetic variation, we aggregated tower GPP by local solar time within each vegetation type and extracted the corresponding GPP values from each product for comparison. As shown in Fig. 6, EGO showed a clear advantage in reproducing diurnal GPP curve across all vegetation types, whereas FLUXCOM and X-BASE generally underestimated GPP throughout the day. EGO GPP (red line) closely followed the diurnal patterns of tower-based GPP (black line), particularly for NF, CRO, and BF. This advantage extended to water-limited biomes like SAV and SHR (Fig. 6e and f). In these ecosystems, GPP is not only lower on average but also often peaks before 12:00 LST due to environmental stress. EGO was the product that mostly accurately captured this earlier peak, aligning closest with the reference curves, while the other products tended to produce lower value and inconsistent diurnal shapes that failed to reflect these patterns. As expected, comparisons at our six selected representative sites across biomes (Fig. S10) further showed that EGO most closely matched FLUXNET observations and effectively captured how diurnal GPP cycles varied across seasons.

We further selected six sites with continuous 10-year records for each vegetation type and visualized their diurnal GPP dynamics using fingerprint plots (Fig. 7). These plots simultaneously display hourly variations of tower GPP, EGO GPP, and their differences for each month across multiple years, clearly illustrating the seasonal evolution and the daily progression of photosynthesis from dawn to dusk. For instance, a crop site (CH-Oe2, Fig. 7a) and a forest site (US-Ha1, Fig. 7b) exhibited persistently high photosynthetic activity throughout the daytime, maintaining substantial GPP values nearly year-round, whereas grass site such as US-SRG (Fig. 7d) and shrubland site such as US-Jo1 (Fig. 7e) showed lower GPP levels and pronounced seasonality. Across all sites, EGO GPP demonstrated high consistency with observations, with differences typically below 1 µmol CO₂ m⁻² s⁻¹ at a given time. Importantly, no systematic overestimation or underestimation was apparent in either morning or afternoon periods, indicating its reliability in capturing the overall shape of diurnal GPP curves without introducing temporal bias.

To systematically quantify diurnal GPP patterns, we conducted experiments reconstructing GPP curves by controlling different metric combinations (Sect. S6). We selected a representative diurnal curve from an arid site (US-Wkg, August 2013) as an example (pink curve in Fig. 8a), which exhibited clear midday depression and pronounced asymmetry (ΔGPP > 0), with both centroid and peak time occurring before noon. When only the absolute magnitude was constrained (Case 1, blue curve), the reconstructed curve deviated substantially from the reference, with both peak and centroid times shifting toward the afternoon. Constraining both magnitude and asymmetry (Case 2, green curve), produced a more realistic curve overall but still failed to reproduce the midday depression. When asymmetry and midday depression were jointly constrained (Case 3, purple curve), the curve shape closely matched the reference, although GPP magnitude was systematically overestimated. In contrast, when all three metrics were jointly constrained (Case 4, yellow curve), the simulated curve nearly overlapped with the reference. These experiments demonstrate that the combined use of mean GPP, ΔGPP and RMD can comprehensively characterize diurnal photosynthetic dynamic.

Under this framework, we constructed a three-dimensional metric space to evaluate how well different products captured diurnal GPP dynamics. In this space (Fig. 8b), a smaller distance between a product and the observed point indicated a more accurate representation of photosynthesis diurnal pattern. Across nearly all vegetation types, EGO GPP was closest to observations. Among the other products, X-BASE performed relatively well in NF and CRO, while FLUXCOM showed better performance in BF and GRA. Overall, EGO not only reproduces the absolute magnitude of GPP but also more accurately captures the diurnal shape of photosynthesis dynamics.

