Generating the 500 m Global Satellite Vegetation Productivity Phenology Product from 2001 to 2020

Highlights

What are the main findings?

• A novel global 500 m vegetation productivity phenology (VPP) dataset (2001–2020) was developed from the GLASS GPP product.
• The dataset shows substantially higher accuracy and spatial integrity than the widely used MCD12Q2 VGP product.

What is the implication of the main finding?

• GLASS VPP provides a more ecologically meaningful representation of vegetation phenology and carbon dynamics.
• It offers a valuable resource for phenology modeling, carbon cycle research, and ecological forecasting under climate change.

Abstract

Accurate monitoring of vegetation phenology is vital for understanding climate change impacts on terrestrial ecosystems. While global vegetation greenness phenology (VGP) products are widely available, vegetation productivity phenology (VPP), which better reflects ecosystems’ carbon dynamics, remains largely inaccessible. This study introduces a novel global 500 m VPP dataset (GLASS VPP) from 2001 to 2020, derived from the GLASS gross primary productivity (GPP) product. Validation against three ground-based datasets—Fluxnet 2015, PhenoCam V2.0, and PEP725—demonstrated the dataset’s superior accuracy. Compared to the widely used MCD12Q2 VGP product, GLASS VPP reduced RMSE and bias by 35% and 63%, respectively, when validated against Fluxnet data. It also showed stronger correlations than MCD12Q2 when compared with PhenoCam (195 sites) and PEP725 (99 sites) observations, and it captured spatial and altitudinal phenology patterns more effectively. Overall, GLASS VPP exhibits a higher spatial integrity, stronger ecological interpretability, and improved consistency with ground observations, making it a valuable dataset for phenology modeling, carbon cycle research, and ecological forecasting under climate change.

1. Introduction

As fundamental components of terrestrial ecosystems, plants exhibit high sensitivity to environmental fluctuations [1]. Vegetation phenology, representing the temporal patterns of plants’ physiological responses and adaptations to environmental conditions [2], has been widely employed as a critical proxy for investigating long-term climate changes [3,4]. Over the past decades, vegetation phenology studies have evolved from traditional ground-based observations, such as those collected by the Pan European Phenology Project (PEP725) [5] and eddy covariance flux towers [6], to satellite-based remote sensing approaches. Compared to traditional ground-based phenology observations, remote sensing phenology offers the distinct advantage of extensive spatial coverage, making it particularly suitable for large-scale climate change analyses [7]. Consequently, it has been broadly applied in various research fields, including land cover classification [8,9,10,11], climate change detection [7,12,13], crop yield prediction [14,15], and quantification and determination of water cycles, energy balance, and carbon fluxes [16,17].

Significant efforts have been dedicated to generating land surface phenology (LSP) products using diverse remote sensing observations, substantially advancing the study of vegetation phenology (Table 1). Primarily derived from spectrally reflected information captured by remote sensors [18], these data sources exhibit varying characteristics of vegetation physiological processes, leading to distinct phenological interpretations [19]. Early satellite-based phenology research predominantly utilized the Normalized Difference Vegetation Index (NDVI), owing to its ease of calculation and robustness against variations in topography, illumination, cloud interference, and atmospheric conditions [20]. However, NDVI exhibits reduced sensitivity in dense vegetation due to its tendency to saturate under high biomass conditions [21]. To mitigate the influence of soil background and reduce the saturation effects associated with NDVI, the Enhanced Vegetation Index (EVI) was developed and has since been widely applied in LSP studies [22,23]. However, it also presents limitations, including a sensitivity to sun-sensor geometry that leads to BRDF effects, reduced responsiveness in low-biomass regions, and dependency on specific sensor configurations [24,25,26]. The Leaf Area Index (LAI), as a biophysical parameter indicative of canopy structure and vegetation density, provides higher ecological relevance and has been shown to yield more accurate and reliable phenological estimates than traditional vegetation indices-based approaches [27]. However, to effectively retrieve LSP from LAI time series, temporal smoothing is required to reduce noise and extract seasonal signals, and differences between LAI products must be carefully considered to ensure consistency [28]. Despite their respective limitations, these indicators remain central to the development of global phenology products (Table 1). Unlike previous indicators that primarily reflect canopies’ greenness or structure, both gross primary productivity (GPP) and sun-induced chlorophyll fluorescence (SIF) are more directly linked to photosynthetic activity, offering improved sensitivity to vegetation phenology [29]. Field observations have shown strong consistency between canopy phenology and GPP-derived phenological metrics [30,31], while SIF exhibits a robust linear relationship with GPP throughout the growing season, supporting its capacity to bridge canopy and satellite measurements [32,33]. Moreover, SIF has demonstrated superior accuracy in capturing key phenological transitions such as the onset of spring and the decline in photosynthetic activity in autumn across various biomes and regions [34,35]. These findings highlight the need to better utilize photosynthesis-based metrics with a focus on GPP, to improve our understanding of vegetation phenology and its role in the terrestrial carbon cycle.

