Next Article in Journal
SAR-ESAE: Echo Signal-Guided Adversarial Example Generation Method for Synthetic Aperture Radar Target Detection
Previous Article in Journal
Spatiotemporal Dynamics of Annual Precipitation and Future Projections of China’s 400 mm Isohyet
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Remote Sensing
Volume 17
Issue 17
10.3390/rs17173079
remotesensing-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Geostationary Satellite-Derived Diurnal Cycles of Photosynthesis and Their Drivers in a Subtropical Forest

by
Jiang Xu
†,
Xi Dai
†,
Zhibin Liu
,
Chenyang He
,
Enze Song
and
Kun Huang
*
Zhejiang Tiantong Forest Ecosystem National Observation and Research Station, Research Center for Global Change and Ecological Forecasting, Institute of Eco-Chongming (IEC), School of Ecological and Environmental Sciences, East China Normal University, Shanghai 200241, China
*
Author to whom correspondence should be addressed.
†
These authors contributed equally to this work.
Remote Sens. 2025, 17(17), 3079; https://doi.org/10.3390/rs17173079
Submission received: 18 July 2025 / Revised: 28 August 2025 / Accepted: 2 September 2025 / Published: 4 September 2025
Browse Figures
Versions Notes

Abstract

Tropical and subtropical forests account for approximately one-third of global terrestrial gross primary productivity (GPP), and the diurnal patterns of GPP strongly regulate the land–atmosphere CO2 interactions and feedback to the climate. Combined with ground eddy-covariance (EC) flux towers, geostationary satellites offer significant advantages for continuously monitoring these diurnal variations in the “breathing of biosphere”. Here we utilized half-hourly optical signals from the Himawari-8 Advanced Himawari Imager (H8/AHI) geostationary satellite and tower-based EC flux data to investigate the diurnal variations in subtropical forest GPP and its drivers. Results showed that three machine learning models well estimated the diurnal patterns of subtropical forest GPP, with the determination coefficient (R2) ranging from 0.71 to 0.76. Photosynthetically active radiation (PAR) is the primary driver of the diurnal cycle of GPP, modulated by temperature, soil water content, and vapor pressure deficit. Moreover, the effect magnitude of PAR on GPP varies across three timescales. This study provides robust technical support for diurnal forest GPP estimations and the possibility for large-scale estimations of diurnal GPP over tropics in the future.

1. Introduction

Forests critically influence climate through energy, water, and carbon dioxide (CO2) exchanges while storing roughly 77% of global terrestrial carbon [1]. Terrestrial ecosystems are capable of absorbing and sequestering approximately one-third of global carbon emissions, with forest ecosystems accounting for about 90% of this contribution [2]. Subtropical evergreen broadleaf forests exhibit a higher carbon sequestration capacity compared with other regions at similar latitudes, accounting for approximately 8% of the total carbon uptake by global forest ecosystems. They play a vital role in both regional and global terrestrial carbon cycles [3]. Gross primary productivity (GPP) is a key indicator of the carbon uptake process; therefore, accurately estimating GPP is essential for studying carbon flux dynamics in subtropical forests. However, the diurnal variations in carbon uptake and the impact of key meteorological factors on GPP in the tropical–subtropical forests remain insufficiently and ambiguously studied [4,5]. Although recent studies have advanced sub-daily GPP estimation at regional to global scales [6,7,8], research on the diurnal dynamics of carbon fluxes in specific ecosystems remains limited. In particular, subtropical evergreen broadleaf forests in East Asia are still underrepresented in existing sub-daily GPP analyses. Characterizing GPP at sub-daily resolution is crucial for understanding vegetation responses to environmental disturbances [9,10,11], and is especially important for capturing transient phenomena such as mid-day depression [12].
The ability to continuously and accurately monitor carbon flux variations is critical for understanding the diurnal patterns of forest carbon sinks [13]. In situ measurements at the canopy and ecosystem scales, such as those obtained using portable instruments [14] and eddy-covariance (EC) techniques [15], have been extensively employed to examine diurnal variations in plant functions and ecosystem processes [16]. Specifically, EC sites provide continuous, sub-hourly measurements of ecosystem carbon and water exchanges [17], meteorological variables, and soil moisture. However, these in situ measurements are spatially sparse [18], and the results derived may not be applicable to other locations [19]. In contrast, satellites can offer direct observations of vegetation at regional or global scales [20], measuring plant status, canopy coverage, productivity, or other environmental variables [13]. Traditional polar-orbiting satellites, such as Landsat and the Moderate Resolution Imaging Spectroradiometer (MODIS) [21], have limited temporal resolution with observation intervals ranging from daily to multi-day [22]. This limitation often hinders their ability to capture rapid diurnal variations [23]. Compared with traditional observation satellites, emerging satellites can provide continuous observations throughout the day and night [24]. The Himawari-8 Advanced Himawari Imager (H8/AHI) geostationary satellite, positioned at 140° E, continuously observes East Asia and Southeast Asia. The wide instantaneous coverage of the Asia–Pacific region facilitates cross-regional comparisons of ecological and meteorological processes, with broad applications in areas including wildfire detection and environmental change monitoring [25,26].
EC sites provide direct, high-frequency measurements of ecosystem CO2 fluxes, offering reliable ground truth for calibrating and validating remote sensing-based GPP estimates. By integrating EC data with satellite observations, studies have bridged the gap between point measurements and spatially continuous products [27]. While the extent to which machine learning models can enhance estimative skills is uncertain, they are expected to integrate EC and remote sensing observations [28], potentially further improving estimates of terrestrial carbon fluxes [29]. This study employs H8/AHI geostationary satellite data together with carbon flux and meteorological data from the Tiantong forest flux site to assess the diurnal cycle of GPP in a subtropical forest. The objectives of this study are to (1) assess the feasibility of using geostationary satellites to estimate diurnal photosynthesis of subtropical forests; (2) evaluate the performance of machine learning models in characterizing GPP diurnal dynamics; and (3) identify the key drivers of subtropical forest photosynthesis.

2. Materials and Methods

2.1. Site Information

This study was conducted at the Tiantong Forest Ecosystem Flux Observation Site (Figure 1a) (29°48′45″N, 121°47′11″E) within the Tiantong National Forest Park, Ningbo, Zhejiang Province. The park covers a total area of approximately 349 ha and is characterized primarily by hilly landforms (Figure 1b), with an average elevation of about 300 m and a maximum elevation of 653.3 m [30]. The site is part of the China Flux Observation and Research Network (ChinaFLUX) and is located in the evergreen broadleaf forest ecological zone of eastern subtropical China. The evergreen broadleaf forest ecosystem was dominated by Castanopsis fargesii, Schima superba, Castanopsis carlesii and Cyclobalanopsis sessilifolia [31], which reflected the typical characteristics of low-elevation forests in the subtropical region of eastern China. The annual mean temperature was 16.2 °C, with a maximum of 28.1 °C in July and a minimum of 4.2 °C in January. The annual average precipitation was 1551 mm, primarily concentrated in the period of May to August. The annual evaporation is approximately 1320 mm, and the relative humidity can reach up to 85%. The total annual sunshine duration is about 2010 h [32]. The soil belonged to Mountain yellow red soil [33], with 6.8% gravel, 55.5% silt and 37.7% clay. The soil pH ranged from 4.4 to 5.1 [34].

2.2. Eddy-Covariance Measurements

The net ecosystem exchange (NEE) data were continuously measured with an open-path eddy-covariance system. The EC sensors primarily included a 3D ultrasonic anemometer (CSAT-3, Campbell Scientific, Logan, UT, USA) and an open-path infrared CO2/H2O gas analyzer (Li-7500; LI-COR, Lincoln, NE, USA) (Figure 1d), installed at a height of 40 m on the flux tower (Figure 1c). The carbon flux footprint refers to the spatial source distribution of carbon flux signals measured by the eddy-covariance observation system, which is controlled by factors such as wind speed, wind direction, atmospheric stability, and surface roughness. It determines the ground contribution area that the flux tower can monitor in different directions [35]. The carbon flux footprint of the TTS flux tower can extend up to 1800 m from the tower, covering an area of approximately 10 km2. Notably, 50% of the core carbon flux contribution is concentrated within 200 m of the tower (Figure 1e), and the spatial distribution of carbon flux exhibits a pattern of concentrated contributions in the core area and sparser contributions in the periphery [36]. These instruments continuously monitored CO2, water vapor, and energy fluxes at a sampling frequency of 10 Hz. The flux tower was also equipped with probes to measure conventional meteorological parameters. Two sets of air temperature (Tair) and relative humidity (RH) probes (HMP45C, Campbell Scientific) were installed on the flux tower at heights of 40 m and 10 m, respectively. Precipitation was recorded using a tipping bucket rain gauge (TR-525M; Texas Electronics, Inc, Dallas, TX, USA). PAR was measured using a PAR quantum sensor (LI-190R; LI-COR).
The original 10 Hz flux data from 2021 to 2024 were processed using the EddyPro (7.0.6) software, and half-hourly fluxes were calculated. Quality-control protocols included 3D coordinate rotation [37], time lag correction [38], and Webb–Pearman–Leuning (WPL) correction [39]. Half-hourly fluxes were filtered to exclude false values caused by sensor malfunctions and disturbances such as rain. Nighttime anomalies were filtered using a friction velocity (u*) threshold. Evapotranspiration (ET) is directly obtained by eddy covariance through the covariance of instantaneous vertical wind speed and water vapor concentration fluctuations; the output files from the EC system include ET data calculated in this way. To avoid relying on potentially problematic nighttime data [40], we employed a daytime flux partitioning method using the R package REddyProc (1.3.3) to separate NEE into GPP and ecosystem respiration (RECO) (equations in Supporting Materials). Data quality control was performed using the eddy-covariance data quality-control flagging system. The NEE data from TTS during 2021–2024 were partitioned into corresponding annual GPP and RECO, which were then combined with H8/AHI data over the same period for further analysis.