We first produced global maps to examine the spatial patterns of EGO GPP at different local solar time (Figs. 9 and S11). Photosynthesis commenced after sunrise when solar radiation became available. GPP increased throughout the morning (Fig. 9a–c) as light availability improved under favorable temperature and moisture conditions, peaked around noon (Fig. 9d–f), and declined in the afternoon as photosynthetic activity weakened. After sunset, GPP approached zero in the absence of sunlight (Fig. 9g–i). Spatially, regions such as the Amazon, central Africa, and Southeast Asia exhibited consistently high GPP, with midday values exceeding 25 µmol CO₂ m⁻² s⁻¹. In contrast, high-latitude regions, the western United States, and Australia showed lower photosynthetic activity throughout the day, rarely surpassing 10 µmol CO₂ m⁻² s⁻¹. Beyond these general patterns, Fig. S12 reveals substantial seasonal variations in diurnal photosynthesis. During June–August, photosynthetic activity intensified across the Northern Hemisphere, with midday GPP exceeding 25 µmol CO₂ m⁻² s⁻¹ in boreal and temperate forests. In December–February, the epicenter of global photosynthetic activity shifted southward, with enhanced photosynthesis in the Southern Hemisphere, particularly in the Amazon and southern Africa. The seasons transition reflects the gradual latitudinal migration of global photosynthesis belt driven by changing light availability. In terms of uncertainty (Fig. S13), the standard deviation of EGO GPP remains below 1 µmol CO₂ m⁻² s⁻¹ across most regions and times, even in high-GPP areas.

Within the selected ROIs, the Amazon represents a core region of strong photosynthetic activity, where EGO and the other products all captured persistently high GPP values (∼15 µmol CO₂ m⁻² s⁻¹) even in the late afternoon (Fig. 10e). Moreover, EGO delineates finer spatial variations within forest types and transition zones, whereas the other products appeared overly smoothed (Fig. 10c, h, and m). This enhanced spatial realism is attributed to the use of sub-grid fractional PFT cover instead of discrete coarse vegetation class labels, which better represents gradual transitions in land cover. Similar advantages in spatial detail are evident in the central North America (Fig. S14) and Europe (Fig. S15).

We then presented the multiyear growing-season averages of four diurnal metrics derived from EGO GPP including RMD, ΔGPP, centroid time, and peak time (Fig. 11). Together, these metrics reveal distinct regional characteristics of diurnal photosynthetic dynamics. High RMD values occur mainly in tropical and arid regions, particularly in tropical drylands such as the Somali Peninsula in East Africa, where RMD exceeded 30 %. This indicates a pronounced midday depression of photosynthesis, strong physiological downregulation under intense radiation and severe water and heat stress. In contrast, high-latitude vegetation exhibited much weaker midday depression (RMD < 10 %) due to lower radiation and more favorable hydrothermal conditions. The spatial patterns of ΔGPP, centroid time, and peak time are strongly consistent. Globally, ΔGPP values are mostly negative, indicating that photosynthetic activity is generally stronger in the afternoon than in the morning. Positive ΔGPP values are concentrated in typical drylands such as the western United States, southern Africa, and Australia, where afternoon photosynthesis was strongly suppressed by water and heat stress. These regions also showed earlier centroid and peak times (before 12:00 LST), associated with rapid morning increases and earlier declines in photosynthetic activity. Conversely, in high-latitude and tropical rainforest regions, ΔGPP is small or negative, and both centroid and peak time occur near or after noon, indicating that photosynthesis remains relatively active in the afternoon.

To further assess the performance of EGO GPP under extreme events, we examined two representative cases in the central North America and Europe and analyzed their impacts on diurnal vegetation photosynthesis. As shown in Fig. 12, vegetation in the northwestern U.S. experienced severe environmental stress in June 2021 relative to the multi-year average: soil moisture decreased by about 40 %, while TA and VPD increased by approximately 30 % and 50 %, respectively. Against this backdrop, GPP across most of the U.S. declined markedly, particularly in the western and north-central U.S., where vegetation suffered the strongest stress and GPP losses, approached 50 %. In both reference and extreme months, GPP showed a general pattern of lower values in the western U.S. and higher in the east (Fig. S16). We then evaluated EGO's performance using observations from 39 flux towers with valid records during extreme and normal June (Fig. S17a–c). EGO maintained robust performance during the extreme month (R2= 0.69), comparable to its accuracy in normal Junes (R2= 0.80). Importantly, EGO effectively captured the GPP anomalies induced by the drought, consistent with ground-based observations (R2= 0.51).