GPP, which is defined as the rate of carbon dioxide uptake by plants through photosynthesis and the subsequent conversion of light energy into chemical energy [36], is a critical parameter in carbon cycle modeling [37]. Annual terrestrial GPP is highly sensitive to plant phenology metrics, such as the start and end of the growing season (SOS and EOS) [38,39]. Vegetation productivity phenology (VPP) metrics, nowadays primarily extracted from ground-based GPP time series using flux tower observations [40], provides valuable insights into terrestrial carbon cycle dynamics due to its clear ecological interpretability [41,42,43,44]. Xu et al. analyzed long-term trends in phenological and productivity metrics—such as the start, end, and length of the growing season, peak GPP, and seasonal/annual cumulative GPP—using a dataset comprising 440 site-years of eddy covariance measurements. Based on these metrics, they developed a phenological–physiological statistical model to predict long-term variations in annual GPP [45]. Wu et al. examined the interannual anomalies of net ecosystem productivity (NEP) across 212 site-years of flux observations in North America and Europe. Their findings highlighted that autumn phenological indicators derived from carbon fluxes play a more significant role than spring indicators in determining year-to-year variability in forest NEP [46]. Migliavacca et al. demonstrated that incorporating canopy-level productivity phenology substantially improves the modeling performance of respiration dynamics in deciduous forest ecosystems, especially regarding the temporal patterns of ecosystem respiration [47]. Given the greater sensitivity of vegetation photosynthesis to climate variations compared to greenness [48], VPP is more suitable for assessing the ecosystem’s carbon uptake duration and climate anomaly-induced changes in the terrestrial carbon cycle than VGP [49]. Ground-based VPP metrics have been extensively employed to validate remote sensing-derived LSP datasets [19,50,51,52,53]. Compared to traditional VGP, which primarily reflects changes in canopy color or structure, VPP offers a more ecophysiologically relevant representation of plants’ functional dynamics [1]. Ground-based flux tower observations have demonstrated that GPP-derived phenological metrics—such as the timing of photosynthetic onset and cessation—align more closely with carbon flux variations than VI-based phenology [39]. While VGP products are widely available and useful for large-scale ecosystem monitoring, they tend to lag behind or misrepresent true photosynthetic activity due to their indirect relationship with carbon assimilation [54]. In contrast, VPP captures vegetation responses that are more directly linked to carbon uptake, thus providing a stronger explanatory power in carbon budget modeling and climate feedback assessments [55]. Increasing evidence from site-level and regional studies supports the integration of VPP metrics as a more accurate and meaningful indicator in terrestrial carbon cycle research [56]. However, a systematic validation and comparative evaluation between VPP and VGP using ground-based datasets remains essential for understanding their respective strengths and limitations across spatial and temporal scales [57]. The VPP dataset developed by Xu et al. is limited in that phenological parameters were inferred through a modeling framework based on MODIS GPP rather than being directly extracted from the GPP time series [57]. Despite ongoing efforts to develop satellite-based VPP products, the absence of a publicly accessible, long-term global dataset significantly limits our understanding of the large-scale responses of vegetation’s carbon uptake capacity to climate change.

The increasing demand for GPP data, coupled with advancements in GPP retrieval algorithms, has facilitated the production of large-scale GPP datasets [58]. These developments enable the generation of global, long-term satellite-based VPP products. This study proposes a synthesized productivity phenology derivation scheme. Applying this scheme to the GPP product of the Global Land Surface Satellite (GLASS) products suite [59], which was proven to perform stably across different ecosystems [37,60,61], we generated a global 500 m GLASS VPP dataset spanning 2001 to 2020. To rigorously assess the accuracy of our VPP product, we conducted a comprehensive validation using ground-based observations from Fluxnet, PhenoCam, and the PEP725 (Pan European Phenology Project) network. Additionally, an intercomparison with the widely used MCD12Q2 vegetation phenology product was performed to evaluate their respective capabilities in characterizing surface plant growth patterns.

Table 1. Several representative global LSP products and their key characteristics.

Product	Period	Spatial Extent	Spatial Resolution	Data Source	Methodology	References
MCD12Q2 v061	2001–2021	Global	500 m	MODIS NBAR-EVI2	Threshold	[62]
VIPPHEN_NDVI v004	1981–2014	Global	0.05°	AVHRR LTDR-V4 NDVI (1981–1999) MODIS C5 NDVI (2000–2014)	Threshold	[12]
VIPPHEN_EVI2 v004	1981–2014	Global	0.05°	AVHRR LTDR-V4 EVI2 (1981–1999) MODIS C5 EVI2 (2000–2014)	Threshold	[12]
VNP22Q2 v001	2013–present	Global	500 m	VIIRS NBAR EVI2	Derivative	[63]
VNP22C2 v001	2013–present	Global	0.05°	VIIRS NBAR EVI2	Derivative	[63]
MOD09A1P_NDVI	2001–2020	America	500 m	MOD09A1 NDVI	Threshold	[64]
MOD09A1P_EVI	2001–2020	America	500 m	MOD09A1 EVI	Threshold
MOD09Q1P_NDVI	2001–2020	America	250 m	MOD09A1 NDVI	Threshold
MOD09Q1P_EVI	2001–2020	America	250 m	MOD09A1 EVI	Threshold
MOD15PHN	2001–2020	America	1 km	MCD15A2 LAI	Threshold	[64]
MSLSP30NA1	2016–2018	North America	30 m	HLS EVI	Threshold	[65]
ChinaCropPhen	2000–2015	China	1 km	GLASS LAI	Threshold/Derivative	[66]
Qinghai–Tibet Plateau Phenology Dataset	2000–2016	Qinghai–Tibet Plateau	500 m	MOD09A1 EVI	Threshold	[67]
MODIS Sanjiangyuan Phenological Phase Dataset	2001–2020	Sanjiangyuan	250 m	MOD13A2 NDVI	Threshold/Derivative	[68]

Table 1. Several representative global LSP products and their key characteristics.

