Gross Primary Production Products Research Articles

Why models underestimate West African tropical forest primary productivity

Tropical forests dominate terrestrial photosynthesis, yet there are major contradictions in our understanding due to a lack of field studies, especially outside the tropical Americas. A recent field study indicated that West African forests have among the highest forests gross primary productivity (GPP) yet observed, contradicting models that rank them lower than Amazonian forests. Here, we show possible reasons for this data-model mismatch. We found that biometric GPP measurements are on average 56.3% higher than multiple global GPP products at the study sites. The underestimation of GPP largely disappears when a standard photosynthesis model is informed by local field-measured values of (a) fractional absorbed photosynthetic radiation (fAPAR), and (b) photosynthetic traits. Remote sensing products systematically underestimate fAPAR (33.9% on average at study sites) due to cloud contamination issues. The study highlights the potential widespread underestimation of tropical forests GPP and carbon cycling and hints at the ways forward for model and input data improvement.

Terrestrial photosynthesis inferred from plant carbonyl sulfide uptake.

Terrestrial photosynthesis, or gross primary production (GPP), is the largest carbon flux in the biosphere, but its global magnitude and spatiotemporal dynamics remain uncertain1. The global annual mean GPP is historically thought to be around 120 PgC yr-1 (refs. 2-6), which is about 30-50 PgC yr-1 lower than GPP inferred from the oxygen-18 (18O) isotope7 and soil respiration8. This disparity is a source of uncertainty in predicting climate-carbon cycle feedbacks9,10. Here we infer GPP from carbonyl sulfide, an innovative tracer for CO2 diffusion from ambient air to leaf chloroplasts through stomata and mesophyll layers. We demonstrate that explicitly representing mesophyll diffusion is important for accurately quantifying the spatiotemporal dynamics of carbonyl sulfide uptake by plants. From the estimate of carbonyl sulfide uptake by plants, we infer a global contemporary GPP of 157 (±8.5) PgC yr-1, which is consistent with estimates from 18O (150-175 PgC yr-1) and soil respiration ( PgC yr-1), but with an improved confidence level. Our global GPP is higher than satellite optical observation-driven estimates (120-140 PgC yr-1) that are used for Earth system model benchmarking. This difference predominantly occurs in the pan-tropical rainforests and is corroborated by ground measurements11, suggesting a more productive tropics than satellite-based GPP products indicated. As GPP is a primary determinant of terrestrial carbon sinks and may shape climate trajectories9,10, our findings lay a physiological foundation on which the understanding and prediction of carbon-climate feedbacks can be advanced.

Intercomparison of gross primary productivity in spatio-temporal distribution over a typical tropical region based on different datasets.

Gross Primary Productivity (GPP) plays a vital role in the carbon cycle of terrestrial ecosystems. For the purpose of assessing the performance of various GPP products in a typical tropical area, this study conducted an intercomparison of seven different GPP products over Hainan Island, China, and analyzed spatiotemporal characteristics of GPP in different elevation regions and vegetation types over study area. The results showed that although the GPP spatial distribution pattern over Hainan Island, as presented by different GPP products, is similar- showing that GPP values in the southwest are higher than those in the northeast- there are significant differences in the spatial distributions and magnitudes of each GPP product. The multi-annual mean GPP values from the seven GPP products exhibited similar temporal variation trends, with GOSIF being obviously higher than other GPP products. The GPP products indicated significant differences in different elevation and vegetation type groups. The GPP gap between GLASS AVHRR and BESS V2 decreases slightly with the increase in elevation. Most GPP products show a consistent change trend over time. There were also significant differences in the results of significance tests for the change trends of GPP products. Though GOSIF is significantly higher than other GPP products, it demonstrated a relatively high correlation with all GPP products, while AGPP exhibited relatively low correlations with all GPP products. The quantitatively comparative analysis of seven different GPP products in this study is valuable for improving GPP retrieval models and can also serve as a scientific reference for the application of GPP products in tropical region vegetation studies.