To further examine diurnal responses, we aggregated observed and EGO GPP across dryland and non-dryland sites in the U.S. In drylands (Fig. 13a), intense water stress in June 2021 (black curve) led to a pronounced shift of the GPP peak, advancing by approximately 2 h compared with reference years (gray curve). Consistently, RMD increased from 6.81 % to 13.14 %, and diurnal asymmetry strengthened (ΔGPP from -0.03 to 0.12). Non-dryland sites (Fig. 13b) showed more moderate changes: overall GPP decreased modestly, while RMD increased from 4.91 % to 5.52 %, and ΔGPP shifted from -0.04 to 0.02. Across both drylands and non-drylands, EGO GPP closely matched flux-tower observations. More importantly, EGO closely matched flux-tower observations, and successfully reproduced both the direction and magnitude of changes in key diurnal metrics under drought stress.

Spatial analyses (Fig. 14) further revealed systematic differences between dryland and non-dryland ecosystems across the U.S. The western U.S., characterized by extensive drylands, exhibited higher RMD and ΔGPP values alongside earlier centroid and peak times (before 12:00 LST) compared to eastern U.S. Under the extreme drought event, these patterns intensified, particularly in the severely affected northwestern U.S., where RMD increased by about 15 %, ΔGPP rose by roughly 0.1, the diurnal centroid advanced by nearly 0.2 h, and peak time shifted earlier by 1 to 2 h. Strengthening of both midday depression and diurnal asymmetry was more pronounced in drylands than in non-drylands. Diurnal metrics derived from EGO and tower-based GPP (Fig. S18), showed strong agreement for both drylands and non-drylands.

A similar analysis for the August 2003 European heatwave yielded consistent conclusions. The heatwave caused temperature increases of up to 40 % and strongly reduced GPP across all hours, with anomalies reaching about 45 % (Figs. S20 and S21). The heatwave advanced GPP diurnal peaks by about 1 h earlier across Europe, with more pronounced changes in drylands (Fig. S21): RMD increased from 4.51 % to 9.67 %, and ΔGPP shifted from -0.06 to 0.02. Non-drylands showed gentler responses, with limited changes in morning GPP changes but notable midday depression (RMD up to 8.33 %). Although ΔGPP increased across most regions, diurnal asymmetry remained weak outside Mediterranean drylands (Fig. S22), indicating comparatively higher thermal resistance in non-dryland ecosystems. EGO again showed reliable performance during this event, capturing GPP anomalies (R2= 0.40) and associated changes in diurnal metric (Figs. S17d–f and S23). Taken together, these findings indicate that EGO can reliably characterize the diurnal GPP cycle across a range of hydroclimatic conditions, including both normal and extreme years and both dryland and non-dryland ecosystems, providing a useful dataset for future studies of sub-daily vegetation carbon dynamics.

In this study, we developed EGO, a global hourly GPP product generated using a causal knowledge-driven machine learning framework that integrates the PCMCI causal discovery algorithm with an XGBoost model. EGO demonstrated strong performance across multiple dimensions: it achieved high accuracy in simulating tower-based hourly GPP (Sect. 4.1) and reconstructing diurnal GPP curves (Sect. 4.2), reliably characterized diurnal GPP dynamic at the global scale (Sect. 4.3), and accurately captured photosynthetic diurnal responses under extreme drought and heatwave events (Sect. 4.4). The robust performance of EGO GPP stems from several advances in both model architecture and input variables.