Product	Period	Spatial Extent	Spatial Resolution	Data Source	Methodology	References
MCD12Q2 v061	2001–2021	Global	500 m	MODIS NBAR-EVI2	Threshold	[62]
VIPPHEN_NDVI v004	1981–2014	Global	0.05°	AVHRR LTDR-V4 NDVI (1981–1999) MODIS C5 NDVI (2000–2014)	Threshold	[12]
VIPPHEN_EVI2 v004	1981–2014	Global	0.05°	AVHRR LTDR-V4 EVI2 (1981–1999) MODIS C5 EVI2 (2000–2014)	Threshold	[12]
VNP22Q2 v001	2013–present	Global	500 m	VIIRS NBAR EVI2	Derivative	[63]
VNP22C2 v001	2013–present	Global	0.05°	VIIRS NBAR EVI2	Derivative	[63]
MOD09A1P_NDVI	2001–2020	America	500 m	MOD09A1 NDVI	Threshold	[64]
MOD09A1P_EVI	2001–2020	America	500 m	MOD09A1 EVI	Threshold
MOD09Q1P_NDVI	2001–2020	America	250 m	MOD09A1 NDVI	Threshold
MOD09Q1P_EVI	2001–2020	America	250 m	MOD09A1 EVI	Threshold
MOD15PHN	2001–2020	America	1 km	MCD15A2 LAI	Threshold	[64]
MSLSP30NA1	2016–2018	North America	30 m	HLS EVI	Threshold	[65]
ChinaCropPhen	2000–2015	China	1 km	GLASS LAI	Threshold/Derivative	[66]
Qinghai–Tibet Plateau Phenology Dataset	2000–2016	Qinghai–Tibet Plateau	500 m	MOD09A1 EVI	Threshold	[67]
MODIS Sanjiangyuan Phenological Phase Dataset	2001–2020	Sanjiangyuan	250 m	MOD13A2 NDVI	Threshold/Derivative	[68]

2. Materials and Methods

2.1. Data

2.1.1. GLASS GPP Product

We used the GLASS GPP V60 dataset from 2000 to 2021 as the basic data source to derive the key phenology metrics and generate the global VPP product in this study [69]. The GLASS GPP dataset has a spatial resolution of 500 m and provides vegetation productivity information for the global land surface every eight days. The GLASS GPP dataset was generated with the improved EC-LUE model, which was proven to perform stably across different ecosystems [37,60,61]. The GLASS GPP product integrates multiple satellite-derived datasets. Specifically, it utilizes net radiation (Rn), air temperature (T), relative humidity (Rh), and photosynthetically active radiation (PAR) obtained from MERRA (Modern-Era Retrospective Analysis for Research and Applications). Additionally, global 8-day MODIS LAI data (MOD15A2) and 16-day MODIS NDVI data (MOD13A2) are incorporated to support vegetation-related inputs. These data were obtained from a total of 54 flux tower sites provided by the AmeriFlux (http://public.ornl.gov/ameriflux (accessed on 26 August 2025)) and EuroFlux networks (https://www.europe-fluxdata.eu/ (accessed on 26 August 2025)) [37]. The GLASS GPP products and their generation algorithms are extensively used in determining spatiotemporal variations in ecosystem productivity and global carbon sink assessments [70,71,72].

2.1.2. Ground Observations

To comprehensively assess the quality of the GLASS VPP product generated in this study, we collected three kinds of ground observations, including the in situ measured GPP from the Fluxnet program, the PhenoCam Dataset V2.0, and the plant phenology observations from PEP725 network. The spatial distribution of these ground observation sites is shown in Figure 1.

Fluxnet sites and phenology from daily GPP. To assess the quality of the GLASS VPP metrics, we leveraged the daily GPP observation from the latest Fluxnet 2015 dataset. The Fluxnet program employs the eddy covariance technique to measure the exchange of water, carbon, and energy between ecosystems and the atmosphere [40,73], providing valuable daily GPP observations. We applied the phenology extraction algorithm outlined in Section 2.2.2 and Section 2.2.3 to the Fluxnet GPP time series to derive ground vegetation productivity phenology (VPP_Ground). Using the daily GPP data from 2001 to 2014, a total of 574 Fluxnet phenology metric samples over 102 sites were obtained and employed to assess the accuracy of the satellite-derived phenology products. It is worth noting that although the GLASS GPP product was calibrated using 54 flux tower sites (mainly from AmeriFlux and EuroFlux), our study utilized a total of 102 sites from the FLUXNET2015 dataset for validation. The majority of these sites were not involved in the GLASS GPP model’s development, ensuring a high degree of independence between the validation data and the original calibration dataset. However, we also acknowledge that the spatial distribution of high-quality flux sites is uneven globally. In particular, sites located in the Southern Hemisphere and tropical regions are relatively sparse and often suffer from issues such as short observation periods, data discontinuities, or inconsistent temporal quality. These limitations constrain the representativeness of ground-based validation in these regions and highlight the need for enhanced flux observation capacity in future global phenology research.

PhenoCam and phenology from daily GCC (Green Chromatic Coordinate). This study employed the digital image statistics product from the PhenoCam dataset v2.0, which provides a time series of visible band reflected information at various global sites from 2000 to 2014. For each site, the dataset provides statistical values of a predefined region of interest (ROI) for a specific vegetation type. We calculated the GCC index with these three visible bands [74] to characterize the vegetation greenness dynamics, where R, G and B represent the red, green, and blue color channel values, respectively. This index normalizes green intensity against total brightness, thereby minimizing the effects of varying illumination conditions. For each image, we extracted the region of interest (ROI) and calculated the average GCC value across all pixels within the ROIs. The GCC time series was smoothed, and low-quality data were manually removed to extract ground vegetation greenness phenology (VGP_Ground). Finally, a total of 745 PhenoCam VGP_Ground samples over 195 camera sites were obtained.

$$GCC={\displaystyle \frac{G}{R+G+B}}$$

(1)

The plant phenology observation from the PEP725 network. Plant phenological data were obtained from the PEP725 network. This ground-based European database provides long-term, spatially dense observations of key plant phenophases, including leaf unfolding, flowering’s onset, and leaf fall [5,71,75]. Phenological stages were determined using the BBCH (Biologische Bundesanstalt, Bundessortenamt und Chemische Industrie) scale. The BBCH scale is a standardized phenological coding system that describes the growth stages of plants throughout their entire life cycle, from germination to senescence [76]. The full flowering periods of four fruit trees in Germany, Malus domestica, Prunus cerasus, Prunus avium, and Pyrus communis, were selected in this study, due to their relatively concentrated and brief flowering periods, ensuring high data quality. Full flowering was defined as the stage at which 50% of flowers were open, with the possibility of initial petal fall. PEP725 serves as a reliable benchmark for evaluating the quality of various remote sensing phenology products and enables robust inter-dataset comparisons [77].