Improving global gross primary productivity estimation using two-leaf light use efficiency model by considering various environmental factors via machine learning

Distinguishing gross primary productivity (GPP) into sunlit (GPPsu) and shaded (GPPsh) components is critical for understanding the carbon exchange between the atmosphere and terrestrial ecosystems under climate change. Recently, the two-leaf light use efficiency (TL-LUE) model has proven effective for simulating global GPPsu and GPPsh. However, no known physical method has focused on integrating the overall constraint of intricate environmental factors on photosynthetic capability, and seasonal differences in the foliage clumping index (CI), which most likely influences GPP estimation in LUE models. Here, we propose the TL-CRF model, which uses the random forest technique to integrate various environmental variables, particularly for terrestrial water storage (TWS), into the TL-LUE model. Moreover, we consider seasonal differences in CI at a global scale. Based on 267 global eddy covariance flux sites, we explored the functional response of vegetation photosynthesis to key environmental factors, and trained and evaluated the TL-CRF model. The TL-CRF model was then used to simulate global eight-day GPP, GPPsu, and GPPsh from 2002 to 2020. The results show that the relative prediction error of environmental stress factors on the maximum LUE is reduced by approximately 52 % when these factors are integrated via the RF model. Thus the accuracy of global GPP estimation (R2 = 0.87, RMSE = 0.94 g C m−2 d−1, MAE = 0.61 g C m−2 d−1) in the TL-CRF model is greater than that (R2 = 0.76, RMSE = 2.18 g C m−2 d−1, MAE = 1.50 g C m−2 d−1) in the TL-LUE model, although this accuracy awaits further investigation among the released GPP products. TWS exerts the greatest control over ecosystem photosynthesis intensity, making it a suitable water indicator. Furthermore, the results confirm an optimal minimum air temperature for photosynthesis. Overall, these findings indicate a promising method for producing a new global GPP dataset, advancing our understanding of the dynamics and interactions between photosynthesis and environmental factors.

Negative Asymmetric Response of Pantropical Gross Primary Productivity to Precipitation Anomalies

AbstractThe carbon sink in pantropical biomes play a crucial role in modulating the inter‐annual variations of global terrestrial carbon balance and is threatened by extreme climate events. However, it has not been carefully examined whether an increase in tropical gross primary productivity (GPP) can compensate the decrease during precipitation anomalies. Using the asymmetry index (AI) and multiple GPP products, we assessed responses of pantropical GPP to precipitation anomalies during 2001–2022. Positive AI indicates that GPP increases are greater than GPP decreases during precipitation anomalies, and vice versa. Our results showed an average negative pantropical GPP asymmetry, that is, GPP decreases exceeded the GPP increases during precipitation anomalies. In addition, a positive AI was found in tropical hyper‐arid and arid regions, which is opposite to the negative AI observed in tropical semi‐arid, sub‐humid, and humid regions. This suggest that tropical GPP asymmetry changes from positive to negative as the moisture increases. Notably, a significant decreasing trend of negative AI was observed over the entire tropical region, indicating that the negative effect of inter‐annual precipitation variations on pantropical vegetation productivity has enhanced. Considering the model predicted increasing climate variability and extremes, the negative impact of precipitation variability on tropical carbon cycle may continue to intensify. Lastly, the divergence in AI estimates among multiple GPP products highlight the need to further improve our understanding of the response of tropical carbon cycle to climate changes, especially for the tropical humid regions.

Contrasting responses of vegetation productivity to intraseasonal rainfall in Earth system models