Firstly, EGO employs a semi-empirical modelling framework based on PCMCI-XGBoost model, enabling strong nonlinear representation while explicitly accounting for causal effect of predictor variables on GPP (Garcia-Leon et al., 2021). Identifying robust causal structures is critical for data-driven machine learning, as the magnitude of hourly GPP response to environmental changes strongly depends on the strength of specific causal effects rather than mere statistical correlations (Galytska et al., 2023; Krich et al., 2020; Yuan et al., 2022). It has been successfully demonstrated in other complex ecological tasks, such as estimating wetland methane emissions (Yuan et al., 2024) and ecosystem respiration (Zhao et al., 2025; Zhao and Zhu, 2025). The superior performance observed in our evaluations also confirms the effectiveness of this causal modelling framework. Secondly, EGO integrates multidimensional predictors encompassing vegetation, radiation, both long-term and short-term water availability, and other environmental constraints, thereby allowing the model to better learn the drivers of the diurnal GPP cycle (Anav et al., 2015; Hou et al., 2022; Pei et al., 2025; Wang et al., 2023). Unlike previous products that primarily relied on short-term water variability predictors, EGO explicitly introduced the aridity index as a proxy for long-term climatic water balance which plays a pivotal role in regulating the diurnal hysteresis (Li et al., 2023; Liu et al., 2024). As shown in Fig. S9, incorporating aridity index as a predictor substantially improved hourly GPP simulation accuracy especially in drylands. This improvement directly mitigates the uncertainties of existing products in these regions. Thirdly, we used sub-pixel vegetation endmember abundance data rather than discrete vegetation type as model inputs. Traditional machine learning upscaling often assigns a single dominant vegetation class to each pixel, which neglects surface heterogeneity and introduces mixed-pixel bias (Bodesheim et al., 2018; Nelson et al., 2024). By contrast, the endmember abundances enable EGO to resolve continuous transitions among vegetation types and capture detailed local variations in photosynthesis (Qin et al., 2023). Finally, EGO adopted a 10-fold cross-validation ensemble strategy, where predictions from all folds were averaged to produce the final GPP output. This ensemble integration not only stabilized the learning process against random sample partitioning but also improved generalization by aggregating information learned from diverse training subsets (Han et al., 2025; Kang et al., 2025). Consequently, the ensemble output demonstrated higher accuracy than any single-fold model and reduced the influence of potential overfitting. To our knowledge, EGO represents the highest spatial resolution (0.05°) global hourly GPP product currently available and may extends the potential applicability for addressing localized ecological questions and refining global carbon budget assessments. Together, these advances in modelling strategy and data integration underpin the strong performance of EGO GPP and its capability to reproduce sub-daily vegetation carbon dynamics across biomes and climatic conditions.

We developed quantitative metrics to characterize the morphology of GPP diurnal cycles, describing the curves in terms of absolute magnitude, asymmetry, and midday depression. Beyond providing site-scale diagnostics, we mapped the global distributions of these metrics and analyzed their variations under extreme drought and heatwave events, thereby offering a holistic perspective for assessing how vegetation photosynthesis responds to environmental perturbations. This framework offers a new analytical perspective for evaluating hourly GPP products, shifting the focus beyond traditional accuracy performance to diurnal dynamics.

Previous studies on diurnal photosynthetic dynamics have mostly been confined to regional or site scales (Lin et al., 2019; Zhu and Zhu, 2025). Our earlier study reported that the 2020 U.S. heatwave substantially altered diurnal GPP patterns especially in western dryland ecosystems, causing widespread afternoon depression and earlier diurnal centroids (Li et al., 2023). Zhang et al. (2023a) used OCO-3 and tower observations to reveal distinct morning-afternoon responses of Amazon forests to atmospheric warming and drying, where enhanced morning photosynthesis partially compensated for afternoon declines driven by high VPD. Xiao et al. (2021) combined Himawari-8 AHI with tower observations to demonstrate substantial afternoon GPP reductions during Australian heatwaves. These findings collectively highlight that vegetation can rapidly adjust to short-term environmental stresses through physiological processes such as stomatal regulation and photoprotective mechanisms which cannot be adequately captured by traditional daily or monthly GPP data (Novick et al., 2024; Xiao et al., 2025). With the increase in the frequency of extreme climate events, midday and afternoon depression of photosynthesis may intensify globally, underscoring the need for high-resolution GPP products capable of resolving these fine-scale dynamics (Li et al., 2021; Zhang et al., 2023c). Our systematic evaluation confirms EGO's capability in this regard. During the two representative extreme events in this study, EGO accurately captured stress-induced changes in diurnal metrics, including widespread GPP reductions, earlier peaks and centroids, and intensified asymmetry and midday depression.