2.1.3. MCD12Q2 Phenology Dataset

To assess the performance of the GLASS VPP product, we conducted an intercomparison with the satellite-derived vegetation greenness phenology (VGP) from the MCD12Q2 v6.1 product. The MCD12Q2 phenology metrics were derived from the Enhanced Vegetation Index (EVI) using a threshold-based method [62]. The MCD12Q2 VGP product has been widely utilized for determining long-term changes in vegetation phenology [78,79,80]. Since the definition of phenology parameters of MCD12Q2 differs from our VPP metrics, the corresponding relationships between the two products’ phenology metrics are outlined in Figure 2 and

Table 2. The correspondence definition of phenology metrics between VGP (MCD12Q2) and VPP.

VPP (This Study)	VGP (MCD12Q2)
Start of Season (SOS)	Green-up: The first date within the green-up segment where the EVI2 time series crosses 15%
Mid of Green-Up (MOG)	Mid Green-Up: The first date within the green-up segment where the EVI2 time series crosses 50%
Peak of Season (POS)	Peak: The date at which the maximum value of the EVI2 time series is reached
Mid of Brown-Down (MOB)	Mid Green-Down: The first date within the green-down segment where the EVI2 time series crosses 50%
End of Season (EOS)	Dormancy: The first date within the green-down segment where the EVI2 time series crosses 15%

.

2.2. Method

Our objective is to generate the global VPP by integrating two independent methods based on the GLASS GPP product. To achieve this goal, we designed a framework composed of several steps (Figure 3). The detailed technical processes are introduced in the following sections.

2.2.1. Identification of Natural Vegetation with a Single Growing Season

In order to minimize the influence of non-vegetative and agricultural pixels on the GLASS VPP mapping, we used the MCD12Q1 land cover product (University of Maryland classification criteria) to identify natural vegetation distributions [81]. Pixels classified as water bodies, urban and built-up lands, barren, and croplands were excluded. To eliminate pixels with indistinct growing seasons, such as tropical rainforests exhibiting a minimal GPP seasonal variability, we applied a Fast Fourier Transform (FFT) to pixel-level GPP time series to analyze the frequency characteristics. Pixels with a dominant frequency that matched the temporal coverage of the GPP time series from 2000 to 2021 were classified as “seasonally clear” and retained for subsequent phenology extraction. To address asynchronous vegetation growth cycles in the Northern and Southern Hemispheres, the original GPP time series was partitioned based on the characteristics of its maximum amplitude component, ensuring each segment represented a complete growing season. It is important to note that in this study, the term “tropical forest” does not refer to a specific land cover classification (e.g., evergreen, semi-deciduous, or monsoon forests) but rather to areas with persistently high GPP values and no apparent seasonal cycles as determined by frequency–domain analysis. This functional definition is based on vegetation productivity dynamics rather than static land cover labels.

2.2.2. GPP Time-Series Fitting

To derive GLASS VPP metrics, the temporal variation in the GPP time-series dataset for an entire growing cycle required modeling into a continuous function. We employed the double logistic function (DL, Equation (2)), a commonly used curve-fitting method in VI-based LSP determination studies, to simulate continuous vegetation growth variations.

$$f\left(t\right)=\mathrm{ }{\alpha }_1+{\displaystyle \frac{{\alpha }_2}{1+e^{{-\alpha }_1(t-{\beta }_1)}}}-\mathrm{ }{\displaystyle \frac{{\alpha }_3}{1+e^{{-\alpha }_2(t-{\beta }_2)}}}$$

(2)

where f(t) is the observed GPP at a day of the year (DOY), and α₁, α₂, α₃, β₁, β₂, α₁, and α₂ are the parameters of the DL function, which together control the shape of the planted growth curve [82]. The DL function is characterized by seven parameters, allowing for distinct background values on either side of the non-growth season and broader applicability [83].

To evaluate the accuracy of the DL function in capturing observed GPP seasonal variations, we used the coefficient of determination (R²) to measure the reliability of the fitted curve.

$$R^2={\displaystyle \frac{{{\sum }_{i=1}^{46}(f_i^{\prime }-\overline{f})}^2}{{{\sum }_{i=1}^{46}(f_i-\overline{f})}^2}}$$

(3)

2.2.3. VPP Metrics Extraction

Vegetation productivity phenology (VPP) is defined as the transition dates marking changes in the state of vegetation productivity. With the DL fitted function and GPP time series, five key phenology metrics were extracted in this study, including start of growing season (SOS), end of growing season (EOS), mid of green-up (MOG), mid of brown-down (MOB), and peak of growth season (POS). To take advantage of different phenology extraction algorithms to enhance the robustness of VPP’s determination, we developed a synthesized scheme by combining results from the derivative extremum method (DVM) and the threshold-based method (TBM).