Abstract. Correctly representing the response of vegetation productivity to water availability in Earth system models (ESMs) is essential for accurately modelling the terrestrial carbon cycle and the evolution of the climate system. Previous studies evaluating gross primary productivity (GPP) in ESMs have focused on annual mean GPP and interannual variability, but physical processes at shorter timescales are important for determining vegetation–climate coupling. We evaluate GPP responses at the intraseasonal timescale in five CMIP6 ESMs by analysing changes in GPP after intraseasonal rainfall events with a timescale of approximately 25 d. We compare these responses to those found in a range of observation-based products. When composited around all intraseasonal rainfall events globally, both the amplitude and the timing of the GPP response show large inter-model differences, demonstrating discrepancies between models in their representation of water–carbon coupling processes. However, the responses calculated from the observational datasets also vary considerably, making it challenging to assess the realism of the modelled GPP responses. The models correctly capture the fact that larger increases in GPP at the regional scale are associated with larger increases in surface soil moisture and larger decreases in atmospheric vapour pressure deficit. However, the sensitivity of the GPP response to these drivers varies between models. The GPP in NorESM is insufficiently sensitive to vapour pressure deficit perturbations when compared all to other models and six out of seven observational GPP products tested. Most models produce a faster GPP response where the surface soil moisture perturbation is larger, but the observational evidence for this relationship is weak. This work demonstrates the need for a better understanding of the uncertainties in the representation of water–vegetation relationships in ESMs and highlights a requirement for future daily-resolution observations of GPP to provide a tighter constraint on global water–carbon coupling processes.

A Data‐Driven Approach to Assess the Impact of Climate Change on a Tropical Mangrove in India

AbstractAs a potential carbon sink, mangroves play an important role in climate mitigation. India houses several major global mangrove patches, which remain vulnerable to climate change. The ecosystem‐atmosphere CO2 exchange is most accurately measured by the eddy covariance method, whereas satellites provide the biophysical parameters for a wider area. In the present study, the Sentinel‐2 satellite data is used to map the land cover types in the Pichavaram mangrove forest and identify two major dominant species (Rhizophora spp. and Avicennia marina), which indicated more than 95% classification accuracy. We used 2 years (2017 and 2018) of in situ gross primary productivity (GPP) and leaf area index (LAI) measurements and rectified the Moderate Resolution Imaging Spectroradiometer (MODIS) GPP and LAI products from 2010 to 2018. The modified MODIS GPP and LAI products were used to develop machine learning models, that is, Random Forest (RF) and Extreme Gradient Boosting (XGBoost) to study the climate influence on mangrove productivity. The RF model (R2 = 0.85 and root mean square error (RMSE) = 0.2) outperformed the XGBoost model (R2 = 0.75 and RMSE = 0.26) and was used to project the impact of climate change on the mangrove GPP for two extreme climate change scenarios, namely SSP1‐1.26 and SSP5‐8.5. The GPP increases and decreases in future during wet and dry periods, respectively. Overall, the projected GPP indicated a reduction of 3.73%–20.3% from 2050 to 2060 and of 4.82%–28.15% from 2090 to 2100, compared to its current average (from 2010 to 2018).

Incorporating Spatial Autocorrelation into GPP Estimation Using Eigenvector Spatial Filtering

Terrestrial gross primary productivity (GPP) is a critical part of land carbon fluxes. Accurately quantifying GPP in terrestrial ecosystems and understanding its spatiotemporal dynamics are essential for assessing the capability of vegetation to absorb carbon from the atmosphere. Nevertheless, traditional remote sensing estimation models often require complex parameters and data inputs, and they do not account for spatial effects resulting from the distribution of monitoring sites. This can lead to biased parameter estimation and unstable results. To address these challenges, we have raised a spatial autocorrelation light gradient boosting machine model (SA-LGBM) to enhance GPP estimation. SA-LGBM combines reflectance information from remote sensing observations with eigenvector spatial filtering (ESF) methods to create a set of variables that capture continuous spatiotemporal variations in plant functional types and GPP. SA-LGBM demonstrates promising results when compared to existing GPP products. With the inclusion of eigenvectors, we observed an 8.5% increase in R2 and a 20.8% decrease in RMSE. Furthermore, the residuals of the model became more random, reducing the inherent spatial effects within them. In summary, SA-LGBM represents the first attempt to quantify the impact of spatial autocorrelation and addresses the limitations of underestimation present in existing GPP products. Moreover, SA-LGBM exhibits favorable applicability across various vegetation types.