2.3. H8/AHI Geostationary Satellite and Forcing Data

The H8 satellite, launched by the Japan Meteorological Agency in October 2014, is a geostationary meteorological satellite equipped with the Advanced Himawari Imager [41]. The AHI sensor comprises 16 channels covering visible, near infrared, and short-wave infrared bands. The full-disk coverage extends approximately from 80°E to 160°W in longitude and from 60°S to 60°N in latitude [25]. The H8/AHI acquires imagery every 10 min across 16 spectral bands spanning visible to thermal infrared wavelengths [26]. In this study, we utilized the H8/AHI Level 1 (L1) radiance data with a spatial resolution of ~5 km for the period 06:30–18:00 across four years (2021–2024). These products, which also include solar–sensor geometry information (Table 1), are well-suited for capturing the spatial scale of eddy-covariance flux tower footprints, enabling meaningful comparisons between satellite-derived estimates and ground-based observations. Moreover, the 5 km spatial resolution is consistent with that commonly used in global carbon cycle models, thereby facilitating the integration of our results into broader ecological studies [42].
During the study period (2021–2024), the satellite data showed that variability in cloud cover was most pronounced during the summer months and least during the winter months (Figure S1). The cloud cover percentages peak in June and July, correlating with the rainy season and monsoon in the subtropical regions, during which cloud cover is most dense. Conversely, the region experiences clearer and drier weather during the winter months, as indicated by the increased cloud-free data in these periods. Cloud cover is a common and unavoidable phenomenon that can impact the usability of optical remote sensing imagery [43]. To ensure data quality, we used the official Level 2 (L2) cloud mask products provided by the H8/AHI satellite. These L2 cloud products are generated using a combination of spectral tests and threshold algorithms, which classify each pixel as either cloudy, cloud-shadowed, or cloud-free. Specifically, the algorithm considers brightness temperature differences, reflectance in multiple spectral bands, and spatial context to accurately identify cloud and shadow pixels. For our analysis, we selected only the cloud-free pixels as indicated by the L2 product, matched with the corresponding L1 level product, effectively excluding all cloudy and shadowed areas., thereby ensuring the reliability of the remote sensing data input to the model [44,45].
We utilized H8/AHI products with consistent spatial resolution and temporal frequency to calculate vegetation photosynthesis-related proxies, including near-infrared reflectance of vegetation (NIRvP) [46], near-infrared reflectance of vegetation (NIRv) [46], the normalized difference vegetation index (NDVI) [47], and the enhanced vegetation index (EVI) [48]. For the environmental variables, we used the H8/AHI bands to calculate land surface temperature (LST) [49] and obtained photosynthetically active radiation (PAR) directly from the H8/AHI products. In addition, to further explore the optimal combination of model input parameters, we subsequently trained and tested the model using Bidirectional Reflectance Factor (BRF) features, Cloud and Moisture Imagery (CMI) features, and a combination of BRF and CMI features. Specifically, the BRF includes Himawari-8 data from bands 1 to 6, CMI includes bands 7 to 16, and ALL includes all bands from 1 to 16 (Table 1).
L S T = C + A 1 B T i + A 2 B T i − B T j + A 3 ε + D B T i − B T j sec θ − 1
In this approach [49], B T i and B T j denote the brightness temperatures from two thermal infrared bands provided by the H8/AHI L1 data. In this study, B T i and B T j correspond to the brightness temperatures of bands 14 and band 15, respectively. The coefficients C , A 1 , A 2 , A 3 , and D are obtained from the literature by [50]. These L S T coefficients were originally developed for geostationary operational environmental satellite-advanced baseline imager (GOES-ABI); given that the H8/AHI exhibits essentially identical spectral and radiometric characteristics in bands 14 and 15 [26,51], they can be applied as a reference for the H8 data. The emissivity ( ε ) is dynamically calculated based on NDVI: for NDVI > 0.5, ε is set to 0.98; for 0.2 ≤ NDVI ≤ 0.5, ε is set to 0.92 + N D V I − 0.2 / 5 ; and for NDVI < 0.2, ε is set to 0.92. The satellite zenith angle ( θ ) is provided by the H8 L1 data ( S A Z ).

2.4. Machine Learning

In the estimation of ecosystem carbon fluxes, machine learning models are widely applied to handle complex nonlinear relationships and high-dimensional data [52]. In this study, we employed four machine learning models—Random Forest (RF), Gradient Boosting Regression (GBR), Support Vector Regression (SVR), and Multi-Layer Perceptron (MLP)—to estimate subtropical forest GPP using satellite data products as input parameters. We systematically evaluated the performance of these four models and subsequently selected the best-performing one to analyze diurnal cycle dynamics of photosynthesis in subtropical forests.
The vegetation indices and environmental variables served as input parameters for our machine learning algorithms. Our total sample size consisted of 6868 half-hourly daytime observations, after excluding measurements affected by cloud cover and failing quality-control criteria (Figure S1). Finally, the model outputs were compared with EC-derived GPP obtained from ground-based flux tower measurements to evaluate model performance.
To compare and evaluate the simulation performance of the four models, we utilized two key statistical indicators: the coefficient of determination ( R 2 ) and the root mean square error ( R M S E ). We employed a grid search algorithm and Table S1 presents the parameter search space defined for each machine learning model. In Equation (2), T S S represents the total sum of squares, and R S S denotes the residual sum of squares. In Equation (3), the index i refers to the individual data points, and N is the number of non-missing data points. Here, x i denotes the actual observed values in the time series, and x ^ represents the estimated time series.
R 2 = 1 − R S S T S S
R M S E = ∑ i = 1 N x i − x ^ N
We randomly partitioned all 6868 half-hour daytime samples into 90% training and 10% validation sets [53,54,55] and repeated this split across 50 distinct random seeds (seeds = 1~50) to assess model robustness and mitigate the effects of any single partition [56]. This approach to using machine learning models allows for improved estimations of ecosystem carbon fluxes, providing a potential for near real-time carbon cycle monitoring, especially when geostationary satellites’ high temporal resolution data are integrated with these machine learning techniques [57].

2.5. Pearson Correlation and SHAP Values Analysis

To evaluate the response of carbon flux to meteorological factors, we conducted a Pearson correlation analysis using nearly four years of flux data and constructed a network graph using Mantel correlation and Bray distance algorithms [5]. And we employed SVR as the predictive model to characterize the nonlinear relationships between environmental factors and carbon fluxes. Model stability was assessed using cross-validation. To interpret the model, we applied Kernel SHAP, which estimates Shapley values by approximating the marginal contributions of each feature through sampling. This approach allows quantification of the contribution of individual predictors in each sample, thereby linking feature attributions directly to the SVR predictions. The SHAP value is computed as follows:
ϕ i f = ∑ S ⊆ N \ i S ! ⋅ N − S − 1 ! N ! ⋅ f S ∪ i − f S
ϕ n o r m = ϕ i ∑ ϕ j
In Equation (4), N denotes the set of all meteorological features, and S represents a subset of features corresponding to GPP, NEE, or RECO. Here, ϕ i f is the SHAP value for feature i , and f S is the output of the gradient boosting model given the feature subset S . In Equation (5), ϕ n o r m is the normalized SHAP value for the i − t h feature, where ϕ i represents the absolute value of the SHAP value for feature i and ∑ ϕ j is the sum of the absolute SHAP values for all features.

3. Results

3.1. The Diurnal Dynamics of In Situ Eddy-Covariance Observations

Figure 2 reveals seasonal dynamics in carbon uptake within this subtropical forest ecosystem. The GPP consistently exhibited bimodal peaks around day-of-year 150 and 270, reaching maximal values during summer months with relative declines in spring and winter periods—a pattern that corresponds closely with seasonal variations in solar radiation and temperature. Over the entire observation period, the Tiantong forest consistently acted as a net carbon sink, and the annual mean GPP was 2057 g C m−2 y−1 during the study period.

