GPP of a Chinese Savanna Ecosystem during Different Phenological Phases Simulated from Harmonized Landsat and Sentinel-2 Data

Abstract

Savannas are widespread biomes with highly valued ecosystem services. To successfully manage savannas in the future, it is critical to better understand the long-term dynamics of their productivity and phenology. However, accurate large-scale gross primary productivity (GPP) estimation remains challenging because of the high spatial and seasonal variations in savanna GPP. China’s savanna ecosystems constitute only a small part of the world’s savanna ecosystems and are ecologically fragile. However, studies on GPP and phenological changes, while closely related to climate change, remain scarce. Therefore, we simulated savanna ecosystem GPP via a satellite-based vegetation photosynthesis model (VPM) with fine-resolution harmonized Landsat and Sentinel-2 (HLS) imagery and derived savanna phenophases from phenocam images. From 2015 to 2018, we compared the GPP from HLS VPM (GPP_HLS-VPM) simulations and that from Moderate-Resolution Imaging Spectroradiometer (MODIS) VPM simulations (GPP_MODIS-VPM) with GPP estimates from an eddy covariance (EC) flux tower (GPP_EC) in Yuanjiang, China. Moreover, the consistency of the savanna ecosystem GPP was validated for a conventional MODIS product (MOD17A2). This study clearly revealed the potential of the HLS VPM for estimating savanna GPP. Compared with the MODIS VPM, the HLS VPM yielded more accurate GPP estimates with lower root-mean-square errors (RMSEs) and slopes closer to 1:1. Specifically, the annual RMSE values for the HLS VPM were 1.54 (2015), 2.65 (2016), 2.64 (2017), and 1.80 (2018), whereas those for the MODIS VPM were 3.04, 3.10, 2.62, and 2.49, respectively. The HLS VPM slopes were 1.12, 1.80, 1.65, and 1.27, indicating better agreement with the EC data than the MODIS VPM slopes of 2.04, 2.51, 2.14, and 1.54, respectively. Moreover, HLS VPM suitably indicated GPP dynamics during all phenophases, especially during the autumn green-down period. As the first study that simulates GPP involving HLS VPM and compares satellite-based and EC flux observations of the GPP in Chinese savanna ecosystems, our study enables better exploration of the Chinese savanna ecosystem GPP during different phenophases and more effective savanna management and conservation worldwide.

1. Introduction

One-third of terrestrial net primary production is stored in savannas annually, which constitute an important component of the overall land carbon sink [1]. The climate and photosynthesis in savannas are characterized by distinct wet and dry seasons. Climate change is predicted to profoundly affect savanna ecosystems in terms of productivity and light–water use efficiency [2]. As reported in previous studies [3,4], savannas play an important role in the cycle of several greenhouse gases and in energy fluxes (latent and sensible heat fluxes) within the context of climate change. Investigating the distinct responses of carbon and water fluxes to environmental factors in savannas is critical [5]. The gross primary production (GPP) dynamics in Chinese savanna ecosystems and in other regions worldwide have also been studied because of the need to better understand the effects of climate change on regional and global biogeochemical cycles [2,6]. Notably, this approach allows us to estimate the contributions of savannas to regional and global climate controls, including the relationships among climatic factors, plant productivity, phenological feedback, and greenhouse gas exchange within the atmosphere [7]. Compared with areas such as Africa and Australia [8], where savannas are more widely distributed, the savanna area in China is relatively small and sparse. Yunnan Province, Southwest China, has a monsoonal climate suitable for savannas and encompasses vast areas with savanna physiognomy. However, the savanna ecosystem in Yunnan Province also faces fragmentation and is in urgent need of protection [9]. Our research focused on formally recognized savannas in the Yuanjiang dry valleys of Yunnan Province in Southwest China. The Yuanjiang savanna is the most typical savanna ecosystem in China [10] and is the only savanna ecosystem site among the multiple existing carbon flux observatories in China.

Both eddy covariance (EC) and remote sensing data are extremely important for providing reliable information on CO₂ exchange in savannas where the flux tower coverage is limited or nonexistent. In traditional remote sensing-based GPP simulation methods, the GPP is calculated via light use efficiency (LUE) models that are based on vegetation indices (such as the enhanced vegetation index (EVI)) and meteorological data (such as solar radiation and temperature) as inputs. Numerous LUE models have been developed with different combinations of vegetation indices and expressions for environmental stress [11]. In several studies, GPP data derived from LUE models have been compared with GPP EC data obtained from EC flux tower sites [12,13]. However, savanna vegetation comprises both C3 and C4 plants, of which the GPP is more difficult to simulate and study than that in ecosystems of a single plant type (C3 or C4 composition). Among these LUE models, the satellite-based vegetation photosynthesis model (VPM) was the first to consider the fraction of photosynthetically active radiation (PAR) absorbed by vegetation chlorophyll (FPAR_chl) to estimate the GPP [14]. The VPM is frequently used for GPP estimation because of its clear structure and simple parameters. Recently, this model has been further developed and applied to various ecosystems [15], including forests [16], grasslands [17], and savannas [18]. With the use of the VPM, ref. [15] estimated the GPP of different vegetation types globally, including savannas, by setting different photosynthetic use efficiency values for C3 and C4 plants. Geographically, the VPM has been employed to simulate the GPP in Mediterranean-type savannas, savannas in North America, and the cerrado in South America [18,19]. However, the satellite data used in the VPM are typically Moderate-Resolution Imaging Spectroradiometer (MODIS) remote sensing data, whereas VPM simulations with high-resolution data, such as Sentinel-2 and harmonized Landsat and Sentinel-2 (HLS) data, are rare.

High-temporal-resolution products (e.g., Moderate-Resolution Imaging Spectrometer (MODIS) and Medium-Resolution Imaging Spectrometer (MERIS) products) are generally too coarse (from 250 m to a few kilometers) to capture tropical ecosystem heterogeneity [20]. Moreover, owing to their long revisit time (16 days), medium-spatial-resolution products (e.g., Landsat products with a 30 m spatial resolution) potentially lack observations at critical phenological stages [21]. The combination of data from different ongoing satellite missions, such as Landsat-7/8 and Sentinel-2, provides an unprecedented opportunity to estimate the GPP at the regional scale [22]. However, harmonizing L7, L8, and S2 data remains a challenging process that requires several procedures. Whether HLS data can be used to simulate the GPP in savannas is unknown because the required procedures are data-demanding and computationally intensive [23]. The Google Earth Engine (GEE) has significantly streamlined data management and accelerated the analysis process by providing access to planetary-scale archives of remote sensing data, including Landsat-7, Landsat-8, and Sentinel-2 data. The inherent functions and algorithms embedded within the GEE platform involve numerous preprocessing tasks, enabling users to focus on interpreting fundamental algorithms [24]. With the GEE platform, there is a new opportunity for savanna GPP simulation.