Spatiotemporal Variability of Gross Primary Productivity in Türkiye: A Multi-Source and Multi-Method Assessment

We investigated the spatiotemporal variability of remotely sensed gross primary productivity (GPP) over Türkiye based on MODIS, TL-LUE, GOSIF, MuSyQ, and PMLV2 GPP products. The differences in various GPP products were assessed using Kruskal–Wallis and Mann–Whitney U methods, and long-term trends were analyzed using Modified Mann–Kendall (MMK), innovative trend analysis (ITA), and empirical mode decomposition (EMD). Our results show that at least one GPP product significantly differs from the others over the seven geographic regions of Türkiye (χ2 values of 50.8, 21.9, 76.9, 42.6, 149, 34.5, and 168; p < 0.05), and trend analyses reveal a significant increase in GPP from all satellite-based products over the latter half of the study period. Throughout the year, the average number of months in which each dataset showed significant increases across all study regions are 6.7, 8.1, 5.9, 9.6, and 8.7 for MODIS, TL-LUE, GOSIF, MuSyQ, and PMLV2, respectively. The ITA and EMD methods provided additional insight into the MMK test in both visualizing and detecting trends due to their graphical techniques. Overall, the GPP products investigated here suggest ‘greening’ for Türkiye, consistent with the findings from global studies, but the use of different statistical approaches and satellite-based GPP estimates creates different interpretations of how these trends have emerged. Ground stations, such as eddy covariance towers, can help further improve our understanding of the carbon cycle across the diverse ecosystem of Türkiye.

An Improved Gross Primary Production Model Considering Atmospheric CO2 Fertilization: The Qinghai–Tibet Plateau as a Case Study

Involving the effect of atmospheric CO2 fertilization is effective for improving the accuracy of estimating gross primary production (GPP) using light use efficiency (LUE) models. However, the widely used LUE model, the remote sensing-driven Carnegie–Ames–Stanford Approach (CASA) model, scarcely considers the effects of atmospheric CO2 fertilization, which causes GPP estimation uncertainties. Therefore, this study proposed an improved method for estimating GPP by integrating the atmospheric CO2 concentration into the CASA model and generated a long time series GPP dataset with high precision for the Qinghai–Tibet Plateau. The CASA model was improved by considering the impact of atmospheric CO2 on vegetation productivity and discerning variations in CO2 gradients within the canopy and leaves. A 500 m monthly GPP dataset for the Qinghai–Tibet Plateau from 2003 to 2020 was generated. The results showed that the improved GPP estimation model achieved better performances on estimating GPP (R2 = 0.68, RMSE = 406 g C/m2/year) than the original model (R2 = 0.67, RMSE = 499.32 g C/m2/year) and MODIS GPP products (R2 = 0.49, RMSE = 522.56 g C/m2/year). The GPP on the Qinghai–Tibet Plateau increased significantly with the increase in atmospheric CO2 concentration and the gradual accumulation of dry matter. The improved method can also be used for other regions and the generated GPP dataset is valuable for further understanding the ecosystem carbon cycles on the Qinghai–Tibet Plateau.

Comparative assessment of leaf photosynthetic capacity datasets for estimating terrestrial gross primary productivity