3.2. Assessment of the Machine Learning Models Performance and Feature Importance Analysis

These models were evaluated based on statistical metrics like R2 and RMSE to quantify their estimative ability. Figure 3a illustrates the performance comparison of four machine learning methods in estimating GPP. The estimation results from the four models indicate that the SVR model performed the best, achieving an R2 value of approximately 0.76 and an RMSE of 0.17 μmol m−2 s−1 for GPP simulation, while the MLP model exhibited the weakest performance, with an R2 value of only 0.51 and an RMSE of 0.25 μmol m−2 s−1 (Table 2). Notably, the RF and GBR models demonstrated consistent performance and ranked just below SVR.
For NEE estimation, all four modeling methods performed poorly, with the best fit being achieved by SVR, whose R2 was only 0.28. Regarding RECO estimation, SVR performed the best too, with an average R2 of 0.59 and an RMSE of 0.43 μmol m−2 s−1, while RF exhibited comparable performance with an average R2 of 0.58 and an RMSE of 0.44 μmol m−2 s−1 (Table S2). The above results indicate that three machine learning models achieved reasonably good performances in simulating GPP for subtropical forests; however, their performance in simulating RECO and NEE was suboptimal, with the SVR model performing the best for GPP estimation. Based on the test and cross-validation outcomes, there are significant differences in the fitting performance among the four models. Figure 3b shows that the best fit for GPP is achieved with the combination of BRF and CMI features, yielding an R2 of 0.71 and an RMSE of 0.19 μmol m−2 s−1 (Table S3).
The importance score, calculated via permutation importance, quantifies the influence of each feature on the estimative capability of the model [58]. Higher scores indicate greater feature importance for model estimations. Results (Figure S2) indicate that for the SVR model, the most important input features are derived from CMI, followed by BRF features and PAR; notably, bands 5 and 6 achieved the highest scores among all bands (Table S4). In the RF model, the ranking of feature importance is PAR first, followed by CMI and then the BRF, while the GBR model exhibits a similar ranking: PAR, CMI, and the BRF. For the SVR model, the least important feature is NIRv, whereas in both the RF and GBR models, the least important feature is band 13. Across all three models, PAR emerges as the most important feature, with importance scores of 0.11, 0.27, and 0.47, respectively (Table S5). In contrast, the models’ performance for NEE was poor for all three feature combinations, suggesting that these features may not be closely related to NEE values. For RECO, the results differ somewhat: the CMI features produced the highest R2 (0.67) and an RMSE of 0.39 μmol m−2 s−1, while the BRF and CMI combination closely followed with an R2 of 0.65 and an RMSE of 0.39 μmol m−2 s−1 (Table S6).

3.3. The Performance of the SVR Model and the Diurnal Cycle of Photosynthesis in Subtropical Forests

The comparison of linear regression R2 values over the four years reveals that the SVR model generally captures observed forest GPP well. During all the study years, using the linear regression, the model’s performance in capturing diurnal GPP variations varies with year. The fit R2 values estimated and observed GPP values ranged from 0.69 to 0.86. the fit varies (R2 = 0.69–0.86), reflecting interannual differences in model performance, while the combined four-year data yield R2 = 0.74 (Figure 4), indicating that the model explains most of the observed variability and performs consistently over the multi-year period.
As depicted in Figure 5a–d, SVR model exhibits varying degrees of performance in capturing the diurnal cycles of GPP across the four seasons in subtropical forests. The model demonstrates accuracy during spring and summer (R2 = 0.78), while achieving the highest explanatory power in autumn (R2 = 0.84). Its estimative capability for winter GPP (R2 = 0.83) is slightly lower than that for autumn. Notably, RMSE in summer (2.64 μmol m−2 s−1) is significantly higher than in spring (2.14 μmol m−2 s−1), while winter exhibits the lowest RMSE (2.11 μmol m−2 s−1). This indicates that the diurnal GPP range is largest in summer, with higher peak values and longer durations. In contrast, winter GPP displays the lowest daily peak values, with a relatively flat daily curve and reduced amplitude.
Figure 5e–h illustrates the diurnal variations in observed GPP and SVR-estimated values across four years (2021–2024) during daylight hours. Although the curves exhibit a generally similar unimodal pattern, there are notable interannual differences between the observed and estimated values in terms of R2 and RMSE. The SVR model’s GPP estimations for 2021 are comparable to those for 2022, with R2 values of 0.84 and 0.86, respectively. However, the RMSE for 2022 is the highest among the four years at 1.36 μmol m−2 s−1. This may be attributed to the heatwave events in southern China during 2022 [59] and a higher amount of missing ground flux data compared with other years. The SVR model demonstrates excellent GPP estimation performances for 2023 and 2024, with R2 values of 0.94 and 0.93, and RMSE values of 0.85 μmol m−2 s−1 and 0.84 μmol m−2 s−1, respectively, reflecting smaller and more comparable errors.
Overall, subtropical forests exhibit a consistent single-peak pattern in their diurnal photosynthetic cycle, with notable seasonal and interannual variations. Diurnal GPP in subtropical forests peaks during summer and reaches its lowest levels in winter, with intermediate values observed in spring and autumn. This pattern closely corresponds to the vegetation growth cycle and light availability in subtropical forests. The annual peak GPP observed in 2023 may be attributed to climatic factors (e.g., precipitation, temperature) or vegetation health conditions. The diurnal variations in EC-derived and satellite-derived GPP exhibit strong consistency, demonstrating that the SVR model based on the AHI/H8 data can effectively retrieve GPP in subtropical forests.

3.4. Influencing Factors on the Diurnal Cycle of Photosynthesis in Subtropical Forests

Figure S3 shows that the diurnal variations in satellite-derived GPP and the AHI/H8-based products—PAR, EVI, and LST—across the four seasons. Across all seasons, GPP and PAR exhibit highly consistent unimodal diurnal patterns. PAR provides light energy during daylight hours, acting as the most direct driver of photosynthesis. In spring and autumn, while PAR remains a dominant driver, the moderated GPP peaks suggest additional modulation by temperature regimes and phenological factors. For GPP, within the optimal temperature range, photosynthetic rates increase with rising LST. However, excessive temperatures may induce physiological stress, leading to GPP suppression. The diurnal fluctuation patterns of EVI across the four seasons exhibit similarities, with their peaks slightly preceding the estimated GPP peaks.
We utilized four years of ground-based data—including carbon flux and meteorological measurements—to analyze the factors influencing the diurnal cycle of photosynthesis in subtropical forests (Figure 6). The heatmap illustrates the correlations between GPP and various meteorological factors over the past four years. The color and thickness of the connecting lines correspond to the direction and magnitude of correlations. In the heatmap, PAR exhibits the highest correlation with GPP (r = 0.63), followed by the vapor pressure deficit (VPD), Tair, and RH with correlation coefficients of 0.58, 0.57, and 0.54, respectively (Table S7). The result indicates that PAR exerts the strongest positive driving effect on GPP in subtropical forests, while VPD, Tair, and RH also have significant influences on GPP.
Using half-hourly observational data from 2021 to 2024 as input to the SHAP model, we calculated the relative importance of each feature by normalizing the absolute SHAP values, dividing each feature’s absolute SHAP value by the sum of all absolute SHAP values. The results (Figure 7) indicate that PAR exerts the greatest influence on the model output, with a normalized importance value of approximately 0.7 (Table S8). This finding underscores the critical role of radiation in estimating GPP.
In this study, the diurnal GPP of subtropical forests was divided into three time periods: morning (07:00–10:00), noon (11:00–14:00), and afternoon (15:00–17:00). Using four years of data from the flux tower site, we processed these datasets separately for each time period. The results (Figure 8a) show that PAR exhibited the highest normalized SHAP values in the morning and afternoon (0.53 and 0.51, respectively), highlighting its dominant role in driving GPP during these periods. In contrast, at noon, the contribution of PAR declined markedly (0.22), and Tsoil emerged as the most influential factor (0.28). Meanwhile, the relative contributions of SWC and VPD also increased compared with the morning and afternoon (Table S9).
At both the seasonal and interannual scales, PAR is the most influential factor affecting GPP model outputs. However, the contribution rate of PAR to GPP varies among different years and seasons (Figure 8b). At the seasonal scale, PAR exhibited the lowest contribution in spring (0.45) and the highest in autumn (0.59), while the contributions of Tair, Tsoil, and SWC also varied substantially across seasons (Table S10). On an annual scale (Figure 8c), the contribution rate of PAR in 2021 was the lowest among the four years, with a normalized SHAP value of 0.42, while the influences of Tair and SWC on GPP increased relative to the other three years (Table S11).