Much information on the response of savannas to climate change can be obtained from monitoring savanna phenology [25]. Studying the relationship between phenology and GPP changes can enhance our understanding of the functions and characteristics of diverse vegetation ecosystems worldwide [26]. To better understand seasonal patterns and the response of forest canopies to interannual and long-term changes in climate, various approaches have been adopted to obtain temporal canopy information. Satellite remote sensing techniques have been widely applied in vegetation phenology research to detect the timing of phenological events in greenness-related vegetation indices [27]. However, savannas constitute complex assemblages of tree–shrub–grass layers, each with distinct phenological responses to seasonal climate and environmental variables, and accurate and robust retrievals of savanna phenology information via remote sensing methods remain a challenge [28]. Recently, automatically captured digital camera images, a form of near-surface remote sensing, have been reported as a promising new tool for accurately monitoring vegetation phenology at the ecosystem scale [29]. A digital camera (also referred to as a phenocam) can continuously provide imagery over time and is robust to variations in illumination conditions. Phenocams are rarely obscured by clouds and can collect high-temporal-resolution data with lower operating costs than can remote sensing satellites [30]. Digital repeat photos provide quantitative values of color indices, which allows researchers to detect changing phenological patterns. Moreover, phenocams are widely used at CO₂ flux tower monitoring sites to interpret the seasonal variability in GPP [31]. The color indices derived from photos are correlated with the tower flux-based GPP and can be used to estimate the spatial and temporal distributions of photosynthetic productivity [25].

Research on productivity patterns and drivers in savannas is urgently needed because it is central to understanding the response of savannas to climate change [32]. Further evaluation of VPM simulations with phenological analysis for other savannas is necessary before this model can be applied to estimate the GPP in savanna ecosystems at the continental scale [33]. Recently, it has been recognized that leaf phenology is an important driver of canopy gas exchange, water use efficiency, and photosynthetic capacity changes [34]. Vegetation phenology is an important indicator that is closely connected to changes in forest ecosystem productivity. Globally, phenology-based seasonal gross carbon uptake by terrestrial ecosystems exhibits asymmetric spatiotemporal dynamics. Phenology-based definitions of seasonality might offer a better starting point for studying spatiotemporal patterns of seasonal GPP dynamics than definitions of meteorological seasons do [35]. Hence, considering canopy phenophase information in the GPP estimation process could improve accuracy [36]. Limited research has been conducted in tropical water- or light-limited/driven environments, in contrast to studies of temperature-driven trends in vegetation phenology. Grasslands and trees in savannas exhibit different phenological strategies, and such strategies directly affect the relative contributions of grasses and trees to net primary production [37]. By analyzing data from 3 EC towers in a Mediterranean-type savanna, ref. [38] reported that phenology affects the uncertainty in carbon flux estimates and that spatial heterogeneity significantly contributes to differences in savanna flux estimates. Moreover, ref. [21] identified that the scarcity of valid Landsat data for Brazilian forest vegetation limited phenological monitoring and suggested that the combination of Landsat data with Sentinel-2 data should be enhanced in future studies. Studying the GPP in savannas during different phenophases within the context of global vegetation phenological shifts and increased interannual variability in climatic drivers due to global climate change is particularly important [2]. It has been shown that LUE models perform differently in modeling the GPP at different phenological stages [16,39], but the types of remote sensing-based savanna GPP changes during different phenophases using the VPM remain uncertain, especially when combined with high-resolution remote sensing data.

It is essential to couple technological advances with scientific questions in the study of phenology and savanna GPP [40]. In this study, phenocam images were combined with EC and satellite data to provide insights into the dynamics of phenological processes and their relationships with VPM GPP estimates for a typical Chinese savanna ecosystem. The main objective of this research was to simulate the GPP in this savanna and evaluate the modeling results via in situ data from our flux tower station in Yuanjiang, Yunnan, China. Furthermore, since HLS data have recently been proposed as very promising remote sensing data and since there are no readily available HLS data products that cover the study period, we aimed to harmonize Landsat and Sentinel-2 data by leveraging the advantages of the built-in functions, coherent data structures, and computational power of the GEE. We extracted essential variables as key VPM inputs by exploiting the GEE capabilities via remote sensing HLS data and estimated the GPP in the selected savanna via the VPM. We aimed to answer the following questions: (1) Is the VPM suitable for simulating the GPP in savannas in China? (2) How does the VPM perform using high-resolution HLS data? (3) How does the estimated GPP in Chinese savannas behave during different phenophases?

2. Materials and Methods

2.1. Study Area

Located in Yuanjiang County, Yunnan Province, China, the Yuanjiang Savanna Ecosystem Research Station (YSERS; 23°28′ N, 102°10′ E; elevation: 481 m above sea level (asl)) of the Xishuangbanna Tropical Botanical Garden, Chinese Academy of Sciences, was chosen as the study site (Figure 1). The climate in this region is dry and hot, with a long-term mean annual temperature of 24.0 °C and a mean annual precipitation of 786.6 mm (during the last 36 years), with rainfall during the rainy season (May–October) accounting for approximately 81.0% of the total precipitation. The soil depth is less than 35 cm, and approximately 65% of the soil is gravel. The soil is a type of Ferralic Cambisol according to the soil classification system. The vegetation in the Yuanjiang savanna mainly comprises trees, shrubs, and herbaceous plants. Lannea coromandelica and Polyalthia cerasoides are the main tree species, Campylotropis delavayi is the main shrub species, and Heteropogon contortus is the herbaceous plant species. The Yuanjiang savanna is the most typical and characteristic savanna ecosystem in China [41].

2.2. Workflow Overview

We summarized the general workflow of this research into five main steps (Figure 2): (1) Landsat and Sentinel-2 data harmonization, (2) HLS VPM simulation, (3) model evaluation between the HLS VPM and MODIS VPM, (4) phenocam image analysis and phenophase extraction and (5) model evaluation between the different phenophases. Notably, vegetation indices and the land surface water index (LSWI) were extracted from the HLS time series by synthesizing the HLS data in the GEE. Similarly, the vegetation indices and LSWI were extracted from the MODIS data. These data were entered into the VPM as input variables along with meteorological data obtained from the flux tower to obtain the GPP based on HLS data (hereafter referred to as the GPP_HLS-VPM) and the GPP based on MODIS data (hereafter referred to as the GPP_MODIS-VPM). These two types of GPP data were evaluated using linear regression statistical analysis of the model performance with different input data. The R package xROI was used to analyze the color indices derived from the phenocam images, after which the phenopix package was employed for phenological curve fitting and phenological phase derivation. The GPP_HLS-VPM and GPP_MODIS-VPM were matched to the corresponding phenophase by date, and the VPM was then evaluated on the basis of the considered phenophases. The MODIS GPP product (GPP_MOD17A2) was also compared with the GPP_HLS-VPM, considering the wide range of product applications.

2.3. Flux and Meteorological Data

Starting in May 2013, climatic variables were simultaneously measured via an open-path EC system, which directly provides a high-temporal-resolution dataset of the energy/carbon/H₂O fluxes between the ecosystem and atmosphere [42]. A triaxial sonic anemometer (CSAT3, Campbell Scientific Inc., USA) and a high-frequency open-path CO₂/H₂O infrared gas analyzer (Li-7500, LI-COR Inc., Lincoln, NE, USA) that comprised the EC system were mounted at a height of 13 m above the ground on the flux tower. The flux measurements were complemented by a series of meteorological measurements, such as relative humidity, air temperature (Ta), soil temperature, incoming and outgoing shortwave and longwave radiation amounts, and incoming and reflected PAR amounts.

Both FLUXNET and ChinaFLUX have created a set of standard quality assurance and quality control methodologies to control and ensure the quality of flux data. Standard quality assessment and control (QA/QC) protocols were followed in accordance with their application, as suggested by [42], with the instrument setup and data processing details provided in [43]. In summary, processing included despiking high-frequency data, applying coordinate rotations to align horizontal wind to the mean wind direction, correcting the effects of air density fluctuations via Webb–Pearman–Leuning (WPL) calibration, filtering fluxes with a low friction velocity (0.2 m/s), and identifying and rejecting outliers.