The maximum Rubisco carboxylation rate normalized to 25 °C (Vcmax25) is a key parameter in terrestrial biosphere models for simulating carbon cycling. Recently, global distributions of Vcmax25 have been derived through various methods and different data, including field measurements, ecological optimality theory (EOT), leaf chlorophyll content (LCC), and solar-induced chlorophyll fluorescence (SIF). However, direct validation poses challenges due to high uncertainty arising from limited ground-based observations. This study conducted an indirect evaluation of four Vcmax25 datasets by assessing the accuracy of gross primary productivity (GPP) simulated using the Biosphere-atmosphere Exchange Process Simulator (BEPS) at both site and global scales. Results indicate that, compared to utilizing Vcmax25 fixed by plant functional types (PFT) derived from field measurements, incorporating Vcmax25 derived from SIF and LCC (SIF + LCC), or solely LCC, into BEPS significantly reduces simulated errors in the annual total GPP, with a 23.2 %–25.1 % decrease in the average absolute bias across 196 FLUXNET2015 sites. Daily GPP for evergreen needleleaf forests, deciduous broadleaf forests, shrublands, grasslands, and croplands shows a 7.8 %–27.6 % decrease in absolute bias, primarily attributed to reduced simulation errors during off-peak seasons of vegetation growth. Conversely, the annual total GPP error simulated using EOT-derived Vcmax25 increases slightly (2.2 %) compared to that simulated using PFT-fixed Vcmax25. This is primarily due to a significant overestimation in evergreen broadleaf forests and underestimation in croplands, despite slight increased accuracy for other PFTs. The global annual GPP simulated using Vcmax25 with seasonal variations (i.e., LCC Vcmax25 and SIF + LCC Vcmax25) yields a 4.3 %–7.3 % decrease compared to that simulated using PFT-fixed Vcmax25. Compared to FLUXCOM and GOSIF GPP products, the GPP simulated based on SIF + LCC Vcmax25 and LCC Vcmax25 demonstrates better consistency (R2 = 0.91–0.93, RMSE = 314.2–376.6 g C m−2 yr−1). This study underscores the importance of accurately characterizing the spatiotemporal variations in Vcmax25 for the accurate simulation of global vegetation productivity.

Spatiotemporal patterns and driving factors of gross primary productivity over the Mongolian Plateau steppe in the past 20 years

The Eurasian steppe is the largest temperate grassland in the world. The grassland of the Mongolian Plateau (MP) represents an important part of the Eurasian steppe with high climatic sensitivity. Gross primary productivity (GPP) is a key indicator of the grassland's production, status and dynamic on the MP. In this study, we calibrated and evaluated the grassland-specific light use efficiency model (GRASS-LUE) against the observed GPP collected from nine eddy covariance flux sites on the MP, and compared the performance with other four GPP products (MOD17, VPM, GLASS and GOSIF). GRASS-LUE with higher R2 (0.91) and lower root mean square error (RMSE = 0.99 gC m−2 day−1) showed a better performance compared to the four GPP products in terms of model accuracy and dynamic consistency, especially in typical and desert steppe. The parameters of the GRASS-LUE are more suitable for water-limited grassland could be the reason for its outstanding performance in typical and desert steppe. Mean grassland GPP derived from GRASS-LUE was higher in the east and lower in the west of the MP. Grassland GPP was on average 205 gC m−2 over the MP between 2001 and 2020 with mean annual total GPP of 322 TgC yr−1. 30 % of the MP steppe showed a significant GPP increase. Growing season precipitation is the main factor affecting GPP of the MP steppe across regions. Anthropogenic factors (livestock density and population density) had greater effect on GPP than growing season temperature in pastoral counties in IM that take grazing as one of main industries. These findings can inform the status and trend of the productivity of MP steppe and help government and scientific research institutions to understand the drivers for spatial pattern of grassland GPP on the MP.

Contrasting responses of relationship between solar-induced fluorescence and gross primary production to drought across aridity gradients

Satellite solar-induced chlorophyll fluorescence (SIF) can provide valuable insights on terrestrial gross primary production (GPP) and its response to climate from region to global scales. However, the relationship between SIF and GPP under different climate conditions (such as drought) is not yet fully understood and remains a topic of debate in the literature. Here we examined the responses of the SIF–GPP relationship (i.e., the slope k of the linear SIF–GPP relationship, GPP = k × SIF) to drought along the aridity gradient using seven satellite SIF datasets, three GPP products, and 62 flux towers. The results revealed divergent SIF–GPP relationships under different aridity and drought conditions based on different SIF and GPP datasets. Overall, we found that with increasing aridity index (AI) (more humid), k derived from SIF with afternoon overpass (dailySIFPM) decreased, while k from SIF with morning overpass (dailySIFAM) increased. When there is a drought, for arid and semi-arid regions with AI from 0 to 0.4, kPM (the slope of dailySIFPM–GPP) increased, but kAM (the slope of dailySIFAM–GPP) decreased. The dailySIFPM also showed larger decreases than dailySIFAM and GPP during the drought. The contrasting responses among two groups of SIF may result from different SIF overpass time which detected different levels of water stresses on photosynthesis. The dryland ecosystems may exhibit strong midday and afternoon depression in photosynthesis under stress, which led to underestimation of daily SIF that was converted from midday/afternoon-observed SIF by assuming ideal symmetrical diurnal cycle of solar radiation. Our study highlights that in addition to multiple previously reported factors, different observation time of SIF may also affect the SIF–GPP relationship.