4. Discussion

4.1. Advantages of Integrating Eddy-Covariance Data with Geostationary Satellite Observations

Geostationary satellite observations hold enormous potential for monitoring the diurnal cycles of photosynthesis and respiration as well as for advancing carbon cycle research [60]. Geostationary satellite data offer several distinct advantages over polar-orbiting satellite observations when it comes to characterizing the diurnal variations in photosynthesis and respiration in forest ecosystems [61,62]. These sub-daily estimates enable the investigation of ecological processes that polar-orbiting satellites (e.g., MODIS, Landsat) cannot capture, such as mid-day depression, flash droughts, and canopy energy balance dynamics. They also provide valuable inputs for spatiotemporal gap-filling and data fusion with polar-orbiting observations. Geostationary satellites exhibit a higher temporal resolution [63]. They can capture observations at intervals as short as 5 to 10 min, allowing for continuous monitoring of the same region throughout the day [64]. This continuous coverage is crucial for detecting the full diurnal cycle and transient events [65], such as sudden changes in light intensity or temperature that directly affect photosynthesis and respiration. The high-frequency data acquired by geostationary satellites facilitate the identification of short-term environmental fluctuations that significantly influence plant physiological processes [19].
EC sites provide continuous, sub-hourly measurements of ecosystem carbon and water exchanges [66], meteorological variables, and soil moisture, and these measurements have been widely applied in carbon flux research worldwide [17]. The integration of in situ techniques with geostationary satellite observations enables the effective combination of high temporal resolution data with ground-based eddy-covariance measurements [67]. EC data provide high-frequency, site-scale measurements of carbon fluxes, enabling accurate capture of the instantaneous responses of forest photosynthesis and its diurnal variations [68]; when combined with satellite observations, EC data can reflect the physical coupling mechanisms between soil temperature, air temperature, and vegetation responses [69]. Meanwhile, although previous studies [6,7,8] represent important advances in sub-daily GPP estimation at regional and global scales—providing global sub-daily estimates and demonstrating the potential of machine learning models to predict GPP at half-hourly resolution based on flux observations—their insights into the diurnal photosynthetic dynamics of subtropical evergreen broadleaf forests in East Asia remain limited, and these ecosystems are still underrepresented in existing datasets. Our study addresses this gap by integrating high-frequency reflectance information from the H8/AHI geostationary satellite with machine learning models to reconstruct half-hourly diurnal GPP dynamics, which are further validated against multi-year eddy-covariance observations. This approach not only enhances the accuracy of ecosystem carbon flux estimations but also expands the spatial coverage of research and monitoring [19], thereby facilitating a more robust understanding of carbon dynamics within forest ecosystems. Together, these advantages enable more precise monitoring and modeling of the diurnal cycles of carbon exchange [70].

4.2. Evaluation of Machine Learning Models Performance and Diurnal Cycle Characteristics of Photosynthesis in Subtropical Forests

Using H8/AHI data in conjunction with four machine learning models (SVR, RF, GBR, and MLP), our analysis demonstrates that SVR (R2 = 0.76), RF (R2 = 0.71), and GBR (R2 = 0.71) all achieve significantly higher accuracy in estimating GPP compared with MLP (R2 = 0.51) (Table 2). Among these methods, SVR yields the highest precision, while GBR performs similarly to RF [57]. A comparison between RF and SVR in GPP estimation reveals differences in their strengths and suitable application scenarios. RF, an ensemble learning method based on decision trees, offers strong nonlinear modeling capabilities and robustness to noise, particularly when feature variables exhibit multicollinearity and noise [58]. Numerous studies have shown that RF can effectively extract features and reduce overfitting when handling high-dimensional, multi-source remote sensing data [71]. In contrast, SVR demonstrates strong generalization ability under high-dimensional, small-sample conditions, capturing complex nonlinear relationships and maintaining a high predictive accuracy when sample size is limited [72]. Furthermore, we found that RF requires longer processing times for the same amount of remote sensing data; given the relatively limited size of our dataset, we chose SVR for further research due to its higher computational efficiency and better simulation performance.
Seasonally, the SVR model achieves high estimative accuracy in autumn and winter (R2 > 0.80), while spring and summer exhibit reduced performance (Figure 5c). The model’s superior performance in autumn and winter may stem from more stable environmental conditions (e.g., temperature, solar radiation), reduced climatic disturbances, and decreased vegetation physiological activity. The availability of cloud-free H8/AHI data is lower during these seasons (Figure S2), leading to reduced data quality compared with autumn and winter, collectively increasing the complexity of GPP estimations. Moreover, due to the increased variability in meteorological factors in summer—characterized by longer and more intense periods of solar radiation and higher temperatures compared with other seasons [5]—the diurnal range and peak values of GPP in subtropical forests are significantly greater in summer [73], as evidenced by the substantially higher RMSE of 2.64 μmol m−2 s−1 (Figure 5b) observed during this season. The greater variability in summer data aligns with heightened climatic and physiological dynamics, while reduced fluctuations in winter reflect environmental stability and suppressed metabolic activity. Additionally, it is noteworthy that data gaps and heatwave events in southern China [74] led to an increased estimation error for diurnal GPP in 2022, with an RMSE of 1.36 μmol m−2 s−1—the highest among the four years (Figure 5f). These findings are consistent with previous research indicating that extreme weather conditions can alter the relative contributions of environmental drivers to ecosystem carbon fluxes [45]. SVR model effectively estimates the dynamic diurnal variations in photosynthesis in subtropical forests using H8/AHI data, and it also demonstrates strong estimative capability at both seasonal and interannual scales.

4.3. Meteorological Factors Influencing the Diurnal Cycle of Carbon Uptake in Subtropical Forests

PAR remains the dominant driver of the GPP model output throughout the day, but its relative influence declines at mid-day as Tsoil, Tair, and VPD gain prominence. This observation is consistent with the phenomenon of mid-day depression of photosynthesis [14] and underscores the amplified regulatory effect of environmental factors under high-temperature and arid conditions [75]. This phenomenon may be attributed to “mid-day depression of photosynthesis,” which results from various physiological and environmental constraints. During mid-day, despite high light intensity, factors such as increased VPD, excessive leaf temperature, and photoinhibition lead to stomatal closure, thereby reducing photosynthetic efficiency [7,42]. Consequently, the relative influence of temperature-related factors (Tsoil and Tair) increases as they mediate stress responses and respiration rates [43]. Our SHAP-based analysis reveals distinct diurnal sensitivity to environmental drivers, notably the mid-day declines in PAR contribution and increased impact of VPD and temperature, which reflects physiological mid-day depression. These intra-day shifts are ecologically meaningful and underscore the importance of capturing GPP at sub-daily resolution.
At seasonal and interannual scales, our analysis shows that in spring, the influence of PAR on GPP is relatively reduced while the impact of SWC is enhanced, suggesting that photosynthesis in subtropical vegetation during spring is more constrained by soil moisture availability. This discrepancy may reflect the combined effects of early vegetation phenology and elevated environmental variability during spring [5]. Moreover, in 2022, the contribution rate of PAR was the lowest among the four years, while the influences of Tair, Tsoil, and SWC on GPP were relatively higher compared with the other three years, likely due to extreme climatic events—such as heatwaves [76]—that occurred in 2022; from Figure S6, it can also be observed that GPP at Tiantong forest decreased to some extent during the heatwave period around day 210 in 2022. These findings provide critical insights into the diurnal cycling mechanisms of carbon fluxes in subtropical forests and their seasonal modulation. Overall, GPP exhibits a typical single-peak diurnal cycle in all seasons. In summer (Figure 5b), the peak is the highest, indicating that forest photosynthesis is more vigorous during periods characterized by higher temperatures, intense radiation, and a greater leaf area index [56]. In contrast, the winter season (Figure 5d) shows relatively lower peak values. Furthermore, each year displays a similar single-peak diurnal pattern, with slight interannual shifts in the timing of the peak—either advancing or delaying—which reflects the influence of external climate and phenological changes [4].

4.4. Limitations and Future Implications

Although geostationary satellite observations offer high temporal resolution compared to polar-orbiting satellites, they tend to have relatively poorer spatial resolution. The high temporal resolution of sub-hourly H8/AHI data may sometimes be compromised by cloud cover [77,78]. Certain optical bands can only be effectively observed during daytime, making nighttime data unavailable and thereby limiting the ability to monitor some ecological processes continuously over a 24-h period. In the subtropical region under study, the extended rainy season significantly increases the duration of cloud cover [5], which reduces the amount of cloud-free data available for ground-based carbon flux inversion, and data gaps consequently lead to a decline in model estimation accuracy.
Furthermore, the scarcity of ground-based EC flux towers poses a critical constraint—this study relied solely on flux data derived from a single EC tower [79,80], and its findings are confined to a single forest ecosystem type, limiting broader ecological extrapolation. We acknowledge that relying solely on a single EC site may introduce spatial bias and limit model generalization. To address this, we have initiated cross-site model evaluations incorporating additional EC towers in subtropical regions (e.g., Puding, Jurong) (Figure S4), with preliminary results showing that the SVR model maintains stable performances across sites (Figure S5). This ongoing effort aims to build a transferable, regionally applicable GPP prediction framework. These sites span a range of ecosystem types for the subtropical climate zone. Leveraging this multi-site dataset, we applied the SVR model in conjunction with geostationary satellite observations to assess its performance across diverse ecosystem types throughout subtropical China. We plan to conduct an in-depth study of the mechanisms by which extreme events such as heatwaves and drought affect GPP, using high-frequency H8/AHI data to capture GPP dynamics under these conditions and validating model performance with eddy-covariance data.

5. Conclusions

This study demonstrates that utilizing sub-hourly data from the H8/AHI geostationary meteorological satellite, combined with EC observations, can effectively capture diurnal variations in photosynthesis within subtropical forests. A comparative analysis of multiple machine learning models for estimating GPP in forest ecosystems revealed that the SVR model achieved the highest estimative accuracy, with the greatest coefficient of determination (R2 = 0.76) and the lowest RMSE. Further analysis revealed that subtropical forests exhibit a consistent unimodal pattern in their diurnal photosynthetic cycle, characterized by significant seasonal and interannual variability. Furthermore, through SHAP value analysis and correlation assessments, we elucidated the regulatory mechanisms of photosynthesis by various environmental factors. PAR consistently emerged as the dominant driver of GPP variability. However, during mid-day, its relative influence diminished due to the “mid-day photosynthetic depression” phenomenon, while the effects of temperature- and moisture-related factors became more pronounced. Seasonal and interannual comparisons further revealed that the proportional contributions of these environmental factors underwent significant shifts under extreme climatic events or across different phenological stages.