VPP metrics’ determination with DVM. Following the method proposed by Gonsamo et al. [83], the five key VPP metrics could be derived from the fitted DL function with these algorithms in Equations (4)–(8). SOS and EOS were identified as the transition dates of the first maximum value and last minimum value of third-order derivation of the fitted DL function, respectively. MOG and MOB were identified as the transition dates of the first maximum value and last minimum value of first-order derivation of the fitted DL function, respectively. POS was calculated as the mean of MOG and MOB.

$${SOS}_{DVM}={\beta }_1-{\displaystyle \frac{4.562}{2{\alpha }_1}}$$

(4)

$${EOS}_{DVM}={\beta }_2-{\displaystyle \frac{4.562}{2{\alpha }_2}}$$

(5)

$${SOP}_{DVM}={\beta }_1$$

(6)

$${EOP}_{DVM}={\beta }_2$$

(7)

$${POS}_{DVM}={\displaystyle \frac{{\beta }_1+{\beta }_2}{2}}$$

(8)

VPP metrics’ determination with TBM. The VPP metrics from TBM were derived with the simulated daily GPP time series by the DL fitting model for an entire growing season. As outlined in Equations (9) and (10), f(t) represents the simulated GPP on a given day (t); F_m denotes the maximum daily GPP value within the annual GPP time series, with its corresponding date designated as T_m; and F_dl and F_dr represent the mean dormancy productivity before and after the growing season, respectively. The phenology metric, P_TBM, is defined as the date when the simulated GPP time series reaches a specified threshold K_p.

$${\theta }_l\left(t\right)={\displaystyle \frac{f\left(t\right)-F_{dl}}{F_m-F_{dl}}}t<T_M$$

(9)

$${\theta }_r\left(t\right)={\displaystyle \frac{f\left(t\right)-F_{dr}}{F_m-F_{dr}}}t\ge T_M$$

(10)

Generation of synthesized scheme. Reasonable selection of K_p is crucial for extracting P_TBM and effectively synthesizing results from both TBM and DVM. The fundamental principle of K_p‘s determination is to ensure that the phenology metrics extracted by TBM closely correspond to those derived from DVM. This alignment allows for a more robust composition of the phenology metrics from both methods. To establish K_p values for the derivation SOS, EOS, MOG and MOB, a random selection of 100,000 GPP time-series points from regions north of 50°N in 2018 was conducted. Prioritizing pixels (8641) with high goodness of fitting (R² ≥ 0.99) values from the DL model fitting ensured the reliability of these GPP time-series samples. Subsequently, a comparative analysis identified the K_p values that yielded the highest consistency between P_TBM and P_DVM for each phenology metric, establishing the final thresholds.

Given the complementary strengths and limitations of the DVM and TBM in phenology metric determination [84], we developed a hybrid approach to optimize VPP’s estimation. Using the methods outlined in the previous steps, pixel-level phenology metrics were extracted separately from both DVM (P_DVM) and DBM (P_TBM). The absolute difference (Δ) between these metrics was then calculated for each pixel. If the Δ exceeded twice the temporal resolution of the original input GPP time series (8-day in this study), the uncertainty of P_DVM was deemed excessive, and only the P_TBM result was retained. Otherwise, the mean value of P_DVM and P_TBM was used as the final synthesized VPP estimate.

2.2.4. Validation of VPP Metrics

To assess the accuracy of our VPP product, phenology metrics were validated against ground observations from Fluxnet, PhenoCam V2.0, and PEP725. Numerous studies have used PhenoCam and Fluxnet datasets to validate remote sensing phenology [74,85,86,87,88,89]. Correlation coefficients (R), root mean square error (RMSE), and bias were calculated using Equations (11)–(13). The widely used MCD12Q2 VGP dataset was also synchronously evaluated and compared to VPP. For VPP pixels with multiple ground observations, mean values were calculated to represent the ground data.

$$RMSE=\sqrt{{\displaystyle \frac{{\sum }_{i=1}^N{\left(y_i^{\prime }-y_i\right)}^2}{N}}}{\displaystyle (}\mathrm{d}\mathrm{a}\mathrm{y}{\displaystyle )}$$

(11)

$$Bias=mean{\displaystyle (}{y}^{\prime }{\displaystyle )}-mean{\displaystyle (}y{\displaystyle )}{\displaystyle (}\mathrm{d}\mathrm{a}\mathrm{y}{\displaystyle )}$$

(12)

$$R={\displaystyle \frac{{\sum }_{i=1}^N(y_i-\overline{y})(y_i^{\prime }-\overline{y^{\prime }})}{\sqrt{{\sum }_{i=1}^N{(y_i-\overline{y})}^2}\sqrt{{\sum }_{i=1}^N{(y_i^{\prime }-\overline{y^{\prime }})}^2}}}$$

(13)

where y_i is the phenology metrics from ground observation, y_i^’ represents the satellite derived phenology metrics, and N is the number of samples employed for validation.

3. Results

We firstly developed the VPP metric extraction scheme by combining two independent methods. Then, using the synthesized algorithm, we produced the global VPP product from 2001 to 2020 based on the GLASS GPP data. Both GLASS VPP and MCD12Q2 VGP metrics were subsequently validated against ground phenology observations.

3.1. Determination of the Synthesized Phenology Extraction Scheme

Based on the scheme depicted in Section 2.2.3, a sensitivity analysis was conducted to assess impact of varying K_p values on the consistency of phenology metrics derived from both methods. The results, presented in Figure 4 and

Table 3. Comparative analysis of phenology metrics extracted using TBM and DVM, employing both optimal and final determined K_p values.

	SOS		EOS		MOG		MOB
Kp	9.5%	10%	8.8%	10%	48.8%	50%	46.6%	50%
R2	99.9%	99.8%	99.8%	98.4%	99.7%	99.6%	99.6%	97.3%
Bias (day)	0.112	0.62	−0.1	−1.671	0.179	0.59	−0.108	−1.619
MAD (day)	0.364	0.631	0.413	1.671	0.491	0.674	0.551	1.62
RMSD (day)	0.609	0.795	0.686	1.768	0.605	0.765	0.676	1.679

, indicate that the optimal K_p values for SOS, MOG, EOS, and MOB are 9.5% (left), 48.8% (left), 8.8% (right), and 46.6% (right), respectively.