Gap filling and partitioning procedures were applied to the flux data through an online method recommended by FLUXNET [44]. This method is established as the standard protocol by EUROFLUX and is maintained by the Max Planck Institute, accessible at http://www.bgc-jena.mpg.de/5622399/REddyProc, accessed on 1 April 2024. Net ecosystem exchange (NEE) values were partitioned to determine ecosystem respiration (Re) following the methodology outlined by [45], and the GPP was derived as the sum of the NEE and Re. Further information on the data and their processing can be found in other studies [6].

2.4. Phenocam Data and Canopy Phenology Analysis

2.4.1. Phenocam Settings

A digital camera (i.e., the phenocam) was securely mounted on the flux tower and positioned horizontally with a panoramic view of the canopy. Starting in 2006, the Xishuangbanna Tropical Botanical Garden of Chinese Academy of Sciences (XTBG) spent 5 years completing the flux tower site selection and scientific demonstration of the ecological station in the entire Yuanjiang Ecological Reserve after a detailed and thorough site selection and demonstration process. Therefore, the phenocam mounted on the flux tower has a strong regional representation. The camera was enclosed in a waterproof enclosure and directed toward the northwest. A Ricoh Caplio R4 camera was used to capture JPEG images at 1.5 h intervals each day. The images exhibited a resolution of 3072 × 2304 pixels, featuring three 8-bit red–green–blue (RGB) color channels (digital numbers (DNs)) ranging from 0 to 255. The white balance was adjusted in accordance with the camera settings. The camera was operated in automatic aperture and exposure modes.

2.4.2. Image Analysis and Identification of Phenophase Transition Dates

The time series of images were rapidly visually inspected to identify camera shifts in the field of view (FOV), which, notably, remained largely unchanged. Image analysis was performed on a daily timescale. For all the daytime photographs obtained, the time range of 11:30 am to 1:00 pm was first screened and selected. Selecting images during this period minimizes the impact of illumination geometry changes on the images [30]. The images were processed as described in [6] considering a separate set of regions of interest (ROIs) for tree canopies at different distances. The R interactive framework xROI package was employed to select the ROIs from the images [46]. After determining that low-quality or even misleading data were related to camera FOV shifts [46], the FOV shift monitoring module in the xROI package was applied to our four-year (2015–2018) canopy images to rapidly detect potential FOV shifts. In all the image data, there were only minor or major FOV shifts but never a complete FOV change. We regenerated the mask of the ROI in the corresponding image in which an FOV shift was identified.

Canopy greenness time series data were obtained by calculating the green chromatic coordinates (GCCs) of each ROI via Equation (1), as described by [47].

$$\mathrm{GCC}=\frac{{\mathrm{G}}_{\mathrm{DN}}}{{\mathrm{R}}_{\mathrm{DN}}+{\mathrm{G}}_{\mathrm{DN}}+{\mathrm{B}}_{\mathrm{DN}}}$$

(1)

where GCC is a positive ratio that expresses the mean green color fraction of all pixel values in a given ROI, and R_DN, G_DN, and B_DN are the average digital numbers of the three separate color channels (red, green, and blue DN values, respectively) across the ROI. The consistency of the interannual GCC variation was visually assessed to determine whether the inflection point of the change in the index and trends in savanna color change in the photographs were met. Importantly, near-surface camera data were used exclusively for phenological phase division. These data were not incorporated into the GPP modeling process. Instead, HLS and MODIS data served as the primary inputs for GPP estimation via the VPM.

2.4.3. Identifying Phenophase Transition Dates from GCC Values

The R package phenopix was employed for image analysis and phenophase extraction, as documented in [48]. The annual GCC values derived from the digital photographs were processed using phenopix functions to assess phenology patterns. The seasonal GCC trajectories were fitted with a double logistic curve. The estimated dates of important phenophases of canopy development were calculated via the PhenoExtract function. The range of dates associated with notable green-up of the canopy in the ROI was defined as the start of the season (SOS). The peak of the season position (POP) is associated with the maximum canopy green vigor, which corresponds to the peak production. The end of the season (EOS) was defined as autumn, when the canopy becomes less vigorous and green, leaves begin to senesce, and the chlorophyll content decreases. To obtain estimates of important dates in canopy development (aka phenophases), we fit a double logistic curve to the seasonal GCC trajectories, with the equation proposed by [49]. Phenophases are computed via the method proposed by [50]. Specifically, we define the SOS and EOS as the days of the year when the first local maximum and the last local minimum in the rate of change in curvature (k1) occur, respectively. K1 is calculated on the basis of first and second derivatives of the double logistic equation, further defined in [50,51] and graphically illustrated in [48]. The POP is defined as the day of the year when the seasonal maximum of the GCC is reached. Analysis of the GCC trajectories enabled the identification of two phenological stages (start and end of the active growth phase, i.e., the green-up and green-down phases, respectively). Notably, the SOS, POP, and EOS data were retrieved as the day of year (DOY) during the growing season.

Specifically, considering the characteristic color index patterns of savannas, where there was no plateau phase following a color index increase (instead, the savannas directly transitioned from peak values to the decline phase), we segmented the phenological fitted line into two stages: the green-up (spring and summer growing periods, i.e., dates from the SOS to the POP) and green-down (autumn and winter senescence periods, i.e., dates from the POP to the EOS) phases. This simple phenophase division not only facilitates comparisons with the VPM results but also ensures consistency with the seasonal division of savannas commonly observed in other studies [2,25]; namely, dry and wet seasons are generally distinguished.

To address the different spatial resolutions of HLS (10–60 m) and MODIS (500 m) data, we validated the phenocam-derived phenophases against multiple representative points within the satellite imagery by visually comparing the phenology curves for consistency. This approach confirmed the applicability of phenocam data across the study area, ensuring spatial representativeness.

2.5. Harmonization of Landsat and Sentinel-2 in the Google Earth Engine

2.5.1. Cloud and Cloud Shadow Masking

Clouds impede the collection of land surface information from optical images, and the presence of clouds is considered background noise that must be screened out prior to time series analysis. Therefore, we used the pixel quality attribute bands of the Landsat-7, Landsat-8, and Sentinel-2 top of atmosphere (TOA) products for the cloud and cloud shadow masking of each image. Pixel quality attribute bands were provided to indicate the state of each pixel in a given image, such as clouds, shadows or clear pixels.

2.5.2. Radiometric Difference Correction of Multisensor TOA Data

Although Landsat-7, Landsat-8, and Sentinel-2 data exhibit similar spectral band configurations, radiometric differences in the corresponding bands do occur [52]. Therefore, cross-sensor transformation functions were first applied to the TOA data of the three sensors to correct for radiometric differences. The transformation models used in this study were proposed by [52]. These bandwise linear regression models were created on the basis of samples covering the entire contiguous United States (CONUS) with TOA observation pairs one day apart obtained from multispectral instrument (MSI), Enhanced Thematic Mapper Plus (ETM+), and Operational Land Imager (OLI) data. We converted the ETM+ and OLI data into MSI data via the functions provided in

Table 1. Cross-sensor transformation functions for the MSI, ETM+, and OLI TOA data used in this study.

Spectral Bands	From OLI to MSI	From ETM+ to MSI
Blue	MSI = −0.0107 + 1.0946 × OLI	MSI = −0.0139 + 1.1060 × ETM+
Green	MSI = 0.0026 + 1.0043 × OLI	MSI = 0.0041 + 0.9909 × ETM+
Red	MSI = −0.0015 + 1.0524 × OLI	MSI = −0.0024 + 1.0568 × ETM+
Near-infrared (NIR)	MSI = −0.0021 + 1.0283 × OLI	MSI = −0.0140 + 1.1515 × ETM+
Shortwave infrared (SWIR1)	MSI = 0.0065 + 1.0049 × OLI	MSI = 0.0041 + 1.0361 × ETM+
SWIR2	MSI = 0.0046 + 1.0002 × OLI	MSI = 0.0086 + 1.0401 × ETM+

. The data of the three sensors could be integrated after applying the transformation models.

2.5.3. SIAC Atmospheric Correction

The TOA reflectance is usually affected by the atmosphere. To obtain reflectance values that represent the spectral features of the land surface, the atmospheric effect should be corrected to derive the surface reflectance (SR). Despite the availability of Landsat-7, Landsat-8, and Sentinel-2 SR products in the GEE, the atmospheric correction algorithms are diverse, allowing for variations in the retrieved SR results. In this study, we used the uniform sensor-invariant atmospheric correction (SIAC) algorithm [53] to obtain the SRs of the three sensors. The SIAC algorithm solves this problem within the Bayesian probabilistic framework for medium-resolution data. The emulated radiative transfer model with the inputs of the aerosol optical thickness (AOT) and total column of water vapor (TCWV) of the Copernicus Atmosphere Monitoring Service (CAMS) and the bidirectional reflectance distribution function (BRDF) of the MODIS MCD43 product were used to estimate state parameters at the atmospheric correction stage [53]. The SIAC algorithm is currently being developed as a tool module that can be called from within the GEE. In this study, we applied the SIAC algorithm to the radiometrically consistent TOA data of the three sensors to obtain harmonized Landsat-7, Landsat-8, and Sentinel-2 SR data.

The harmonized Landsat and Sentinel-2 time series SR data obtained from the above processes were used to calculate various vegetation indices (VIs), including the normalized difference vegetation index (NDVI), EVI, and LSWI (Equations (2)–(4)), which were subsequently used for GPP estimation.

$$\mathrm{NDVI}={\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{e}\mathrm{d}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}\mathrm{e}\mathrm{d}}}$$

(2)

$$\mathrm{EVI}=2.5\times {\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\mathrm{e}\mathrm{d}}{\mathrm{N}\mathrm{I}\mathrm{R}+6\times \mathrm{R}\mathrm{e}\mathrm{d}-7.5\times \mathrm{B}\mathrm{l}\mathrm{u}\mathrm{e}+1}}$$

(3)

$$\mathrm{LSWI}={\displaystyle \frac{\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}{\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}}}$$

(4)

2.6. MODIS Land Surface Reflectance and GPP Product Data

The MODIS land science team provides a suite of eight-day composite products, including an eight-day SR product (MOD09A1) and an eight-day GPP product (MOD17A2). The MOD09A1 Collection 6 product provides 500 m spatial resolution and 8-day estimates of the SR for seven spectral bands that are used for vegetation and land surface measurements. A detailed description of MOD09A1 is given at https://lpdaac.usgs.gov/products/mod09a1v006/, accessed on 3 April 2024. After data quality control of MOD09A1, the land SR could be obtained, and the NDVI, EVI, and LSWI (Equations (2)–(4)) were calculated in the GEE.

GPP_MOD17A2 was included in this study for comparison with other remote sensing data sources. Over the study period, which was centered on the EC flux site, MOD17A2H products were selected and accessed through the GEE. The GPP of MOD17A2H corresponded to the sum of the GPP values during an 8-day period, so the original MOD17A2 GPP values were transformed from kg C m⁻² 8-day⁻¹ to g C m⁻² day⁻¹ to enable pairwise comparisons with the daily flux tower measurements.

2.7. VPM and Model Parameters

The VPM aims to estimate the daily GPP by using the product of the FPAR_chl of the vegetation canopy and the light use efficiency [54]. The parameters of the VPM can be divided into 2 parts: the vegetation index and meteorological parameters (Figure 2). The air temperature (Ta), EVI, LSWI, and PAR are the input data of the VPM. The meteorological data for Ta and PAR are both based on daily measurements from the Yuanjiang flux tower. This model can be expressed as follows:

GPP = LUE × FPARchl × PAR

(5)

where FPARchl is the fraction of the PAR absorbed by chlorophyll, PAR is the photosynthetically active radiation, the product FPARchl×PAR refers to the absorbed photosynthetically active radiation (APAR), and LUE is the light use efficiency (g C mol⁻¹ photosynthetic photon flux density (PPFD)).

FPARchl can be calculated as a linear function of the EVI with adjusted coefficients of 0.1 and 1.25 according to previous studies [15]:

FPARchl = (EVI − 0.1) × 1.25

(6)

The light use efficiency (ε_g) can be obtained from the maximum LUE (ε₀) scaled by the effects of temperature (T_scalar) and water (W_scalar) (Equation (7)).

ε_g = ε₀ × T_scalar × W_scalar

(7)

T_scalar is a measure of the sensitivity of photosynthesis to the air temperature, which can be calculated for each time interval as follows (Equation (8)):

$${\mathrm{T}}_{\mathrm{scalar}}={\displaystyle \frac{(\mathrm{T}-{\mathrm{T}}_{\mathrm{m}\mathrm{i}\mathrm{n}})(\mathrm{T}-{\mathrm{T}}_{\mathrm{m}\mathrm{a}\mathrm{x}})}{[(\mathrm{T}-{\mathrm{T}}_{\mathrm{m}\mathrm{i}\mathrm{n}})(\mathrm{T}-{\mathrm{T}}_{\mathrm{m}\mathrm{a}\mathrm{x}})]-{(\mathrm{T}-{\mathrm{T}}_{\mathrm{o}\mathrm{p}\mathrm{t}})}^2}}$$

(8)

where T, T_max, T_min, and T_opt are the daytime mean, maximum, minimum, and optimum temperatures for photosynthesis, respectively. The biome-specific values for the latter three parameters can be obtained from a lookup table for a T_max value of 48 °C and T_min and T_opt values of 1 °C and 30 °C, respectively [15].

The sensitivity of photosynthesis to water (W_scalar) can be estimated via the LSWI, and during the growing season, the maximum LSWI was assumed to match the LSWI_max:

$${\mathrm{W}}_{\mathrm{scalar}}={\displaystyle \frac{1+\mathrm{L}\mathrm{S}\mathrm{W}\mathrm{I}}{1+{\mathrm{L}\mathrm{S}\mathrm{W}\mathrm{I}}_{\mathrm{m}\mathrm{a}\mathrm{x}}}}$$

(9)

C3 and C4 plants in savannas exhibit contrasting responses to rising CO₂ concentrations, as well as differing responses to solar radiation and temperature changes [2]. The proportions of the net primary productivity (NPP) of C3 and C4 vegetation in the Yuanjiang savanna were surveyed [55], in which the proportion of C3 vegetation to the total NPP reached 82.57%, whereas the proportion of C4 vegetation was 17.43%; moreover, NPP field biomass surveys were linked to vegetation proportions [56]. We adopted this value as the approximated proportion of C3 and C4 vegetation. We recalculated the maximum LUE value according to the proportion of C3 and C4 vegetation in the savanna. Following the latest VPM GPP products, the LUEmax values of C3 and C4 plants were set to 0.42 and 0.63 g C/mol APAR, respectively [15]. Furthermore, to reduce the model uncertainty [11], we fitted a curve based on two weeks of half-hourly flux data during the 2015–2018 peak savanna growth period to obtain the maximum LUE value, i.e., ε₀, via the GPP–PPFD curve relationship method [57] for model calibration (Equation (10)).

$$\mathrm{GPP}={\displaystyle \frac{{\mathsf{\varepsilon }}_0\times {\mathrm{G}\mathrm{P}\mathrm{P}}_{\mathrm{m}\mathrm{a}\mathrm{x}}\times \mathrm{P}\mathrm{P}\mathrm{F}\mathrm{D}}{{\mathrm{\varepsilon }}_0+{\mathrm{G}\mathrm{P}\mathrm{P}}_{\mathrm{m}\mathrm{a}\mathrm{x}}\times \mathrm{P}\mathrm{P}\mathrm{F}\mathrm{D}}}$$

(10)

where GPP_max is the maximum canopy CO₂ uptake rate (μmol m⁻² s⁻¹) at light saturation.

The maximum photosynthetic efficiency obtained by fitting the GPP–PPFD curve (Figure S1) was 0.44 g C/mol APAR, which is close to the maximum LUE value of 0.4566 g C/mol APAR calculated by using the vegetation proportion. In many recent VPM studies, a lookup table has been employed to calculate ε₀. For comparison convenience, the maximum photosynthetic efficiency in VPM parameter calibration was set to 0.4566 g C/mol APAR.

After the VPM was coded in the R language and environment [58], the HLS- and MODIS-based VIs and the meteorological variables were input into the VPM. Thereafter, the GPP_HLS-VPM and GPP_MODIS-VPM were estimated via the VPM.

2.8. VPM Performance Evaluation

To assess the model validity, the EC-estimated GPP_EC and VPM-modeled GPP (GPP_VPM) data were compared. Since the sampling time frequencies of the HLS and MODIS data differed—4.6 days on average for the HLS data and 8 days for the MODIS data—coupled with the phenocam availability at the Yuanjiang station, there was no need to interpolate and fit the GPP curves to extract phenological metrics (Section 2.4.3). Therefore, we directly compared the VPM-modeled GPP with the EC data of the corresponding dates. A linear regression model was created between the GPP_EC and GPP_VPM values. Three statistics were used to assess the model agreement and bias: the root-mean-square error (RMSE) (Equation (11)), regression slope, and coefficient of determination (R²). The RMSE reflects the model’s ability to estimate the variability in measurements around the mean and can be used to assess the change in the GPP_VPM relative to the GPP_EC. The lowest RMSE threshold is 0, indicating full agreement between the GPP_VPM and GPP_EC values. The mean difference percentage between the GPP_VPM and GPP_EC values is represented by a curve with a zero intercept. The RMSE can be calculated as follows:

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{{\sum }_1^{\mathrm{n}}{({\mathrm{GPP}}_{\mathrm{EC}}-{\mathrm{GPP}}_{\mathrm{VPM}})}^2}{\mathrm{n}}}}$$

(11)

where n is the number of observations.

3. Results

3.1. Temperature and Precipitation and the Relationship between the GPP_EC and Remote Sensing VIs at the Yuanjiang Station

The highest average daytime temperature at the Yuanjiang station from 2015 to 2018 exceeded 35 °C (Figure 3a) and ranged from 5 to 36 °C. This average temperature range is similar to that of African and Mediterranean savannas [18,59]. In contrast to Mediterranean savannas, where rainfall is mainly concentrated in fall and winter, rainfall in the Yuanjiang savanna is mostly concentrated in summer from July to October, with distinct wet and dry seasons (Figure 3b).

The VIs extracted in the GEE from the HLS and MODIS satellite remote sensing data and the GCC values extracted from the near-surface remote sensing data are plotted as time series. In general, all the remote sensing-based VIs exhibited distinct seasonal variations and were consistent with the temporal variability in the GCC values (Figure 4 and Figure 5). The annual EVI amplitudes for the HLS data ranged from 0.1 to 0.6, whereas those for the MODIS data ranged from 0.2 to 0.6. Both the HLS and MODIS data exhibited large annual EVI amplitudes that exceeded the minimum requirement for reliable phenology detection, with a savanna EVI amplitude of 0.02 [28]. Visually, the HLS data showed a more consistent phenological signal from year to year; i.e., the HLS data were closer to the GCC in terms of temporal resolution and amplitude, so a more complete curve of phenological change could be obtained (Figure 4). However, the MODIS-based EVI and LSWI, which are linked to VPM simulations, were not synchronized in terms of curve change. For example, during the 3-year period from 2016 to 2018, the peaks of the MODIS-based EVI and LSWI lagged behind those of the GCC, whereas the peaks were ahead of those of the GCC in 2015 (Figure 5). In contrast, for the HLS data, the EVI and LSWI were better synchronized in terms of curve change and were also more centralized toward the GCC data. Moreover, the HLS-based EVI and LSWI peaks were more synchronized than those of the GCC (Figure 4). This is one of the reasons why HLS data have been used as a seamless fusion of remote sensing data and phenocam photos in other phenology studies [60]. The time series of all the GCC and VI data provide the basis for all the following comparisons and analyses.

To further evaluate the biological significance and performance of VIs in GPP simulation, simple linear regression analyses between the GPP_EC and HLS-based VIs and between the GPP_EC and MODIS-based VIs were performed (Figure 6). For both sets of satellite data, linear regression analysis revealed that the GPP_EC exhibited better linear relationships with both the LSWI and EVI than with the NDVI. The relationships between the HLS-based VIs (NDVI, LSWI, and EVI) and the GPP_EC (Figure 6a–c, R² = 0.2, 0.32, and 0.32, respectively) were comparable to those between the MODIS-based VIs and the GPP_EC (Figure 6d–f, R² = 0.14, 0.17, and 0.22, respectively). This finding also demonstrates that for high-resolution satellite data, the simple linear regression model approach to GPP model prediction does not provide a comparative advantage [61] and that an alternative model structure (e.g., the VPM) is needed for savanna GPP model simulations to reduce the uncertainty [62].

3.2. Comparison of the EC Dynamics and Trends of the HLS and MODIS VPM Results

The daily GPP (GPP_HLS-VPM) was generated using the HLS data and VPM and validated with the EC-measured GPP (GPP_EC) to evaluate its accuracy. Since the earliest date for Sentinel-2 data on the GEE is 23 June 2015 [63], the data points for the GPP_HLS-VPM differ from those for the MODIS because they started after July 2015 (Figure 7b). Both the MODIS and HLS data can be used to track EC interannual changes and intra-annual logistic curve movements relatively well (Figure 7). The highest EC peak in the Yuanjiang savanna during the 4-year period occurred in August 2018, reaching 7.64 g C m⁻² d⁻¹. The GPP overestimations for both the HLS and MODIS data were obvious, with some data points occurring above the EC curve from 2015 to 2018 (Figure 7a,b). The GPP_HLS-VPM data points, especially for the 2016 to 2018 period, were more complete, and many of these points coincided well with the EC data (Figure 7b).

3.3. Evaluation of the HLS and MODIS VPM Results

The VPM was applied to the processed HLS and MODIS data. The GPP_HLS-VPM and GPP_MODIS-VPM were compared with the GPP measured at the Yuanjiang flux tower (Figure 8;

Table 2. Linear regression coefficient, coefficient of determination (R²), and RMSE of the VPM-based estimates of the GPP with the GPP_EC at the Yuanjiang savanna site.

	GPPHLS-VPM			GPPMODIS-VPM
Year	Slope	R2	RMSE	Slope	R2	RMSE
2015	1.12	0.58	1.54	2.04	0.76	3.04
2016	1.8	0.79	2.65	2.51	0.83	3.1
2017	1.65	0.66	2.64	2.14	0.66	2.62
2018	1.27	0.74	1.80	1.54	0.78	2.49
2015–2018	1.46	0.71	2.25	1.91	0.74	2.83

). During the 2015–2018 period, compared with those of the GPP_EC curves, the slopes of the linear regression curves (ideally 1.00) varied between 1.12 and 1.8 for the GPP_HLS-VPM (Figure 8a–d; Table 2). For the MODIS VPM, the GPP_MODIS-VPM was 91% greater than the GPP relative to the GPP_EC (Table 2, slope = 1.91). The HLS overestimation was 45% lower than the MODIS overestimation (Table 2, slope = 1.46). The VPM results for very few of the MODIS data fell below the 1:1 line (Figure S2a), indicating that the GPP_MODIS-VPM values were higher than the EC observations. The VPM results for the HLS data occurred on both sides of the 1:1 line (Figure S2b), indicating the robustness of the GPP_HLS-VPM in matching the EC-based GPP. The use of the HLS-based VIs in the VPM improved the predictability of the GPP, as represented by the reduced RMSE (decreased from 2.83 to 2.25) of the MODIS VPM results compared with that of the GPP_EC (Table 2). In terms of the coefficient of determination (R²), except for 2015, the differences between the GPP_HLS-VPM and GPP_MODIS-VPM were not significant (Figure 8; Table 2), indicating the need to further evaluate the performance of GPP simulations with data of two distinct spatial resolutions on a finer time scale (i.e., the seasonal phenological phase).

3.4. Comparison of the Conventional MODIS Products and GPP_EC

This study focused on the improvement effect of HLS data on the performance of the VPM. Moreover, we included MODIS data products with a coarse spatial resolution (a spatial resolution of 500 m), which differs from that of the VPM algorithm, for comparison to better understand the performance of conventional MODIS products in estimating the GPP in the Yuanjiang savanna.

Compared with those of the HLS and MODIS VPM results, the interannual and seasonal trends in MOD17A2 were very inconsistent with those in the EC data (Figure 9), and MOD17A2 could not suitably track the variations in the EC data (Figure 7). The comparison between the predicted GPP from the MOD17A2 algorithm and the EC GPP revealed an average overestimation of 32% across all years and overestimations of 39%, 57%, 14%, and 26% in 2015, 2016, 2017, and 2018, respectively. However, the GPP from MOD17A2 did not explain the variability in the EC GPP very well (

Table 3. Linear regression coefficient, coefficient of determination (R²), and RMSE of the GPP MOD17A2 with the GPP_EC at the Yuanjiang savanna site.

Year	Slope	R2	RMSE
2015	1.39	0.01	2.82
2016	1.57	0.03	2.81
2017	1.14	0.05	3.2
2018	1.26	0.03	3.25
2015–2018	1.32	0.001	3.03

, R² = 0.01~0.05), and the RMSE values were significantly greater than those for the HLS-based VPM, ranging from 2.81 to 3.25 across the years. This highlights the challenges of using MOD17A2 to simulate GPP accurately in the Yuanjiang savanna.

3.5. Extraction of Phenophases and Changes in the GPP during Different Phenophases via Remote Sensing Simulations

3.5.1. Phenophase Division

By using the xROI package, we could extract GCC values from the screened phenocam photos, and the interannual GCC values showed obvious seasonal variations in the Yuanjiang savanna (Figure 10). Notably, the green GCC peak was reached during the summer season of each year, which agrees with our previous study [6].

On the basis of the GCC curve, we further divided the phenophases of the canopy in the Yuanjiang savanna, in which the methods of fitting a phenological curve and dividing phenological stages were compared with the results of various analyses with the phenopix package (Figure S3). Moreover, we selected Beck’s fitting and Klosterman’s classification methods (Figure S3). The phenophase extraction results varied by year, which was the result of the combined effects of the different climatic and soil environmental conditions and species compositions of the savanna community from year to year. We obtained the green-up (spring–summer) and green-down (autumn–winter) phenological periods from 2015 to 2018 (

Table 4. Division of phenological periods on the basis of the GCC value from near-surface remote sensing phenocam images, 2015–2018. The green-up stage represents the period (DOY) from the beginning of the spring growing season to the green peak, and the green-down stage represents the period (DOY) from the green peak to the end of the autumn growing season. SOS_DOY: Start of the season (day of the year). POP_DOY: Peak of production (day of the year). EOS_DOY: End of the season (day of the year).

Year	SOSDOY (Green-Up)	POPDOY	EOSDOY (Green-Down)
2015	128	176	296
2016	105	186	265
2017	102	169	268
2018	91	158	220

), which also provided the basis for the phenological periods when comparing the remote sensing-based GPP simulation results during different phenophases.

3.5.2. Comparison of the Simulated GPP Values during the Different Phenophases

By using the GCC values, we could achieve fine phenophase division. We can further analyze the performance of the VPM and remote sensing data sources with different spatial resolutions during the different phenophases in the Yuanjiang savanna by comparing the HLS and MODIS VPM simulation results with the EC data at the corresponding phenological stages. Specifically, the HLS data for the Yuanjiang savanna in this study were collected from July 2015 onward. For VPM result comparison during the different phenophases, only data from 2015 that matched the date of the phenophase were used (namely, the DOY should match the 128 to 176 range, Table 4). Therefore, during the 2015 green-up phase, the amount of data was very small (Figure 11a). However, to ensure comparison completeness, these data were retained (Figure 11).

After comparison, we found that the HLS VPM simulation results exhibited a better correlation and a lower simulation error for both the green-up (spring and summer seasons) phenophase (slope = 1.91, R² = 0.75) and the green-down (autumn and winter seasons) phenophase (slope = 1.39, R² = 0.78) (Figure 11 and Figure S4). The RMSE values for the HLS VPM simulation were also lower, with green-up RMSEs ranging from 2.23 to 4.94 and green-down RMSEs ranging from 1.03 to 3.62 across the years (Table S1). The MODIS VPM simulation results also demonstrated favorable performance during both the green-up (slope = 2.61, R² = 0.75) and green-down (slope = 1.68, R² = 0.78) phenophases, but the simulation results in autumn and winter were better than those at the spring and summer phenological stages, respectively (Figure 12 and Figure S4). The corresponding RMSE values for the MODIS VPM were greater than those for the HLS, with green-up RMSEs between 2.83 and 5.78, and green-down RMSEs between 2.89 and 4.74 (Table S1).

In contrast, GPP_MOD17A2 did not effectively track the observed phenological changes. The GPP_MOD17A2 model simulations exhibited high instability during the green-up phase, with the 2015 and 2017 simulations indicating underestimation and the 2016 and 2018 simulations indicating overestimation. The RMSE values for the MOD17A2H green-up phase were significantly greater, ranging from 21.8 to 27.7, indicating greater variability and error (Table S1). Moreover, the R² values were low across the years, with the maximum occurring in 2016 (R² = 0.58) and the minimum occurring in 2017 (R² = 0.34). However, during the green-down phase, the model simulation values were overestimated, with a maximum overestimation of 62% in 2016 and a minimum overestimation of 2% in 2017. The RMSE values during the green-down phase for MOD17A2H were also consistently high, ranging from 22.7 to 25.4 (Table S1).

4. Discussion

4.1. HLS Data Play a Significant Role in Improving Savanna GPP Estimation

Coarse-resolution remote sensing data can lead to uncertainty in vegetation productivity estimates [62]. Observations with a finer spatial resolution should enhance regional GPP estimations by more effectively capturing variations in heterogeneous landscapes [64]. While the MODIS VPM and MODIS GPP products provide valuable GPP estimates that are widely available in space and time, GPP estimation in the face of complex and heterogeneous land cover types, such as river valley savannas, benefits from the use of finer-spatial-resolution imagery. Effectively combining observation data from in situ stations with remote sensing satellite data for more accurate GPP estimation is a research goal that has been addressed in many studies [23,65]. In our study, we validated the performance of the use of HLS data in modeling the GPP in a Chinese savanna through the harmonization of satellite data and long-term observation data from our flux station.

We found GPP overestimation in the Yuanjiang savanna both by the VPM using HLS data and by the VPM using MODIS data. The HLS VPM results indicated lower overestimation than the MODIS VPM results did, with an overestimation of 46% from 2015 to 2018 (Table 2). In contrast, the overestimation of the MODIS VPM results reached 91%. Our MODIS VPM results greatly differed from those in Mediterranean savannas [18], and the GPP_MODIS-VPM at the two Mediterranean-type savanna sites was very close to the GPP_EC during the study period (slope: 0.83–1.15) [18]. This difference may be due to multiple factors influencing savanna types, VPM simulations, and EC measurements. The employed VPM differed from that used in [18]. The newer version of the VPM was established by adjusting the EVI coefficients and removing the factor (P_scalar) accounting for the effect of leaf longevity from the earlier model [15], and the new EVI coefficients caused an increase in FPAR_chl estimates for the same EVI values. In addition, and more importantly, the Yuanjiang savanna exhibits low annual productivity [66]. The mean multiyear EC values of the two Mediterranean savannas, namely, US-Ton and ES-LMa, were 1035.58 and 1033.44 g C m⁻², respectively, whereas the mean multiyear EC value at the Yuanjiang station was 596 g C m⁻². Moreover, our previous studies at the Yuanjiang station revealed that this savanna is highly spatially heterogeneous and that very small precipitation variations could cause changes in the structure and composition of plant communities, influence the structure and composition of tropical savannas, and affect the function and stability of the Yuanjiang savanna ecosystem [41]. With respect to VPM, high-resolution remote sensing data inputs could improve estimations for ecosystems with high spatial heterogeneity [62]. With the integration of HLS data from two high-spatial-resolution satellites, namely, Landsat (with a resolution of 30 m) and Sentinel-2 (with a resolution of 10–60 m), the VPM significantly improved the MODIS VPM results. Notably, overestimation was reduced from 91% to 46% (Table 2). The HLS VPM results also indicated a significant improvement in the GPP estimation accuracy for this savanna, which is highly spatially heterogeneous. Seasonal GPP estimation for savannas is likely a function of the data availability and/or quality [19]. Thus, with the release of more high-resolution data in this decade [67], high-spatiotemporal-resolution satellite observations are likely to constitute a continued research direction for GPP estimation and analysis in the future [65].

4.2. Relationship between Phenophases and GPP Remote Sensing Modeling in Savanna Ecosystems

In contrast to broadleaf evergreen forests at the same latitude, savannas are climatically driven ecosystems because of the presence of annual herbs and a significant proportion of deciduous species [59]. Accurate phenological stage division is essential for measuring the dynamics of savanna ecosystems, especially the exchange of carbon and water between the canopy and the atmosphere [35]. In this study, remote sensing satellite data were used to simulate the savanna GPP; however, deriving savanna phenophases via remote sensing satellite data would increase the systematic uncertainty [11]. Therefore, we used more accurate phenocam-derived GCC values to extract phenophases, while recognizing that GPP changes in ecosystems with different phenophases also represent a popular research topic [35]. Numerous studies [6,30] have demonstrated that the GCC can be used to monitor canopy development and identify phenophases, with notable correlations with remote sensing VI time series (Figure 4 and Figure 5). The phenocam-extracted GCC color index can clearly reflect the phenological signals of Yuanjiang savanna vegetation color change (Figure 10). However, one of the important findings of this study is that the VPM performs differently during the different phenophases at the growing and senescence stages of savanna canopy greening. For the VPM performance comparison, the differences between the different remote sensing data sources were significant in terms of the green-up period metrics but not in terms of the senescence period metrics (Figure 12 and Figure S4). For example, during the green-down phenophases, both the HLS VPM (Figure 11e–h) and the MODIS VPM (Figure 12e–h) performed well. The HLS VPM overestimated the GPP in spring and summer more than it did in autumn and winter (Figure 11 and Figure S4). However, the MODIS VPM significantly overestimated the GPP throughout the growing season (Figure 12 and Figure S4). We suggest that the reason for this model performance difference is, first, the different natures of savanna phenology during the two phenophases. Notably, ref. [33] reported that the interannual variability in savanna vegetation phenology during the senescence phase was low in southern Africa, whereas the complexity of savanna cover variability and phenology was much greater during the green-up phase. This complexity of the vegetation phenology and cover change during the spring and summer periods places greater demands on the accuracy of the input data used in the VPM. Notably, the low complexity during the GPP senescence period can be effectively simulated by the HLS and MODIS VPMs, whereas the high spatial resolution of the HLS VPM is superior to that of the MODIS VPM for the more variable green-up growing period [37]. Improving the spatial resolution of vegetation indices can effectively reduce the seasonal error of VPM simulations. Notably, before their use for GPP simulation purposes, other studies have shown that HLS data can be suitably integrated with phenocam-derived GCC data for extracting phenological periods, curve fitting, and delineation [60]. This also shows that HLS data are more suitable than MODIS data when combined with near-surface remote sensing cameras for the study of phenology indicators and GPP (Figure 4).

One of our main concerns was the ability of the HLS and MODIS data to drive the GPP models during the different savanna phenophases but not the absolute accuracy of the prediction model. Our study is consistent with [33] in that the VPM-based GPP estimates (GPP_VPM) can track the seasonal dynamics of the EC-based GPP in semiarid and semihumid savannas in southern Africa. Both the HLS and MODIS VPMs can track the seasonal dynamics of the EC-based GPP during the different phenophases in the Yuanjiang savanna (Figure 6 and Figure S4). Moreover, the structural composition of savanna species in Yuanjiang is similar to that in Africa [59]. HLS data can capture rapid phenological changes during the growing season (Figure 4 and Figure 12), thus reducing errors in annual GPP estimates [65]. Evaluating VPM in different global savanna ecosystems could help increase the understanding of vegetation phenology and model performance in savanna ecosystems [19].

4.3. Advantages of the VPM over Conventional Remote Sensing Products in Savanna GPP Studies

The HLS VPM GPP estimates were more consistent with the ground flux observation-based GPP values than were the MODIS VPM GPP estimates and MOD17 GPP product. MOD17A2, as a relatively mature and stable conventional GPP product, still plays an important role in GPP studies involving large- or global-scale data, and the accuracy of GPP estimation in savannas was verified in this study. Moreover, we considered that this might be related to the underestimation of MOD17A2 offset by the low annual GPP value of the Yuanjiang savanna. This occurred because MOD17A2 has been considered to underestimate the GPP in other savanna types [18,33]. However, MOD17A2 is unsuitable for tracking seasonal GPP dynamics or the relationship between GPP and phenology. The performance of GPP_MOD17A2 during the different phenophases greatly fluctuated, and the coefficients of determination between the simulated values and the EC values were very low (Figure 13). Our findings agree with those of [68]; notably, it is difficult for MODIS satellite GPP products to reflect the seasonal dynamics of savannas. Instead, the VPM plays a major role in small- and medium-scale studies, especially in combined studies of remote sensing satellite sensors and flux stations, and increasing the accuracy of HLS data can improve such small- and medium-scale studies.

Both the VPM and MOD17 products are models based on the LUE concept [11]. However, they exhibit two main differences. First, the LSWI, which reflects the vegetation canopy water content, was used as a water stress factor in the VPM. In contrast, in the MOD17 algorithm, the atmospheric moisture indicator vapor pressure deficit (VPD) is considered [69]. Under drought conditions, the VPM performs better in capturing the effect of water stress on the GPP, which can be attributed to the higher sensitivity of the LSWI (Figure 6). In contrast, the VPD cannot capture the effect of water stress on the GPP well under severe drought conditions because it does not incorporate soil water deficits into canopy gas exchange [69]. Second, the VPM considers the EVI (chlorophyll or foliar greenness), whereas the MOD17A2 algorithm considers the NDVI to estimate the FPAR. Previous studies have shown that the relationship between the EVI and GPP is better than that between the NDVI and GPP (Figure 6) in various ecosystems [33,70]. Compared with VPD-based water stress, LSWI-based water stress better captures the effect of soil moisture on GPP. The observed GPP representation differences may also be associated with the input data of the MOD17A2H product for computing the GPP. The use of meteorological data from the Global Modeling and Assimilation Office (GMAO)/National Aeronautics and Space Administration (NASA) set by the MOD17A2H product may not adequately represent the dynamics of soil moisture in savannas [12]. We found that the HLS VPM is more closely related to the phenological changes in the savanna than the MOD17A2 product is (Figure 11 and Figure 13), which can also indicate the potential of high-resolution satellite data such as HLS data [23] in the study of vegetation and carbon phenologies in savannas.

The value of the maximum photosynthetic efficiency is important for applying the savanna LUE model. The most significant limitation of the MODIS GPP algorithm is its improper characterization of the LUE. It relies on lookup tables to determine the maximum LUE for a given vegetation type, after which this value is adjusted downward on the basis of environmental stress factors via interpolated meteorological data. In other savanna ecosystems in Africa and North America, the MOD17A2 product also exhibited poor performance [18,33]. The LUEmax value for savannas in MOD17A2 is low and fixed. Moreover, the meteorological parameters of its algorithm cannot be flexibly adjusted in combination with the observation data of the local flux tower. The data retrieved during in situ field campaigns could be used to determine the maximum LUE of C3 and C4 plants [55], thus reducing the uncertainties associated with GPP VPM simulations. This is an advantage of the VPM over MOD17A2 because it is more flexible and accurate in calibrating model parameters. Furthermore, the use of a uniform LUEmax, such as that of the MOD17A2 product for GPP estimation, causes high uncertainty, especially for GPP estimation under drought conditions [71]. However, there was no significant drought in the Yuanjiang savanna between 2015 and 2018 (Figure 3), and the performance of MOD17A2 in GPP assessment in the Yuanjiang savanna must still be analyzed in combination with extreme drought events in the future.

4.4. Uncertainties, Limitations, and Implications of GPP Simulation

The simulation of GPP in savanna ecosystems is fraught with uncertainties arising from multiple factors, including model structure, input data, and the inherent variability of the ecosystems being studied [62]. One significant source of uncertainty in our study stems from the LUE models used to estimate GPP. Different representations of factors such as FPAR and environmental stresses (e.g., temperature, water availability) in LUE models can lead to varying GPP estimates. Notably, the LSWI derived from HLS data may provide more accurate representations of water stress conditions than the MODIS-based LSWI does because of the finer spatial resolution of HLS imagery [23]. Previous studies have highlighted discrepancies in the performance of different model structures across various ecosystems, which adds to the uncertainty in GPP estimates [72]. The VPM model structure has been shown to reduce simulation errors in savanna ecosystems more effectively than the MOD17A2H model [19,71].

Our analysis revealed that the effects of input variables on GPP can differ significantly depending on the type of vegetation, meteorological conditions, and hydrological status [62]. For example, in the Yuanjiang savanna, the LUEmax is constant across the VPM and MOD17 models, which could lead to overestimations of GPP. In contrast, the VPM’s LUEmax is calibrated on the basis of flux values during peak periods. However, our previous research revealed that the LUE in savannas varies throughout the growing season, increases in summer and autumn, and decreases in winter and spring [73]. This suggests that LUEmax might also change seasonally. This overestimation is particularly notable in the MOD17A2H product, where the LUEmax is derived from a lookup table on the basis of vegetation type, which fails to account for the seasonal variations in LUE observed in savanna ecosystems. These oversimplifications highlight the need for more dynamic and locally calibrated LUE parameters in future models [39].

The spatial resolution of remote sensing data plays a crucial role in the accuracy of GPP simulations. Studies have shown that the GPP derived from fine-spatial-resolution remote sensing data can reduce GPP simulation uncertainties by 20% [74]. However, the use of traditional vegetation indices (Figure 6) in LUE models may still contribute to GPP overestimations. Emerging technologies such as solar-induced chlorophyll fluorescence (SIF) offer promising alternatives by providing more direct measures of photosynthetic activity, potentially reducing the reliance on conventional vegetation indices [75]. SIF also offers a new perspective for future remote sensing studies of GPP simulations in savanna ecosystems.

Some studies suggest that GPP models may overestimate productivity in low-productivity savanna ecosystems such as the Yuanjiang savanna, likely due to the unique ecological characteristics of these regions where more sterile conditions occur frequently (Figure 3) [76]. Another source of uncertainty is ground validation data. Despite adhering to rigorous flux measurement protocols [66], errors in eddy covariance measurements due to factors such as footprint heterogeneity remain [77]. Expanding the network of flux towers across similar savanna ecosystems and integrating machine learning approaches with repeated model simulations could be a potential solution for addressing these uncertainties in the future.

Since the Yuanjiang Ecological Station is currently the only flux research site in China that represents a savanna ecosystem and given the limited distribution of savannas in China, research on GPP remote sensing simulations in these ecosystems is scarce. As a result, it is challenging to obtain comparative GPP remote sensing validations from other savanna regions in China. Therefore, the presented conclusion concerning the accuracy of HLS GPP simulation may not be transferable to other savanna regions in China or the Indochina Peninsula. However, combining studies of China’s savannas with those in Southeast Asia [3] and using high-resolution satellite remote sensing and near-surface remote sensing to investigate phenology and GPP [6] could open new avenues for future research.

5. Conclusions

A few years after the introduction of the HLS concept, we conducted HLS remote sensing GPP model simulations in the most typical savanna in China. This is the first case study of VPM simulations using HLS remote sensing data and considering different phenological phases for savanna GPP estimation on the basis of phenocam images. We used the GEE cloud computing platform to synthesize HLS data for a typical savanna in China and extracted the meteorological and vegetation parameters needed for VPM simulations. VPM simulations were performed by combining important vegetation parameters with ecosystem flux and meteorological data. The simulation results were compared with MODIS satellite data, which are commonly used in other studies. The consistency of the savanna ecosystem GPP was also verified for the global product MOD17A2. In this study, we found that the HLS VPM provided a notable advantage over the MODIS VPM in estimating the GPP in the savanna, and the HLS VPM attained a satisfactory performance during all phenological phases in this savanna ecosystem. Moreover, this study confirmed the usefulness of the MOD17A2 product for annual large-scale savanna GPP estimation, but this product could not track seasonal savanna GPP dynamics and was not suitable for use in research on the relationship between the savanna GPP and phenology. Our study is the first comparative study of satellite remote sensing data and EC flux observations of the GPP in savanna ecosystems in China, which is important for determining the changes in the GPP in savanna ecosystems during different phenological phases against the background of climate change and for managing and protecting savanna ecosystems.