Supplementary Materials

The following supporting information can be downloaded at https://www.mdpi.com/article/10.3390/rs17173079/s1, Figure S1: The distribution of cloud-free remote sensing data, in 2021–2024; Figure S2: Feature importance derived from the selected SVR, RF, and GBR models, which were trained on four years of data from Tiantong forest site for GPP estimation; Figure S3 (a–d): The diurnal variations in satellite-derived GPP and AHI/H8-based products—PAR, EVI, and LST—across the four seasons; Figure S4: Distribution of vegetation types in China with the geographical location of Tiantong Mountain, Puding site and Jurong site marked; Figure S5: Diurnal cycle GPP performance of the SVR model for Puding site (a–h) and Jurong site (i–p). Figure S6: Daily variations in air temperature, precipitation, and GPP in the Tiantong Forest during 2022. Table S1: Grid search specifications and searching space for parameters of the machine learning models. Table S2: R2 and RMSE for GPP, NEE and RECO Estimation Using 4 Years of Data From one Eddy Covariance Site with Multiple ML Models. Table S3: Using the SVR model with various combinations of input features, the corresponding R2 and RMSE values for estimating GPP was determined. Table S4: Feature importance scores for the SVR, RF, and GBR models, which were trained using four years of EC data for GPP estimation. Table S5: Feature importance scores for the SVR, RF, and GBR models, which were trained using four years of eddy covariance data for GPP estimation. Table S6: Using the SVR model with various combinations of input features, the corresponding R2 and RMSE values for estimating NEE, and RECO were determined. Table S7: Mantel test for spatial coupling between GPP and environmental drivers. Table S8: The normalized SHAP values of meteorological factors based on four years (2021–2024) of data from the Tiantong forest flux site. Table S9: The normalized SHAP values for meteorological factors influencing GPP as measured at the Tiantong forest flux site. Table S10: Normalized SHAP values of GPP with respect to meteorological factors at the Tiantong forest site, grouped by spring, summer, autumn, and winter, using half-hourly data from 2021 to 2024. Table S11: Normalized SHAP values of GPP and meteorological factors at the Tiantong forest site, grouped by 2021, 2022, 2023, and 2024.

Author Contributions

Conceptualization, K.H. and J.X.; methodology, X.D.; software, X.D.; validation, J.X., X.D. and E.S.; formal analysis, X.D.; investigation, X.D.; resources, X.D. and J.X.; data curation, X.D. and J.X.; writing—original draft preparation, J.X.; writing—review and editing, J.X.; visualization, X.D.; supervision, K.H., C.H. and Z.L.; project administration, E.S.; funding acquisition, K.H. All authors have read and agreed to the published version of the manuscript.

Funding

We thank Prof. Xuhui Zhou and Prof. Jianyang Xia for their assistance in the eddy covariance flux tower. This research was funded by the National Key R&D Program of China (2022YFF0802102), the National Natural Science Foundation of China (42471121, 42250205), and the Shanghai Pilot Program for Basic Research (TQ20220102).