To ensure comparability with existing vegetation phenology studies while maintaining the accuracy of our metrics, we adopted commonly used thresholds of 10% and 50% for deriving the start/end and middle status of vegetation green-up and brown-down, respectively. While optimal K_p values might offer slightly improved performance, adhering to these common thresholds facilitates comparison with other studies. As indicated in Table 3, adjusting the Kp values to these standard thresholds had a minimal effect on the model’s performance, as measured by R², bias, mean absolute deviation (MAD), and root mean square deviation (RMSD). Moreover, the consistency of phenology metrics derived from both methods remained within a two-day margin, indicating that our approach effectively balances accuracy and comparability.

3.2. VPP Validation

3.2.1. VPP Metrics’ Validation Against Fluxnet Observations

Figure 5 presents a comparative analysis of our GLASS VPP and MCD12Q2 VGP datasets validated against 574 Fluxnet ground phenology metric samples. The results unequivocally demonstrate the superior accuracy of all five GLASS VPP metrics derived in this study compared to MCD12Q2 VGP. Correlation coefficients between GLASS VPP metrics and Fluxnet observations ranged from 0.59 to 0.82, significantly exceeding the range of 0.45 to 0.73 observed for MCD12Q2 VGP metrics. Moreover, GLASS VPP metrics achieved much lower RMSEs and biases, indicative of improved accuracy and reduced systematic errors. The validating RMSEs of VPP are 23.3, 17.2, 9.5, 17.4, and 22.0 days for SOS, MOG, POS, MOB, and EOS, respectively, with a mean decrease of 35% compared to MCD12Q2 VGP. Additionally, GLASS VPP metrics exhibited significantly lower biases against Fluxnet ground observations, with a mean decrease of 63% compared to MCD12Q2 VGP. The regression lines of GLASS VPP metrics against Fluxnet observations exhibited a closer proximity to the 1:1 line, further underscoring their superior performance relative to the MCD12Q2 VGP dataset.

To mitigate the potential impact of unstable GPP measurements, attributable to environmental noise and interannual variability, on the validation results, we calculated multi-year mean ground phenology metrics (VPP_Ground) and corresponding satellite-derived phenology metrics from GLASS VPP and MCD12Q2 VGP for each site. This approach significantly enhanced the reliability of the validation process. Figure 6 presents the validation results for these five phenology metrics. The results demonstrate that both products exhibited improved accuracy, with GLASS VPP metrics consistently outperforming MCD12Q2 VGP metrics. Correlation coefficients between GLASS VPP metrics and Fluxnet observations ranged from 0.72 to 0.88, significantly exceeding the range of 0.47 to 0.86 for MCD12Q2 VGP metrics. Moreover, GLASS VPP metrics achieved lower RMSEs and biases, indicative of superior accuracy and reduced systematic errors.

3.2.2. VPP Metrics Validation Against PhenoCam Observations

The GLASS VPP and MCD12Q2 VGP metrics were further validated against 745 phenology camera observation samples from PhenoCam 2.0. Ground vegetation greenness phenology metrics were derived from daily GCC time series calculated within predefined regions of interest (ROIs) on PhenoCam images. While MCD12Q2 VGP metrics were theoretically more aligned with PhenoCam observations, GLASS VPP metrics still demonstrated a competitive performance (Figure 7). Correlation coefficients between GLASS VPP and PhenoCam observations consistently exceeded those observed for MCD12Q2 VGP for all five phenology metrics. The validation results for RMSEs and biases revealed a divergent performance among the five metrics. GLASS VPP metrics exhibited lower RMSEs and biases for MOG, POS, and MOB but slightly higher RMSEs and biases for SOS and EOS compared to MCD12Q2 VGP metrics. The lower consistency of thee GLASS VPP SOS and EOS against PhenoCam observations may be primarily attributed to theoretical differences between VPP and VGP, leading to larger biases and subsequently higher RMSE values.

The multi-year mean phenology metrics were also calculated at each PhenoCam site and employed to validate the GLASS VPP and MCD12Q2 VGP dataset (Figure 8). Although both products demonstrated improved validation accuracy after accounting for interannual variability, the improvements were not substantial. The validation results revealed that GLASS VPP metrics exhibited slightly superior performance compared to MCD12Q2 VGP, particularly for the three metrics of MOG, POS, and MOB.

3.2.3. VPP Metrics’ Validation Against PEP725 Observations

The GLASS VPP and MCD12Q2 VGP metrics were further validated against ground phenology records from the PEP725 network. Unlike the other two ground observation datasets, PEP725 provides direct plant phenology records. Given the inherent challenges of directly comparing satellite-derived phenology metrics and ground phenology records for absolute values, this study focused on assessing the consistency of the two datasets in spatial–temporal variations. Full flowering date records for four fruit trees (Malus domestica, Prunus cerasus, Prunus avium, and Pyrus communis) observed from 2001 to 2015 in Germany were used to evaluate the accuracy of SOS and MOG metrics from GLASS VPP and MCD12Q2 VGP. Figure 9 presents the validation results, demonstrating the superior consistency of VPP metrics with PEP725 phenology records compared to MCD12Q2 VGP. Correlation coefficients between GLASS VPP metrics and PEP725 records ranged from 0.58 to 0.64 for SOS and from 0.60 to 0.63 for MOG, significantly exceeding the range of 0.16 to 0.19 for SOS and 0.27 to 0.29 for MOG observed for MCD12Q2 VGP metrics. Moreover, the regression lines of GLASS VPP metrics against PEP725 records exhibited a much closer proximity to the 1:1 line, further underscoring their superior performance relative to the MCD12Q2 VGP dataset. Compared with SOS, the satellite-derived vegetation MOG was more consistent with the full flowering phenology records.