Joint improvement on absorbed photosynthetically active radiation and intrinsic quantum yield efficiency algorithms in the P model betters the estimate of terrestrial gross primary productivity

Terrestrial gross primary productivity (GPP) plays a crucial role in global carbon cycle and budget. A range of light use efficiency (LUE) models have been developed to estimate GPP at different spatial scales. However, large uncertainties still remain in GPP output from such models, mainly owing to the difficulty in the proper determination of maximum light use efficiency (LUEmax). The recently developed P model directly quantifies actual LUE based on environmental and physiological factors, reducing the uncertainty in GPP estimation caused by the ambiguous LUEmax. However, the existing P models still suffer from a potential underestimation in global validation, which might be related to the calculation of absorbed photosynthetically active radiation (APAR) and intrinsic quantum yield efficiency (φ0) in the P model. In this study, we improved the P model by differentiating sunlit and shaded leaves for APAR calculation and by considering the spatial variability of the optimal temperature for φ0 (Topt_phi0). The roles of modified APAR and φ0 calculations in reducing the uncertainty of simulated GPP were assessed through model simulation experiments with monthly tower-based GPP from 120 flux sites globally distributed as the benchmark. The validation indicates that simultaneous improvements on APAR and φ0 algorithms are able to better the performance of the P model. Monthly GPP simulated with this improved P model (i.e., IP model) is in good agreement with tower-based values, with a linear R2, root mean squared error, mean error, mean absolute error, and DISO of 0.76, 1.88 g C m−2d−1, -0.14 g C m−2d−1, 1.77 g C m−2d−1 and 0.60. The total global terrestrial GPP output from the IP model (IPGPP) averaged 134.3 Pg C yr−1 during 1981—2021, in the upper range of existing GPP products.

A new global time-series GPP production: DFRF-GPP

Understanding the global ecosystem carbon cycle requires a quantitative estimation of terrestrial gross primary production (GPP). Although most existing GPP estimates can accurately reflect the spatial distribution, interannual trends and even seasonal cycles of GPP under average conditions to some extent, the GPP estimates of different models under extreme climate conditions are still unsatisfactory. In this study, based on the random forest algorithm, we integrated the multimodel GPP simulation results published by the Multiscale Synthesis and Terrestrial Model Intercomparison Project, the FLUXNET flux-site-observed GPP, the standardized precipitation index (SPI) and the standardized temperature index (STI) to generate a set of global GPP time-series data products from 2001 to 2010. The new GPP product was named DFRF-GPP, referring to the GPP generated by data fusion based on random forest. The quality of DFRF-GPP was evaluated by comparison with current commonly used GPP datasets and GPP observations at FLUXNET flux sites. The results showed that DFRF-GPP is in great agreement with other GPP products and can reproduce the spatial and temporal patterns and interannual trends of global GPP, with optimal performance in regions such as those north of 30° in the Northern Hemisphere where there are more FLUXNET flux sites. DFRF-GPP has higher accuracy and excellent performance on drought and high-temperature samples, with higher sensitivity to drought and high temperature. This study not only provides a set of time-series products of global gross primary production but also provides a convenient and practical solution for enhancing the applicability of GPP estimation products in large-scale extreme climate studies.