Data Availability Statement

The datasets generated and analyzed during the current study are available from the corresponding author on reasonable request.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Makarieva, A.M.; Nefiodov, A.V.; Rammig, A.; Nobre, A.D. Re-appraisal of the global climatic role of natural forests for improved climate projections and policies. Front. For. Glob. Change 2023, 6, 1150191. [Google Scholar] [CrossRef]
  2. Pan, Y.; Birdsey, R.A.; Phillips, O.L.; Houghton, R.A.; Fang, J.; Kauppi, P.E.; Keith, H.; Kurz, W.A.; Ito, A.; Lewis, S.L.; et al. The enduring world forest carbon sink. Nature 2024, 631, 563–569. [Google Scholar] [CrossRef]
  3. Yu, G.; Chen, Z.; Piao, S.; Peng, C.; Ciais, P.; Wang, Q.; Li, X.; Zhu, X. High carbon dioxide uptake by subtropical forest ecosystems in the East Asian monsoon region. Proc. Natl. Acad. Sci. USA 2014, 111, 4910–4915. [Google Scholar] [CrossRef]
  4. Lin, C.; Gentine, P.; Frankenberg, C.; Zhou, S.; Kennedy, D.; Li, X. Evaluation and mechanism exploration of the diurnal hysteresis of ecosystem fluxes. Agric. For. Meteorol. 2019, 278, 107642. [Google Scholar] [CrossRef]
  5. Wang, L.; Yu, M.; Ye, S.; Yan, J. Seasonal patterns of carbon and water flux responses to precipitation and solar radiation variability in a subtropical evergreen forest, South China. Agric. For. Meteorol. 2023, 342, 109760. [Google Scholar] [CrossRef]
  6. Losos, D.; Ranjbar, S.; Hoffman, S.; Abernathey, R.; Desai, A.R.; Otkin, J.; Zhang, H.; Ryu, Y.; Stoy, P.C. Rapid changes in terrestrial carbon dioxide uptake captured in near-real time from a geostationary satellite: The ALIVE framework. Remote Sens. Environ. 2025, 324, 114759. [Google Scholar] [CrossRef]
  7. Nakagawa, R.; Chau, M.; Calzaretta, J.; Keenan, T.; Vahabi, P.; Todeschini, A.; Bassiouni, M.; Kang, Y. Upscaling Global Hourly GPP with Temporal Fusion Transformer (TFT). arXiv 2023, arXiv:2306.13815. [Google Scholar] [CrossRef]
  8. Leng, J.; Chen, J.M.; Li, W.; Luo, X.; Xu, M.; Liu, J.; Wang, R.; Rogers, C.; Li, B.; Yan, Y. Global datasets of hourly carbon and water fluxes simulated using a satellite-based process model with dynamic parameterizations. Earth Syst. Sci. Data 2024, 16, 1283–1300. [Google Scholar] [CrossRef]
  9. Wang, Q.; Imasu, R.; Arai, Y.; Ito, S.; Mizoguchi, Y.; Kondo, H.; Xiao, J. Sub-Daily Natural CO2 Flux Simulation Based on Satellite Data: Diurnal and Seasonal Pattern Comparisons to Anthropogenic CO2 Emissions in the Greater Tokyo Area. Remote Sens. 2021, 13, 2037. [Google Scholar] [CrossRef]
  10. Joiner, J.; Yoshida, Y.; Zhang, Y.; Duveiller, G.; Jung, M.; Lyapustin, A.; Wang, Y.; Tucker, C.J. Estimation of Terrestrial Global Gross Primary Production (GPP) with Satellite Data-Driven Models and Eddy Covariance Flux Data. Remote Sens. 2018, 10, 1346. [Google Scholar] [CrossRef]
  11. Magnani, M.; Baneschi, I.; Giamberini, M.; Mosca, P.; Raco, B.; Provenzale, A. Drivers of carbon fluxes in Alpine tundra: A comparison of three empirical model approaches. Sci. Total Environ. 2020, 732, 139139. [Google Scholar] [CrossRef] [PubMed]
  12. Liu, Y.; Peñuelas, J.; Cescatti, A.; Zhang, Y.; Zhang, Z. Atmospheric Dryness Dominates Afternoon Depression of Global Terrestrial Photosynthesis. Geophys. Res. Lett. 2024, 51, e2024GL110954. [Google Scholar] [CrossRef]
  13. Running, S.W.; Nemani, R.R.; Heinsch, F.A.; Zhao, M.; Reeves, M.; Hashimoto, H. A Continuous Satellite-Derived Measure of Global Terrestrial Primary Production. BioScience 2004, 54, 547–560. [Google Scholar] [CrossRef]
  14. Koyama, K.; Takemoto, S. Morning reduction of photosynthetic capacity before midday depression. Sci. Rep. 2014, 4, 4389. [Google Scholar] [CrossRef] [PubMed]
  15. Xu, H.; Xiao, J.; Zhang, Z. Heatwave effects on gross primary production of northern mid-latitude ecosystems. Environ. Res. Lett. 2020, 15, 074027. [Google Scholar] [CrossRef]
  16. Kivalov, S.N.; Fitzjarrald, D.R. Observing the Whole-Canopy Short-Term Dynamic Response to Natural Step Changes in Incident Light: Characteristics of Tropical and Temperate Forests. Bound. Layer Meteorol. 2019, 173, 1–52. [Google Scholar] [CrossRef]
  17. Baldocchi, D.; Falge, E.; Gu, L.; Olson, R.; Hollinger, D.; Running, S.; Anthoni, P.; Bernhofer, C.; Davis, K.; Evans, R.; et al. FLUXNET: A New Tool to Study the Temporal and Spatial Variability of Ecosystem-Scale Carbon Dioxide, Water Vapor, and Energy Flux Densities. Bull. Am. Meteorol. Soc. 2001, 82, 2415–2434. [Google Scholar] [CrossRef]
  18. Emanuel, R.E.; Riveros-Iregui, D.A.; McGlynn, B.L.; Epstein, H.E. On the spatial heterogeneity of net ecosystem productivity in complex landscapes. Ecosphere 2011, 2, art86. [Google Scholar] [CrossRef]
  19. Xiao, J.; Fisher, J.B.; Hashimoto, H.; Ichii, K.; Parazoo, N.C. Emerging satellite observations for diurnal cycling of ecosystem processes. Nat. Plants 2021, 7, 877–887. [Google Scholar] [CrossRef]
  20. Sansone, C.; Bruno, A.; Piscitelli, C.; Baci, D.; Fontana, A.; Brunet, C.; Noonan, D.M.; Albini, A. Natural Compounds of Marine Origin as Inducers of Immunogenic Cell Death (ICD): Potential Role for Cancer Interception and Therapy. Cells 2021, 10, 231. [Google Scholar] [CrossRef]
  21. Zhao, M.; Heinsch, F.A.; Nemani, R.R.; Running, S.W. Improvements of the MODIS terrestrial gross and net primary production global data set. Remote Sens. Environ. 2005, 95, 164–176. [Google Scholar] [CrossRef]
  22. Niclòs, R.; Valor, E.; Caselles, V.; Coll, C.; Sánchez, J.M. In situ angular measurements of thermal infrared sea surface emissivity—Validation of models. Remote Sens. Environ. 2005, 94, 83–93. [Google Scholar] [CrossRef]
  23. Yao, Y.; Humphrey, V.; Konings, A.G.; Wang, Y.; Yin, Y.; Holtzman, N.; Wood, J.D.; Bar-On, Y.; Frankenberg, C. Investigating Diurnal and Seasonal Cycles of Vegetation Optical Depth Retrieved from GNSS Signals in a Broadleaf Forest. Geophys. Res. Lett. 2024, 51, e2023GL107121. [Google Scholar] [CrossRef]
  24. Tran, N.N.; Huete, A.; Nguyen, H.; Grant, I.; Miura, T.; Ma, X.; Lyapustin, A.; Wang, Y.; Ebert, E. Seasonal Comparisons of Himawari-8 AHI and MODIS Vegetation Indices over Latitudinal Australian Grassland Sites. Remote Sens. 2020, 12, 2494. [Google Scholar] [CrossRef]
  25. Arabi, B.; Salama, M.S.; Pitarch, J.; Verhoef, W. Integration of in-situ and multi-sensor satellite observations for long-term water quality monitoring in coastal areas. Remote Sens. Environ. 2020, 239, 111632. [Google Scholar] [CrossRef]
  26. Bessho, K.; Date, K.; Hayashi, M.; Ikeda, A.; Imai, T.; Inoue, H.; Kumagai, Y.; Miyakawa, T.; Murata, H.; Ohno, T.; et al. An Introduction to Himawari-8/9—Japan’s New-Generation Geostationary Meteorological Satellites. J. Meteorol. Soc. Japan. Ser. II 2016, 94, 151–183. [Google Scholar] [CrossRef]
  27. Jung, M.; Schwalm, C.; Migliavacca, M.; Walther, S.; Camps-Valls, G.; Koirala, S.; Anthoni, P.; Besnard, S.; Bodesheim, P.; Carvalhais, N.; et al. Scaling carbon fluxes from eddy covariance sites to globe: Synthesis and evaluation of the FLUXCOM approach. Biogeosciences 2020, 17, 1343–1365. [Google Scholar] [CrossRef]
  28. Yu, T.; Zhang, Q.; Sun, R. Comparison of Machine Learning Methods to Up-Scale Gross Primary Production. Remote Sens. 2021, 13, 2448. [Google Scholar] [CrossRef]
  29. Diaz-Gonzalez, F.A.; Vuelvas, J.; Correa, C.A.; Vallejo, V.E.; Patino, D. Machine learning and remote sensing techniques applied to estimate soil indicators—Review. Ecol. Indic. 2022, 135, 108517. [Google Scholar] [CrossRef]
  30. Wang, X.; Kent, M.; Fang, X.-F. Evergreen broad-leaved forest in Eastern China: Its ecology and conservation and the importance of resprouting in forest restoration. For. Ecol. Manag. 2007, 245, 76–87. [Google Scholar] [CrossRef]
  31. Song, Y.-C.; Da, L.-J.; Yang, Y.-C. Population Structure and Regeneration Types of Dominant Species in an Evergreen Broadleaved Forest in Tiantong National Forest Park, Zhejiang Province, Eastern China. Chin. J. Plant Ecol. 2004, 28, 376–384. [Google Scholar]
  32. Yang, Q.; Liu, H.; Yang, H.; Wang, Z.; Fang, X.; Ma, Z.; Wang, X. Tiantong Subtropical Forest Dynamics Plot: Tree Species and Their Distribution Patterns; China Forestry Publishing House: Beijing, China, 2020. [Google Scholar]
  33. Wang, J. Global and Transect-Scale Variations of Forest Soil Organic Carbon Residence Time and Its Driving Factors. Master’s Thesis, East China Normal University, Shanghai, China, 2019. [Google Scholar]
  34. Yan, E.-R.; Wang, X.-H.; Huang, J.-J. Shifts in plant nutrient use strategies under secondary forest succession. Plant Soil 2006, 289, 187–197. [Google Scholar] [CrossRef]
  35. Schmid, H.P. Footprint modeling for vegetation atmosphere exchange studies: A review and perspective. Agric. For. Meteorol. 2002, 113, 159–183. [Google Scholar] [CrossRef]
  36. Kljun, N.; Calanca, P.; Rotach, M.W.; Schmid, H.P. A simple two-dimensional parameterisation for Flux Footprint Prediction (FFP). Geosci. Model Dev. 2015, 8, 3695–3713. [Google Scholar] [CrossRef]
  37. Wilczak, J.M.; Oncley, S.P.; Stage, S.A. Sonic Anemometer Tilt Correction Algorithm. Bound.-Layer Meteorol. 2001, 99, 127–150. [Google Scholar] [CrossRef]