To assess the consistency of phenology metrics between satellite products and ground observations in interannual variations, the GLASS VPP and MCD12Q2 VGP metrics were further compared against the PEP725 ground phenology records. To eliminate the impact of spatial variations, the 10th, 50th and 90th percentiles of phenology records for these four plants across all sites in Germany were chosen, along with the corresponding satellite-derived metrics. The results shown in Figure 10 and

Table 4. Correlation coefficient between the 10th, 50th, and 90th percentile of the ground observations and the 10th, 50th and 90th percentile of two LSPs.

PEP725	Satellite Metrics	Malus domestica		Prunus cerasus		Prunus avium		Pyrus communis
		VPP	VGP	VPP	VGP	VPP	VGP	VPP	VGP
10th percentile	SOS	0.87	0.49	0.89	0.49	0.90	0.45	0.86	0.49
	MOG	0.90	0.78	0.91	0.81	0.85	0.75	0.90	0.89
50th percentile	SOS	0.85	0.76	0.89	0.82	0.89	0.78	0.84	0.87
	MOG	0.93	0.87	0.93	0.9	0.93	0.88	0.93	0.94
90th percentile	SOS	0.78	0.69	0.87	0.87	0.91	0.89	0.88	0.87
	MOG	0.93	0.85	0.91	0.84	0.92	0.82	0.93	0.88

demonstrate that the interannual variations in the GLASS VPP SOS and MOG are highly consistent with those of plant phenology records for the full flowering period. Although certain biases exist due to the different ecological significance between ground and satellite phenology metrics, the correlation coefficient between the GLASS VPP SOS/MOG and PEP725 plant full flowering records are all nearly 0.9, significantly higher than those of MCD12Q2, demonstrating the significant improvement of GLASS VPP in depicting phenology variations. In general, the satellite-derived MOG is more consistent with the plant full flowering record.

The consistency of satellite-derived phenology with ground observations was further assessed at each site of the PEP725 network. The spatial distribution of correlation coefficients between GLASS VPP, MCD12Q2 VGP, and PEP725 plant observations for all available sites in Germany is shown in Figure 11. The results shown in Figure 11 indicate that over 55% and 68% of the PEP725 sites exhibit significant correlation coefficients greater than 0.7 between ground full flowing records and the GLASS VPP SOS and MOG, respectively. In contrast, only 5% and 16% of the records over these sites show correlation coefficients greater than 0.7 with the corresponding MCD12Q2 VGP metrics. Notably, over 10% of these sites even show negative correlation coefficients with MCD12Q2 VGP metrics.

3.3. Spatial Pattern Comparison of GLASS VPP and MCD12Q2 VPP

To assess the spatial rationality and integrity of GLASS VPP and MCD12Q2 VPP metrics, we compared their phenology mapping results. Figure 12 illustrates the spatial distribution of phenology metrics derived from both products in 2018 over two representative regions: the Ural Mountains, with a significant vertical vegetation distribution (h20v02), and the Tibetan Plateau with sparse vegetation coverage (h25v05). As depicted in Figure 12(b1–b5,c1–c5), the GLASS VPP SOS more accurately captures the delayed vegetation green-up and advanced brown-down at higher elevations compared to MCD12Q2 VGP. Moreover, GLASS VPP metrics exhibit greater spatial continuity than MCD12Q2 VGP, which frequently determines anomalous phenology metrics in numerous pixels. The spatial patterns of phenology metrics in the Tibetan Plateau (Figure 12(e1–e5,f1–f5)) demonstrate GLASS VPP’s superior integrity in sparsely vegetated areas. In this tile, MCD12Q2 VGP has lost over 50% of phenology information over vegetated areas, while GLASS VPP has preserved it effectively.

3.4. Global Mapping of VPP Metrics

Based on the synthesized vegetation phenology determination scheme, global natural vegetation phenology was derived from the GLASS GPP dataset for the period 2001–2020. Figure 13 illustrates the spatial distribution of the five GLASS VPP metrics (SOS, MOG, POS, MOB, EOS) for 2018. In the Southern Hemisphere, phenology metrics of POS, MOB, and EOS were determined for the previous growing season (2017–2018), while SOS and MOG were derived for the current growing season (2018–2019). The resulting phenology maps demonstrate the rationality of GLASS VPP in capturing the latitudinal, vertical, and ecological zonality variations in natural vegetation phenology. For instance, the gradual delay in SOS and advancement in EOS of Northern Hemisphere vegetation with increasing latitude is accurately reflected by GLASS VPP. Moreover, the later SOS and earlier EOS observed in Tibetan Plateau vegetation compared to lower-elevation regions at similar latitudes are also rationally represented.

This study employed the goodness of fit and the absolute difference between two independent algorithms as indicators of the data quality (Figure 14). Regions with a poor model fit goodness were primarily concentrated in arid regions and low-latitude rainforest areas, which is related to the weaker vegetation seasonal growth signals. Besides rainforests and deserts, areas with a high goodness of fit but large discrepancies between the results of two independent algorithms were observed. These regions were mainly concentrated near the Congo basin and South China, where the growing season of vegetation is relatively long and the rate of growth and decline of plant productivity is slower, leading to large uncertainties.

4. Discussion

Vegetation productivity phenology, which can directly characterize vegetation’s growing process and CO₂ uptake duration, is crucial for assessing the variation in territorial productivity and carbon stock [39]. Compared with traditional greenness phenology derived from canopy greenness indices such as NDVI and EVI, productivity phenology has greater advantages in ecological modeling [17] and can better characterize the carbon uptake processes of vegetation. Traditional satellite-based LSP products, primarily derived from vegetation canopy greenness indices like NDVI and EVI, often fall short in accurately characterizing vegetation’s carbon uptake process. While some concurrent VPP metrics can be obtained from flux tower-based Gross Primary Productivity (GPP) measurements, their spatial coverage is limited. To address this gap, we developed a novel global VPP dataset from 2001 to 2020 based on the GLASS GPP product. We employed a hybrid approach, combining the strengths of the popularly used DVM and TBM models, to optimize VPP’s estimation from a time series of satellite observations. This synergistic method effectively leverages the advantages of both models, resulting in a more reliable VPP product.