Given these strengths, EGO GPP holds strong potential as a reliable data source for investigating global diurnal photosynthesis and exploring ecosystem dynamics under extreme stresses in future research. It should be noted that the dominant controls on GPP may vary with temporal aggregation. For example, soil moisture effects can become more important at longer time scales because of ecosystem water-memory effects, whereas temperature may play a stronger role in regulating annual-scale carbon uptake (Green et al., 2019; Jung et al., 2017). Here, EGO provides a useful opportunity for future studies to examine how the relative controls of radiation, temperature, and soil moisture shift from hourly to seasonal and annual time scales. Moreover, EGO can serve as a valuable benchmark for Earth system models, supporting the development and evaluation of next-generation carbon cycle simulations at hourly time steps (Boucher et al., 2020; Liu et al., 2024; Seferian et al., 2019; Voldoire et al., 2019). We will continue to update and distribute EGO to facilitate advanced understanding of global vegetation dynamics in a changing world.

Although EGO demonstrated robust performance, several limitations remain. Due to the uneven spatial distribution of global flux sites, most observations were concentrated in Europe and America (Fig. 1). As a result, the model is trained primarily on these regions, which may introduce bias at the global scale (Bodesheim et al., 2018; Nelson et al., 2024). While this is a prevalent issue in current studies, we anticipate that these uncertainties will be progressively alleviated by the continuous expansion of flux observation networks and strengthened international collaboration on data sharing (Xiao et al., 2025). Another source of uncertainty arises from the fact that tower-based GPP is not directly measured but inferred from net ecosystem exchange through flux partitioning. Different partitioning approaches may introduce some uncertainty in estimated GPP and its diurnal patterns, as they rely on different assumptions about ecosystem respiration (Keenan et al., 2019; Ranjbar et al., 2026). Moreover, we used MERRA-2 (0.5° × 0.625°) and CarbonTracker (3° × 2°) datasets to represent the effects of diffuse radiation fraction and atmospheric CO₂ concentration on diurnal GPP dynamics. Their spatial resolutions are coarser than the EGO's 0.05° grid, some spatial details may have been smoothed, especially in heterogeneous landscapes. Also, the intrinsic uncertainties associated with these reanalysis products may propagate into the model, ultimately affecting GPP estimation accuracy (Wang, 2026). Nevertheless, to our knowledge, these datasets were currently the good available sources that provide hourly variable, thus remain a reasonable compromise that balance data quality and feasibility. In addition, potential uncertainties may arise in mountainous regions. In complex terrain, topographic effects on solar radiation and satellite observation geometry are difficult to fully correct, potentially impacting model performance (Cao et al., 2025; Hao et al., 2022). Since we used MCD12Q1 product, the model cannot capture GPP variations caused by intra-annual land use transitions (e.g., seasonal deforestation or crop rotation) that occur within a single year. Finally, limited by computational resources and data processing efficiency, the released EGO GPP was currently provided as monthly-averaged hourly data. Despite this, EGO is sufficient to support a wide range of studies focusing on diurnal photosynthesis dynamics. Looking ahead, our roadmap includes leveraging advanced cloud computing to produce a daily-hourly product, which will significantly enhance the capability to resolve short-term ecosystem processes.