  38. Fan, S.-M.; Wofsy, S.C.; Bakwin, P.S.; Jacob, D.J.; Fitzjarrald, D.R. Atmosphere-biosphere exchange of CO2 and O3 in the central Amazon Forest. J. Geophys. Res. Atmos. 1990, 95, 16851–16864. [Google Scholar] [CrossRef]
  39. Webb, E.K.; Pearman, G.I.; Leuning, R. Correction of flux measurements for density effects due to heat and water vapour transfer. Q. J. R. Meteorol. Soc. 1980, 106, 85–100. [Google Scholar] [CrossRef]
  40. Lasslop, G.; Reichstein, M.; Papale, D.; Richardson, A.D.; Arneth, A.; Barr, A.; Stoy, P.; Wohlfahrt, G. Separation of net ecosystem exchange into assimilation and respiration using a light response curve approach: Critical issues and global evaluation. Glob. Change Biol. 2010, 16, 187–208. [Google Scholar] [CrossRef]
  41. Kurihara, Y.; Murakami, H.; Kachi, M. Sea surface temperature from the new Japanese geostationary meteorological Himawari-8 satellite. Geophys. Res. Lett. 2016, 43, 1234–1240. [Google Scholar] [CrossRef]
  42. Dong, S.; Cheng, J.; Shi, J.; Shi, C.; Sun, S.; Liu, W. A Data Fusion Method for Generating Hourly Seamless Land Surface Temperature from Himawari-8 AHI Data. Remote Sens. 2022, 14, 5170. [Google Scholar] [CrossRef]
  43. Li, Y.; Chen, W.; Zhang, Y.; Tao, C.; Xiao, R.; Tan, Y. Accurate cloud detection in high-resolution remote sensing imagery by weakly supervised deep learning. Remote Sens. Environ. 2020, 250, 112045. [Google Scholar] [CrossRef]
  44. Ishida, H.; Nakjima, T.Y.; Yokota, T.; Kikuchi, N.; Watanabe, H. Investigation of GOSAT TANSO-CAI Cloud Screening Ability through an Intersatellite Comparison. J. Appl. Meteorol. Climatol. 2011, 50, 1571–1586. [Google Scholar] [CrossRef]
  45. Ishida, H.; Nakajima, T.Y. Development of an unbiased cloud detection algorithm for a spaceborne multispectral imager. J. Geophys. Res. Atmos. 2009, 114, 2008JD010710. [Google Scholar] [CrossRef]
  46. Badgley, G.; Field, C.B.; Berry, J.A. Canopy near-infrared reflectance and terrestrial photosynthesis. Sci. Adv. 2017, 3, e1602244. [Google Scholar] [CrossRef]
  47. Rouse, J.W.; Haas, R.H.; Schell, J.A.; Deering, D.W. Monitoring vegetation systems in the Great Plains with ERTS. In Proceedings of the Third Earth Resources Technology Satellite (ERTS) Symposium, Washington, DC, USA, 10–14 December 1973; pp. 309–317. [Google Scholar]
  48. Huete, A.; Didan, K.; Miura, T.; Rodriguez, E.P.; Gao, X.; Ferreira, L.G. Overview of the radiometric and biophysical performance of the MODIS vegetation indices. Remote Sens. Environ. 2002, 83, 195–213. [Google Scholar] [CrossRef]
  49. Jacquemoud, S.; Baret, F. PROSPECT: A model of leaf optical properties spectra. Remote Sens. Environ. 1990, 34, 75–91. [Google Scholar] [CrossRef]
  50. Yu, Y.; Tarpley, D.; Privette, J.L.; Flynn, L.E.; Xu, H.; Chen, M.; Vinnikov, K.Y.; Sun, D.; Tian, Y. Validation of GOES-R Satellite Land Surface Temperature Algorithm Using SURFRAD Ground Measurements and Statistical Es-timates of Error Properties. IEEE Trans. Geosci. Remote Sens. 2012, 50, 704–713. [Google Scholar] [CrossRef]
  51. Schmit, T.J.; Li, J.; Li, J.; Feltz, W.F.; Gurka, J.J.; Goldberg, M.D.; Schrab, K.J. The GOES-R Advanced Baseline Imager and the Continuation of Current Sounder Products. J. Appl. Meteorol. Climatol. 2008, 47, 2696–2711. [Google Scholar] [CrossRef]
  52. Varghese, R.; Cherukuri, A.K.; Doddrell, N.H.; Doss, C.G.P.; Simkin, A.J.; Ramamoorthy, S. Machine learning in photosynthesis: Prospects on sustainable crop development. Plant Sci. 2023, 335, 111795. [Google Scholar] [CrossRef]
  53. Song, E.; Zhu, X.; Shao, G.; Tian, L.; Zhou, Y.; Jiang, A.; Lu, J. Multi-Temporal Remote Sensing Inversion of Evapotranspiration in the Lower Yangtze River Based on Landsat 8 Remote Sensing Data and Analysis of Driving Factors. Remote Sens. 2023, 15, 2887. [Google Scholar] [CrossRef]
  54. Bichri, H.; Chergui, A.; Hain, M. Investigating the Impact of Train/Test Split Ratio on the Performance of Pre-Trained Models with Custom Datasets. Int. J. Adv. Comput. Sci. Appl. 2024, 15. [Google Scholar] [CrossRef]
  55. Pawluszek-Filipiak, K.; Borkowski, A. On the Importance of Train–Test Split Ratio of Datasets in Automatic Landslide Detection by Supervised Classification. Remote Sens. 2020, 12, 3054. [Google Scholar] [CrossRef]
  56. Maxwell, A.E.; Warner, T.A.; Fang, F. Implementation of machine-learning classification in remote sensing: An applied review. Int. J. Remote Sens. 2018, 39, 2784–2817. [Google Scholar] [CrossRef]
  57. Ranjbar, S.; Losos, D.; Hoffman, S.; Cuntz, M.; Stoy, P.C. Using Geostationary Satellite Observations and Machine Learning Models to Estimate Ecosystem Carbon Uptake and Respiration at Half Hourly Time Steps at Eddy Covariance Sites. J. Adv. Model. Earth Syst. 2024, 16, e2024MS004341. [Google Scholar] [CrossRef]
  58. Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar] [CrossRef]
  59. Gong, H.; Ma, K.; Hu, Z.; Dong, Z.; Ma, Y.; Chen, W.; Wu, R.; Wang, L. Attribution of the August 2022 Extreme Heatwave in Southern China: Role of Dynamical and Thermodynamical Processes. Bull. Am. Meteorol. Soc. 2024, 105, E193–E199. [Google Scholar] [CrossRef]
  60. Li, X.; Xiao, J.; Fisher, J.B.; Baldocchi, D.D. ECOSTRESS estimates gross primary production with fine spatial resolution for different times of day from the International Space Station. Remote Sens. Environ. 2021, 258, 112360. [Google Scholar] [CrossRef]
  61. Li, X.; Ryu, Y.; Xiao, J.; Dechant, B.; Liu, J.; Li, B.; Jeong, S.; Gentine, P. New-generation geostationary satellite reveals widespread midday depression in dryland photosynthesis during 2020 western U.S. heatwave. Sci. Adv. 2023, 9, eadi0775. [Google Scholar] [CrossRef]
  62. Khan, A.M.; Stoy, P.C.; Joiner, J.; Baldocchi, D.; Verfaillie, J.; Chen, M.; Otkin, J.A. The Diurnal Dynamics of Gross Primary Productivity Using Observations From the Advanced Baseline Imager on the Geostationary Operational Environmental Satellite-R Series at an Oak Savanna Ecosystem. J. Geophys. Res. Biogeosci. 2022, 127, e2021JG006701. [Google Scholar] [CrossRef]
  63. Amoroso, N.; Cilli, R.; Maggipinto, T.; Monaco, A.; Tangaro, S.; Bellotti, R. Satellite data and machine learning reveal a significant correlation between NO(2) and COVID-19 mortality. Environ. Res. 2022, 204 Pt A, 111970. [Google Scholar] [CrossRef]
  64. Shang, H.; Letu, H.; Nakajima, T.Y.; Wang, Z.; Ma, R.; Wang, T.; Lei, Y.; Ji, D.; Li, S.; Shi, J. Diurnal cycle and seasonal variation of cloud cover over the Tibetan Plateau as determined from Himawari-8 new-generation geostationary satellite data. Sci. Rep. 2018, 8, 1105. [Google Scholar] [CrossRef]
  65. Chatzopoulos-Vouzoglanis, K.; Reinke, K.J.; Soto-Berelov, M.; Jones, S.D. One year of near-continuous fire monitoring on a continental scale: Comparing fire radiative power from polar-orbiting and geostationary observations. Int. J. Appl. Earth Obs. Geoinf. 2023, 117, 103214. [Google Scholar] [CrossRef]
  66. Yun, S.-I.; Kang, B.-M.; Lim, S.-S.; Choi, W.-J.; Ko, J.; Yoon, S.; Ro, H.-M.; Kim, H.-Y. Further understanding CH4 emissions from a flooded rice field exposed to experimental warming with elevated [CO2]. Agric. For. Meteorol. 2012, 154–155, 75–83. [Google Scholar] [CrossRef]
  67. Losos, D.; Hoffman, S.; Stoy, P.C. GOES-R land surface products at Western Hemisphere eddy covariance tower locations. Sci. Data 2024, 11, 277. [Google Scholar] [CrossRef]
  68. Baldocchi, D.D. Assessing the eddy covariance technique for evaluating carbon dioxide exchange rates of ecosystems: Past, present and future. Glob. Change Biol. 2003, 9, 479–492. [Google Scholar] [CrossRef]
  69. Richardson, A.D.; Andy Black, T.; Ciais, P.; Delbart, N.; Friedl, M.A.; Gobron, N.; Hollinger, D.Y.; Kutsch, W.L.; Longdoz, B.; Luyssaert, S.; et al. Influence of spring and autumn phenological transitions on forest ecosystem productivity. Philos. Trans. R. Soc. B Biol. Sci. 2010, 365, 3227–3246. [Google Scholar] [CrossRef]
  70. Teng, D.; Gong, X.; He, X.; Wang, J.; Lv, G.; Wang, J.; Yang, X. Impact of meteorological variability on diurnal and seasonal net ecosystem productivity in a desert riparian forest ecosystem. Front. Plant. Sci. 2024, 15, 1332192. [Google Scholar] [CrossRef] [PubMed]
  71. Belgiu, M.; Drăguţ, L. Random forest in remote sensing: A review of applications and future directions. ISPRS J. Photogramm. Remote Sens. 2016, 114, 24–31. [Google Scholar] [CrossRef]
  72. Jung, M.; Reichstein, M.; Margolis, H.; Cescatti, A.; Richardson, A.; Arain, A.; Arneth, A.; Bonal, D.; Chen, J.; Gianelle, D.; et al. Global Patterns of Land-Atmosphere Fluxes of Carbon Dioxide, Latent Heat, and Sensible Heat Derived from Eddy Covariance, Satel-lite, and Meteorological Observations. J. Geophys. Res. 2011, 116. [Google Scholar] [CrossRef]
  73. Wang, M.; Zhang, L. Synchronous Changes of GPP and Solar-Induced Chlorophyll Fluorescence in a Subtropical Evergreen Coniferous Forest. Plants 2023, 12, 2224. [Google Scholar] [CrossRef]
  74. Chen, D.; Qiao, S.; Yang, J.; Tang, S.; Zuo, D.; Feng, G. Contribution of anthropogenic influence to the 2022-like Yangtze River valley compound heatwave and drought event. npj Clim. Atmos. Sci. 2024, 7, 172. [Google Scholar] [CrossRef]