Mango‐GPP: A Process‐Based Model for Simulating Gross Primary Productivity of Mangrove Ecosystems

AbstractMangrove ecosystems are becoming increasingly important in global climate mitigation. However, large gaps still exist in evaluating mangroves' gross primary productivity (GPP) due to reasons such as the specific influences, for example, temperature and salt stresses are poorly described in Earth System Models (ESMs). This study developed a process‐based biogeochemical model (Mango‐GPP) to improve the GPP simulation in natural and restored mangroves. The model integrates mangrove‐specific physiological processes, including the response to salt and temperature stresses, as well as the light‐use efficiency at different growing stages. Eddy covariance flux measurements at two natural sites and one restored site in China were used to calibrate and validate Mango‐GPP. The model was calibrated by inverse analysis approach based on two cases and independently validated against the other cases. The validation results showed that it was generally capable of simulating the seasonal and interannual GPP variations at different sites. The simulated daily and annual GPPs agreed well with the observations and yielded R2 of 0.67 and 0.96, with model efficiency of 0.64 and 0.93, respectively. In comparison, Mango‐GPP showed better performances than many current satellite‐based GPP products and ESMs. The model was more sensitive to solar radiation, carbon dioxide concentration, and leaf traits. Future improvements should focus on enhancing Mango‐GPP's descriptive power of key processes, and further simulating other carbon fluxes at regional scales. This work provides a model foundation for further simulating carbon exchanges between the atmosphere, mangrove, and ocean for studying the coastal wetland restoration on regional carbon neutrality.

A water stress factor based on normalized difference water index substantially improved the accuracy of light use efficiency model for arid and semi-arid grasslands

High-accuracy simulation of gross primary productivity (GPP) is crucial for the monitoring and evaluation of the ecosystem services and the adaptive management of grassland. The light use efficiency (LUE) model is one of the most widely-used methods to simulate GPP, given its simple structure and low input requirements. Current LUE models are less applicable to grasslands than other vegetation types and have lower overall estimation accuracy in arid and semi-arid regions. A grassland-specific light use efficiency model (GRASS-LUE), which optimizes three important parameters (the fraction of absorbed photosynthetically active radiation FPAR, optimum temperature Topt and water stress factor f(W)), has been developed to improve the accuracy of GPP simulation for grasslands along aridity gradients. GPP simulated by the GRASS-LUE agreed well with the eddy covariance (EC) GPP estimates for grasslands along the aridity gradient at 8-day (coefficient of determination (R2) = 0.85, Bias = −0.67 gC m−2 day−1), monthly (R2 = 0.88, Bias = −22.33 gC m−2 month−1) and annual time scales (R2 = 0.95, Bias = −118.91 gC m−2 year−1). Compared with five state-of-the-art GPP products (PML, MOD17, rEC-LUE, VPM and BESS), GRASS-LUE had the best and most stable performance in reproducing EC GPP, especially for semi-arid grassland, with the highest global performance indicator (GPI) value. Sensitivity tests further revealed that: 1) modifying f(W) to be based on the Normalized Difference Water Index (NDWI) substantially improved the model accuracy for arid and semi-arid grasslands and 2) using the minimum of temperature and water stress factors (i.e., min⁡(f(W),f(T))) to represent environmental stress in GRASS-LUE was better than that from the multiplication of temperature and water stress factors (i.e., f(W)×f(T)).

Large Divergence of Satellite Monitoring of Diffuse Radiation Effect on Ecosystem Water‐Use Efficiency

AbstractWater‐use efficiency (WUE), defined as the ratio of gross primary productivity (GPP) and evapotranspiration (ET), is a key indicator of the coupling of carbon and water cycles in terrestrial ecosystems. Diffuse radiation significantly affects WUE as diffuse radiation has stronger fertilization effects on GPP than ET. Remote sensing data‐driven (RS) models were widely used to estimate WUE. However, the performance of RS models in simulating diffuse radiation effect on WUE is poorly understood. We used GPP and ET products from MODIS, PML, BESS, and GLASS to explore the response of WUE to diffuse radiation across different vegetation types. We found that there was a large divergence among the RS models in estimating the response of WUE to FDP. PML model performed better than other RS models in simulating diffuse radiation effect on WUE. This research has significant implications for improving model simulations and understanding global carbon and water cycles.

Machine learning algorithms improve MODIS GPP estimates in United States croplands

Introduction: Machine learning methods combined with satellite imagery have the potential to improve estimates of carbon uptake of terrestrial ecosystems, including croplands. Studying carbon uptake patterns across the U.S. using research networks, like the Long-Term Agroecosystem Research (LTAR) network, can allow for the study of broader trends in crop productivity and sustainability.Methods: In this study, gross primary productivity (GPP) estimates from the Moderate Resolution Imaging Spectroradiometer (MODIS) for three LTAR cropland sites were integrated for use in a machine learning modeling effort. They are Kellogg Biological Station (KBS, 2 towers and 20 site-years), Upper Mississippi River Basin (UMRB - Rosemount, 1 tower and 12 site-years), and Platte River High Plains Aquifer (PRHPA, 3 towers and 52 site-years). All sites were planted to maize (Zea mays L.) and soybean (Glycine max L.). The MODIS GPP product was initially compared to in-situ measurements from Eddy Covariance (EC) instruments at each site and then to all sites combined. Next, machine learning algorithms were used to create refined GPP estimates using air temperature, precipitation, crop type (maize or soybean), agroecosystem, and the MODIS GPP product as inputs. The AutoML program in the h2o package tested a variety of individual and combined algorithms, including Gradient Boosting Machines (GBM), eXtreme Gradient Boosting Models (XGBoost), and Stacked Ensemble.Results and discussion: The coefficient of determination (r2) of the raw comparison (MODIS GPP to EC GPP) was 0.38, prior to machine learning model incorporation. The optimal model for simulating GPP across all sites was a Stacked Ensemble type with a validated r2 value of 0.87, RMSE of 2.62 units, and MAE of 1.59. The machine learning methodology was able to successfully simulate GPP across three agroecosystems and two crops.

A flexible and efficient knowledge-guided machine learning data assimilation (KGML-DA) framework for agroecosystem prediction in the US Midwest

Process-based models are widely used to predict the agroecosystem dynamics, but such modeled results often contain considerable uncertainty due to the imperfect model structure, biased model parameters, and inaccurate or inaccessible model inputs. Data assimilation (DA) techniques are widely adopted to reduce prediction uncertainty by calibrating model parameters or dynamically updating the model state variables using observations. However, high computational cost, difficulties in mitigating model structural error, and low flexibility in framework development hinder its applications in large-scale agroecosystem predictions. In this study, we addressed these challenges by proposing a novel DA framework that integrates a Knowledge-Guided Machine Learning (KGML)-based surrogate with tensorized ensemble Kalman filter (EnKF) and parallelized particle swarm optimization (PSO) to effectively assimilate historical and in-season multi-source remote sensing data. Specifically, we incorporate knowledge from a process-based model, ecosys, into a Gated Recurrent Unit (GRU)-based hierarchical neural network. The hierarchical architecture of KGML-DA mimics key processes of ecosys and builds a causal relationship between target variables. Using carbon budget quantification in the US Corn-Belt as a context, we evaluated KGML-DA's performance in predicting key processes of the carbon cycle at three agricultural sites (US-Ne1, US-Ne2, US-Ne3), along with county-level (627 counties) and 30-m pixel-level (Champaign County, IL) grain yield. The site experiments show that updating the upstream variable, e.g., gross primary production (GPP), improved the prediction of downstream variables such as ecosystem respiration, net ecosystem exchange, biomass, and leaf area index (LAI), with RMSE reductions ranging from 9.2% to 30.5% for corn and 4.8% to 24.6% for soybean. Uncertainty in downstream variables was automatically constrained after correcting the upstream variables, demonstrating the effectiveness of the causality linkages in the hierarchical surrogate. We found joint use of in-season GPP and evapotranspiration (ET) products along with historical GPP and surveyed yields achieved the best prediction for county-level yields, while assimilating in-season LAI observations benefitted the prediction in extreme years. Uncertainty and error analysis of regional yield estimation demonstrated that KGML-DA could reduce prediction error by 26.5% for corn and 36.2% for soybean. Remarkably, the GPU-based tensor operation design makes this DA framework more than 7000 times faster than the PB model with a High-Performance Computing system, indicating the high potential of the proposed framework for in-season, high-resolution agroecosystem predictions.