To assess the quality of our VPP product, we conducted a comprehensive validation with three kinds of ground-based observations. Our results demonstrate that GLASS VPP significantly outperforms the widely used MCD12Q2 VGP product. Specifically, when validated against the Fluxnet ground observations, GLASS VPP exhibited a 35% reduction in mean RMSE and a 63% reduction in mean Bias, along with substantially higher correlation coefficients compared to MCD12Q2 VGP. Even when validated against PhenoCam ground observations, which are inherently more aligned with vegetation greenness-based phenology metrics, GLASS VPP consistently demonstrated superior performance to MCD12Q2 VGP.

A longstanding challenge in LSP research has been the comparability between satellite-derived and in situ phenology records [2]. Traditional LSP products have often struggled to align with ground-based plant phenology records, as exemplified by the poor validation of MCD12Q2 VGP against the PEP725 dataset. In contrast, GLASS VPP exhibits superior alignment with PEP725 records, both in terms of overall validation and spatiotemporal variations. Validation results of GLASS VPP against PEP725 records consistently demonstrate strong comparability, with correlation coefficients approaching 0.9 for SOS/MOG and plant full flowering records. Over 68% of PEP725 sites in Germany exhibit significant correlations greater than 0.7 between the GLASS VPP MOG and plant full flowering records. These findings highlight the substantial potential of GLASS VPP in bridging the gap between satellite-derived and in situ plant phenology observations. These results underscore the broader ecological relevance of VPP metrics. While traditional phenology studies mainly focus on leaf development, recent research suggests that flowering phenology may better represent ecosystems’ reproductive cycles and seasonal carbon allocation [90]. In this regard, the strong consistency between GLASS VPP and flowering dates further highlights the ecological representativeness of GPP-based phenology products.

In addition to improved performance in temperate regions, GLASS GPP also demonstrates advantages in extreme and complex environments. For example, in high-altitude areas such as the Tibetan Plateau, traditional NDVI- and EVI-based methods often suffer from spectral noise caused by snow cover and bare soil, resulting in discontinuous or invalid MCD12Q2 outputs [91,92]. In contrast, the GLASS GPP product, with its enhanced sensitivity to weak photosynthetic signals, significantly increases the valid pixel coverage for phenology extraction and improves the spatial completeness of VPP in such regions [26].

Moreover, certain LSP products like VNP22Q2 face challenges with missing values in areas with high background noise, due to algorithm instability [93,94]. Our hybrid fitting approach, combining the derivative-based method (DVM) and the threshold-based method (TBM), successfully mitigates these issues by balancing sensitivity and robustness. This improves both the accuracy and completeness of phenological event detection [95]. To ensure methodological consistency with flux tower-derived phenology studies, we adopted a double logistic function previously applied by Gonsamo et al. in FluxNet studies, which provides a stable and interpretable basis for VPP extraction. Future improvements may involve tailoring smoothing and extraction techniques to the temporal dynamics of GPP across different ecosystems [96].

Despite these improvements, passive optical remote sensing still faces limitations in capturing phenology in tropical evergreen broadleaf forests (EBFs), where seasonal variations are subtle. MODIS-derived products often perform poorly in such contexts [62,97]. Although we filtered out pixels lacking clear seasonal signals to reduce noise, the remaining EBF phenology curves remained noisy or indistinct, reflecting fundamental sensor limitations rather than data or algorithm flaws. Multi-sensor data fusion strategies could help amplify interannual signals and enhance phenology detection in EBF regions [97]. Future studies could explore whether VPP and VGP exhibit complementarity in characterizing terrestrial vegetation activity and, on this basis, consider their combined use to improve the accuracy of vegetation monitoring. The reliability of GLASS VPP metrics is strongly ecosystem-dependent. Specifically, the dataset performs most robustly in ecosystems with a pronounced seasonality (e.g., temperate and boreal forests, croplands, grasslands), while phenological estimates in arid and tropical evergreen regions are associated with higher uncertainty due to weak or irregular seasonal dynamics. In such regions, VPP-derived SOS/EOS may reflect episodic or noisy signals rather than clear biological transitions and should therefore be interpreted with caution, ideally in combination with complementary datasets (e.g., precipitation records, multi-sensor fusion, or ground observations). These limitations reflect both an ecological reality and a methodological challenge, delineating the scope of applicability of GLASS VPP and providing a direction for future improvement.

5. Conclusions

Based on high-quality GLASS GPP data and a synthesized robust phenology extraction algorithm developed in this study, a global VPP product from 2001 to 2020 was generated, addressing the scarcity of global VPP products. Validation against three independent ground-based phenology datasets demonstrates that the GLASS VPP product significantly outperformed the currently widely used MCD12Q2VGP product. Compared to MCD12Q2 VGP, validating accuracy of GLASS VPP against Fluxnet ground observations exhibited an approximately 35% and 63% reduction in mean RMSE and bias, respectively. Even against the canopy greenness-based PhenoCam phenology ground observation, GLASS VPP also show slightly better performance than the MCD12Q2 VGP product. More notably, GLASS VPP exhibited a high degree of alignment with in situ phenology records, significantly overcoming the limitations faced by traditional VGP products. GLASS VPP is the first global dataset directly characterizing the key phases and duration of vegetation’s photosynthesis and productivity. It provides a valuable foundational data source for analyzing the impact of climate change on terrestrial ecosystems’ carbon uptake.