  75. Oren, R.; Sperry, J.S.; Katul, G.G.; Pataki, D.E.; Ewers, B.E.; Phillips, N.; Schäfer, K.V.R. Survey and synthesis of intra- and interspecific variation in stomatal sensitivity to vapour pressure deficit. Plant Cell Environ. 1999, 22, 1515–1526. [Google Scholar] [CrossRef]
  76. Duan, A.; Zhong, Y.; Xu, G.; Yang, K.; Tian, B.; Wu, Y.; Bai, H.; Hu, E. Quantifying the 2022 extreme drought in the Yangtze River Basin using GRACE-FO. J. Hydrol. 2024, 630, 130680. [Google Scholar] [CrossRef]
  77. Zhao, R.; Li, Z.-w.; Feng, G.-c.; Wang, Q.-j.; Hu, J. Monitoring surface deformation over permafrost with an improved SBAS-InSAR algorithm: With emphasis on climatic factors modeling. Remote Sens. Environ. 2016, 184, 276–287. [Google Scholar] [CrossRef]
  78. Kaced, K.; Larbes, C.; Ramzan, N.; Bounabi, M.; Dahmane, Z.E. Bat algorithm based maximum power point tracking for photovoltaic system under partial shading conditions. Sol. Energy 2017, 158, 490–503. [Google Scholar] [CrossRef]
  79. Pumpanen, J.; Lindén, A.; Miettinen, H.; Kolari, P.; Ilvesniemi, H.; Mammarella, I.; Hari, P.; Nikinmaa, E.; Heinonsalo, J.; Bäck, J.; et al. Precipitation and net ecosystem exchange are the most important drivers of DOC flux in upland boreal catchments. J. Geophys. Res. Biogeosci. 2014, 119, 1861–1878. [Google Scholar] [CrossRef]
  80. Hill, T.; Chocholek, M.; Clement, R. The case for increasing the statistical power of eddy covariance ecosystem studies: Why, where and how? Glob. Change Biol. 2017, 23, 2154–2165. [Google Scholar] [CrossRef]
Remotesensing 17 03079 g001
Figure 1. Tiantong forest site. (a) Distribution of vegetation types in China with the geographical location of Tiantong Mountain (TTS) marked. (b) The Tiantong National Forest Park (c) Eddy-covariance flux tower in the Tiantong forest. (d) Eddy-covariance observation system. (e) The flux footprint of the flux tower.
Figure 1. Tiantong forest site. (a) Distribution of vegetation types in China with the geographical location of Tiantong Mountain (TTS) marked. (b) The Tiantong National Forest Park (c) Eddy-covariance flux tower in the Tiantong forest. (d) Eddy-covariance observation system. (e) The flux footprint of the flux tower.
Remotesensing 17 03079 g001
Remotesensing 17 03079 g002
Figure 2. Multi-year means of daily GPP at Tiantong forest during the study period. The shaded area denotes the deviation values.
Figure 2. Multi-year means of daily GPP at Tiantong forest during the study period. The shaded area denotes the deviation values.
Remotesensing 17 03079 g002
Remotesensing 17 03079 g003
Figure 3. (a) GPP estimation with four machine learning models using four years of the AHI/H8 data. R2- and RMSE-based individual models trained on 4 years of data. (b) A comparison of R2 and RMSE for GPP estimation using up to 4 years of H8/AHI data and the selected SVR model using different BRF and CMI feature configurations.
Figure 3. (a) GPP estimation with four machine learning models using four years of the AHI/H8 data. R2- and RMSE-based individual models trained on 4 years of data. (b) A comparison of R2 and RMSE for GPP estimation using up to 4 years of H8/AHI data and the selected SVR model using different BRF and CMI feature configurations.
Remotesensing 17 03079 g003
Remotesensing 17 03079 g004
Figure 4. Comparison of EC-derived GPP vs. satellite-derived GPP during the study period. Flux tower GPP was partitioned from NEE using the daytime-based method. The satellite-derived GPP was retrieved using the SVR model based on the AHI/H8 data. Different colors represent each year.
Figure 4. Comparison of EC-derived GPP vs. satellite-derived GPP during the study period. Flux tower GPP was partitioned from NEE using the daytime-based method. The satellite-derived GPP was retrieved using the SVR model based on the AHI/H8 data. Different colors represent each year.
Remotesensing 17 03079 g004
Remotesensing 17 03079 g005
Figure 5. (a–d) The estimation performance of the SVR model and the diurnal cycle characteristics of seasonal GPP in subtropical forests. (e–h) Interannual changes in the diurnal cycle characteristics of GPP in subtropical forests and the estimation performance of the SVR model.
Figure 5. (a–d) The estimation performance of the SVR model and the diurnal cycle characteristics of seasonal GPP in subtropical forests. (e–h) Interannual changes in the diurnal cycle characteristics of GPP in subtropical forests and the estimation performance of the SVR model.
Remotesensing 17 03079 g005
Remotesensing 17 03079 g006
Figure 6. Heatmap of Pearson correlation analysis between GPP and meteorological factors at the Tiantong forest site, 2021–2024. ***, **, and * denote statistical significance at p ≤ 0.001, 0.01, and 0.05 levels, respectively.
Figure 6. Heatmap of Pearson correlation analysis between GPP and meteorological factors at the Tiantong forest site, 2021–2024. ***, **, and * denote statistical significance at p ≤ 0.001, 0.01, and 0.05 levels, respectively.
Remotesensing 17 03079 g006
Remotesensing 17 03079 g007
Figure 7. SHAP values illustrating the impacts of meteorological factors on EC-derived GPP (2021–2024). The bar chart displays the normalized SHAP values for individual meteorological factors. The plot displays normalized contributions of factors including VPD, PAR, Tair, Tsoil, RH, soil water content (SWC), evapotranspiration (ET), and Precipitation (P), with positive values enhancing and negative values reducing GPP.
Figure 7. SHAP values illustrating the impacts of meteorological factors on EC-derived GPP (2021–2024). The bar chart displays the normalized SHAP values for individual meteorological factors. The plot displays normalized contributions of factors including VPD, PAR, Tair, Tsoil, RH, soil water content (SWC), evapotranspiration (ET), and Precipitation (P), with positive values enhancing and negative values reducing GPP.
Remotesensing 17 03079 g007
Remotesensing 17 03079 g008
Figure 8. (a) Normalized SHAP values of meteorological factors for GPP at the Tiantong forest site, grouped by morning, noon, and afternoon using four years (2021–2024) of half-hourly data. (b) Normalized SHAP values of GPP with respect to meteorological factors at the Tiantong forest site, grouped by spring, summer, autumn, and winter, using half-hourly data from 2021 to 2024. (c) Normalized SHAP values of GPP and meteorological factors at the Tiantong forest site, grouped by 2021, 2022, 2023, and 2024.
Figure 8. (a) Normalized SHAP values of meteorological factors for GPP at the Tiantong forest site, grouped by morning, noon, and afternoon using four years (2021–2024) of half-hourly data. (b) Normalized SHAP values of GPP with respect to meteorological factors at the Tiantong forest site, grouped by spring, summer, autumn, and winter, using half-hourly data from 2021 to 2024. (c) Normalized SHAP values of GPP and meteorological factors at the Tiantong forest site, grouped by 2021, 2022, 2023, and 2024.
Remotesensing 17 03079 g008
Table 1. H8/AHI Bands Information.
Table 1. H8/AHI Bands Information.
BandNameCenter Wavelength (μm)Spatial Resolution
(km)		Observation Objective
1Blue0.471Vegetation, aerosol detection
2Green0.511Vegetation, aerosol detection
3Red0.640.5Vegetation, low cloud/fog detection
4NIR0.861Vegetation, aerosol detection
5SWIR1.62Cloud phase discrimination
6SWIR2.32Cloud particle size analysis
7MIR3.92Wildfire monitoring
8TIR6.22Upper-level tropospheric water vapor density
9TIR6.92Mid-upper tropospheric water vapor density
10TIR7.32Mid-tropospheric water vapor density
11TIR8.62Sulfur dioxide content analysis
12TIR9.62Ozone content monitoring
13TIR10.42Sea surface temperature (SST) measurement
14TIR11.22Cloud imaging, SST measurement
15TIR12.42Cloud imaging, SST measurement
16TIR13.32Cloud top height estimation
Table 2. R2 and RMSE for GPP estimation using four years of data with four machine learning models.
Table 2. R2 and RMSE for GPP estimation using four years of data with four machine learning models.
GPP
TestValidation
R2RMSER2RMSE
RF0.980.010.710.19
GBR0.860.120.710.19
MLP0.510.250.510.25
SVR0.860.130.760.17
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Xu, J.; Dai, X.; Liu, Z.; He, C.; Song, E.; Huang, K. Geostationary Satellite-Derived Diurnal Cycles of Photosynthesis and Their Drivers in a Subtropical Forest. Remote Sens. 2025, 17, 3079. https://doi.org/10.3390/rs17173079

AMA Style

Xu J, Dai X, Liu Z, He C, Song E, Huang K. Geostationary Satellite-Derived Diurnal Cycles of Photosynthesis and Their Drivers in a Subtropical Forest. Remote Sensing. 2025; 17(17):3079. https://doi.org/10.3390/rs17173079

Chicago/Turabian Style

Xu, Jiang, Xi Dai, Zhibin Liu, Chenyang He, Enze Song, and Kun Huang. 2025. "Geostationary Satellite-Derived Diurnal Cycles of Photosynthesis and Their Drivers in a Subtropical Forest" Remote Sensing 17, no. 17: 3079. https://doi.org/10.3390/rs17173079

APA Style

Xu, J., Dai, X., Liu, Z., He, C., Song, E., & Huang, K. (2025). Geostationary Satellite-Derived Diurnal Cycles of Photosynthesis and Their Drivers in a Subtropical Forest. Remote Sensing, 17(17), 3079. https://doi.org/10.3390/rs17173079

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop