Simulating the Carbon, Nitrogen, and Phosphorus of Plant Above-Ground Parts in Alpine Grasslands of Xizang, China
by
, ,
Mingxue Xiang
1
,
Gang Fu
2,*,
Jianghao Cheng
1,
Tao Ma
1,
Yunqiao Ma
1,
Kai Zheng
1 and
Zhaoqi Wang
1
1
State Key Laboratory of Plateau Ecology and Agriculture, Qinghai University, Xining 810018, China
2
Lhasa Plateau Ecosystem Research Station, Key Laboratory of Ecosystem Network Observation and Modeling, Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing 100101, China
*
Author to whom correspondence should be addressed.
Agronomy 2025, 15(6), 1413; https://doi.org/10.3390/agronomy15061413
Submission received: 5 May 2025
/
Revised: 31 May 2025
/
Accepted: 6 June 2025
/
Published: 9 June 2025
(This article belongs to the Special Issue Advanced Machine Learning in Agriculture)
Abstract
Carbon (C), nitrogen (N), and phosphorus (P) act as pivotal regulators of biogeochemical cycles, steering organic matter decomposition and carbon sequestration in terrestrial ecosystems through the stoichiometric properties of photosynthetic organs. Deciphering their multi-scale spatiotemporal dynamics is central to unraveling plant nutrient strategies and their coupling mechanisms with global element cycling. In the current study, we modeled biogeochemical parameters (C/N/P contents, stoichiometry, and pools) in plant aboveground parts by using the growing mean temperature, total precipitation, total radiation, and maximum normalized difference vegetation index (NDVImax) across nine models (i.e., random forest model, generalized boosting regression model, multiple linear regression model, artificial neural network model, generalized linear regression model, conditional inference tree model, extreme gradient boosting model, support vector machine model, and recursive regression tree) in Xizang grasslands. The results showed that the random forest model had the highest predictive accuracy for nitrogen content, C:P, and N:P ratios under both grazing and fencing conditions (training R2 ≥ 0.61, validation R2 ≥ 0.95). Additionally, the random forest model had the highest predictive accuracy for C:N ratios under fencing conditions (training R2 = 0.84, validation R2 = 1.00), as well as for C pool and P content and pool under grazing conditions (training R2 ≥ 0.62, validation R2 ≥ 0.90). Therefore, the random forest algorithm based on climate data and/or the NDVImax demonstrated superior predictive performance in modeling these biogeochemical parameters.
1. Introduction
The carbon (C), nitrogen (N), and phosphorus (P) contents and stoichiometric ratios (C:N, C:P, N:P) in aboveground plant components are critical indicators for understanding biogeochemical cycles and energy flows in terrestrial ecosystems [1,2]. These parameters directly regulate vegetation productivity, litter decomposition rates, and nutrient resorption efficiency, thereby influencing the carbon sink capacity of ecosystems [3]. Despite their recognized importance, two major limitations persist in the current research: (1) most studies rely on small-scale experimental datasets [4,5] or low-resolution historical surveys [6], failing to capture dynamic variations across large spatiotemporal scales [7,8]; and (2) traditional modeling approaches (e.g., single-source remote sensing retrievals) are confounded by the synergistic interactions of climatic variability and anthropogenic activities [9], leading to inadequate decoupling of drivers. Consequently, developing frameworks that integrate multisource data has emerged as a critical pathway for improving predictive accuracy [10].
The alpine grasslands of the Qinghai–Xizang Plateau, as the world’s highest-altitude terrestrial ecosystem [11], have predominantly been studied for conventional metrics like biomass and soil nutrients, while systematic modeling of plant C/N/P stoichiometric parameters remains underexplored. Although recent advancements have established predictive models for biodiversity and soil moisture, dynamic simulations of aboveground elemental contents, ratios, and pools under contrasting grazing and fencing management regimes are still lacking. Existing models struggle to disentangle climate-driven effects from anthropogenic disturbances, severely limiting the precision of regional ecological restoration strategies.
While big data mining techniques such as random forest (RF) and support vector machines (SVM) have gained traction in ecological modeling, their performance in predicting elemental stoichiometry remains contentious. For instance, gradient boosting models (GBR) excel in non-ecological domains [12] but face uncertainties in adapting to nutrient heterogeneity in alpine grasslands [13]. Similarly, artificial neural networks (ANN) demonstrate proficiency in modeling nonlinear relationships [14], yet may falter due to spatiotemporal autocorrelation inherent in ecological datasets [15]. To resolve these controversies, we hypothesize that the random forest algorithm (RF), leveraging its unique capacity to effectively integrate multi-source data (field observations, NDVI, climatic variables), may offer distinct advantages across different management regimes. These advantages could include overcoming the collinearity constraints that are often encountered in traditional remote sensing models [16].
To test this hypothesis, this study systematically evaluates the predictive accuracy of nine mainstream algorithms (RF, GBR, MLR, etc.) for modeling C/N/P contents, stoichiometric ratios, and pools in the Qinghai–Xizang Plateau’s alpine grasslands. Leveraging field observations, the maximum normalized difference vegetation index (NDVImax), and climatic drivers (mean growing season temperature, total precipitation, and total radiation), we developed simulation frameworks under contrasting fencing and grazing regimes. Temperature regulates enzymatic processes governing photosynthesis and nutrient uptake [1], precipitation determines water-mediated nutrient transport [6], and radiation drives photosynthetic carbon fixation [17], while NDVImax serves as a vegetation productivity proxy capturing peak greenness dynamics [18]. This integrative approach—simultaneously addressing the temperature–precipitation–radiation triad and vegetation response—overcomes collinearity limitations in prior climate-only or spectral-driven models. The study innovates in three dimensions: (1) comparative analysis of nine algorithms’ performance across training and validation phases, (2) accuracy divergence between fencing (stable nutrient dynamics) and grazing (herbivore-driven heterogeneity) conditions, and (3) establishment of a multidimensional evaluation framework (R2, mean square errors, and field-measured deviations). As the most comprehensive algorithmic comparison for C/N/P modeling in Xizang’s alpine grasslands to date, this work provides a methodological paradigm for scaling stoichiometric predictions across heterogeneous landscapes.
2. Materials and Methods
2.1. Study Area and Plant Sampling
The research focused on the alpine grassland ecosystems of Xizang Autonomous Region, China (26°00′–36°32′ N, 78°24′–99°06′ E, Figure A1). This region exhibits a characteristic plateau monsoon climate with distinct seasonal patterns: warm–humid summers (June–August mean temperature: 7.2–14.6 °C) alternating with cold–arid winters (December–February mean temperature: −12.5 to −4.3 °C). Precipitation follows strong intra-annual variability, with 72–89% of the annual total (103–694 mm) concentrated during the summer monsoon season (July–September) [19]. Encompassing approximately 1.2 million km2 of alpine grasslands—representing > 60% of China’s total alpine grassland area—the study region forms one of Earth’s most extensive high-altitude pastoral ecosystems [20]. The vegetation primarily consists of alpine meadow (Kobresia pygmaea communities, dominant species including Stipa capillacea, Carex atrofusca, and Kobresia pygmaean) and alpine steppe (Stipa purpurea communities, dominant species including Kobresia pygmaea, Oxytropis glacialis, Carex moorcroftii, Leontopodium nanum, and Gentiana algida), developing on cryoturbated soils at elevations between 3800 and 5200 m a.s.l. Despite its ecological significance as a key carbon sink and water conservation zone, this fragile ecosystem faces increasing climatic stresses, with recorded warming rates (0.3–0.4 °C per decade–1 since 1960) exceeding global alpine region averages [21].
Plant aboveground parts were systematically sampled across representative alpine meadow and steppe communities between 2009 and 2022 (geospatial distribution is detailed in Figure A1). Standardized quadrats (0.5 m × 0.5 m for meadows; 1 m × 1 m for steppes) were established following the ecosystem-specific sampling protocols. Harvested specimens were immediately stored in dark, ventilated containers to prevent photodegradation and microbial decomposition prior to processing. All samples underwent standardized preparation: oven-drying at 65 °C for 48 h to a constant mass, followed by grinding through a 0.5 mm sieve using a Wiley mill. Carbon and nitrogen concentrations were determined using an elemental analyzer (Vario MACRO cube, Elementar, Hanau, Germany), while phosphorus content was quantified through molybdenum blue spectrophotometry after sulfuric acid digestion. Stoichiometric ratios (C:N, C:P, N:P) and elemental pools (C, N, P) were calculated using established biomass–element concentration relationships [1]. Finally, there were 313 and 341, 326 and 341, 285 and 315, 289 and 341, 277 and 310, 279 and 312, 148 and 264, 157 and 271, and 153 and 266 data for plant aboveground C content, N content, P content, C:N, C:P, N:P, C pool, N pool, and P pool under fencing and grazing conditions, respectively.
2.2. Normalized Difference Vegetation Index and Climate
The maximum normalized difference vegetation index (NDVImax) data were sourced from the National Ecosystem Science Data Center, which operates under China’s National Science and Technology Infrastructure framework [http://www.nesdc.org.cn (accessed on 10 July 2024); https://doi.org/10.12199/nesdc.ecodb.rs.2021.012 (accessed on 10 July 2024)] [22]. The NDVImax dataset features a spatial resolution of 30 m; climatic parameters, including mean growing season temperature, cumulative precipitation, and total solar radiation, were derived from spatially interpolated meteorological surfaces constructed using ground-based measurements from 145 weather stations (Figure A1). The interpolated climate surfaces originally featured 1 km spatial granularity, which underwent systematic downscaling to a 30 m resolution through our geospatial processing pipeline prior to conducting cross-scale analyses. Validation protocols across multiple environmental research applications have confirmed the dataset’s demonstrated robust predictive validity, establishing its suitability for cross-scale ecological modeling applications. Under exclosure scenarios, the three principal bioclimatic drivers (mean growing season temperature, cumulative precipitation, and solar irradiance) constituted the predictor variables in our methodological framework. Conversely, grazing regime models required the incorporation of vegetation productivity dynamics, integrating NDVImax with these three climatic parameters to account for vegetation-mediated feedback mechanisms within pastoral ecosystems.
2.3. Model Methodology
We implemented stratified random sampling using the “createDataPartition” function from the R “caret” package (version 4.2.2 for windows), with stratification based on management regimes (fencing and grazing) and ecological gradients (vegetation composition). This dual stratification preserved proportional representation of both anthropogenic management intensities (75%:25% training/validation split within each stratum) and natural environmental variation. The validation subset (n = 30) maintained equivalent stratum weightings to the training data while preventing spatial autocorrelation. This methodological separation ensured the rigorous prevention of data leakage while maintaining proportional representation of ecological gradients across both subsets. Sample sizes varied across management regimes and measured parameters. Under fencing conditions, aboveground plant community metrics included carbon/nitrogen/phosphorus contents (n = 251/261/229), C:N/C:P/N:P stoichiometric ratios (239/227/229), and carbon/nitrogen/phosphorus pools (119/127/123). Grazing regime datasets comprised 273/273/255 samples for elemental contents, 273/250/250 for stoichiometric ratios, and 214/219/216 for ecological pools, maintaining consistent measurement protocols across treatments.
All modeling frameworks were implemented in R version 4.2.2 [23], leveraging discipline-specific computational packages. The random forest (RF) model utilized the “randomForest package” in R [24], with generalized boosting regression (GBR) executed through the “gbm” package [25]. Support vector machines (SVM) were operationalized via the “e1071” library [26], while recursive regression tree (RRT) employed the “rpart” package [27]. Multiple linear regression (MLR) implementations derived from the native “stats” package within base R. In contrast, four advanced algorithms—artificial neural networks (ANN), generalized linear regression (GLR), conditional inference trees (CIT), and extreme gradient boosting (eXGB)—were systematically integrated through the “rminer” meta-package [28], ensuring computational consistency across heterogeneous model architectures. All package dependencies were validated against version-controlled repositories prior to analytical workflows.
Specifically, RF, handles high-dimensional data, reducing overfitting, fast training, noise resistance, and providing feature importance via parallel regression tree ensembles. GBR optimizes accuracy sequentially by correcting errors of prior trees, capturing complex nonlinear relationships. SVM identifies optimal hyperplanes for classification/regression, maximizing margin and minimizing misclassification. MLR is simple, interpretable, and effective for linear relationships between variables. RRT handles high-dimensional and nonlinear data via recursive partitioning. ANN mimics neural systems to adapt to intricate patterns. GLR extends linear regression to generalized distributions, accommodating non-normal data. CIT splits feature space to minimize conditional entropy/variance. eXGB enhances gradient boosting with efficient tree construction. Together, these models offer diverse strengths for regression and classification tasks [16]. RF and GBR leverage ensemble learning (parallel vs. sequential approaches), while SVM provides robust hyperplane optimization. MLR and GLR prioritize simplicity and adaptability to linear or generalized distributions, whereas RRT and CIT excel in recursive partitioning for interpretable, high-dimensional insights. ANN mimics neural adaptability for nonlinear patterns, and eXGB refines boosting with advanced regularization and efficiency. By combining their strengths—ensembles, flexibility, interpretability, and optimization—these models address varied data challenges, enhancing predictive accuracy and robustness across applications.
2.4. Model Accuracy Evaluation
Methodological divergence in training error quantification emerged across modeling frameworks due to differential R package implementations. The random forest model (RF) optimized mean squared error (MSE) minimization with R2 diagnostics (Table A1), while the generalized boosted regression model (GBR) employed mean training error and cross-validation error thresholds (Table A2). Support vector machines (SVM) utilized residual distributions and hyperplane decision values (Table A3). Multilinear regression (MLR) and regression trees (RRT) relied solely on R2 optimization (Table A4 and Table A5). Generalized linear models (GLR), neural networks (ANN), conditional inference trees (CIT), and extreme gradient boosting (eXGB) adopted proprietary error minimization protocols (Table A6). To ensure cross-model comparability, we implemented methodological harmonization through four standardized validation metrics: relative bias, root mean squared error (RMSE), linear regression slope coefficients, and R2 concordance between model predictions and empirical measurements.
Below, there is a schematic representation of the study design, integrating field ecology, biogeochemical analysis, and machine learning frameworks (Figure 1).
3. Results
3.1. Model Construction
The structural parameters of intrinsic ensemble models (tree aggregations) versus individual learners showed systematic variations across architectures (random forest, generalized boosted regression, support vector machine; Table A1, Table A2 and Table A3). In fencing systems, RF’s decision tree ensembles (543–979 trees) occupied an intermediate position between generalized boosted regression’s near-saturated forests (998–1000 trees) and support vector machine’s kernel-based models (101–208 support vectors). This hierarchy persisted under grazing conditions, with random forest maintaining 84–95% of generalized boosted regression’s tree counts while surpassing support vector machine’s vector quantities by 4.1–8.3-fold.
Intrinsic ensemble modeling (specifically random forest/generalized boosted regression’s tree aggregation) demonstrated superior predictive performance across management regimes (Table A1, Table A4 and Table A5): Under fencing: random forest (R2 = 0.53–0.91) explained 32–86% more variance than multiple linear regression (R2 = 0.06–0.53). In grazing systems, random forest (R2 = 0.61–0.89) outperformed multiple linear regression (R2 = 0.12–0.30) by 2.0–7.4-fold, with NDVI emerging as a key predictor (Table A4: β = 14.10 for carbon content).
Computational homogeneity was observed in training errors across the artificial neural network, generalized linear regression, conditional inference tree, and extreme gradient boosting algorithms (Table A6). Under fenced conditions, the conditional inference tree predicted more accurately with error magnitudes spanning 2.23–3865.5 (median = 175.21). Grazing conditions induced algorithmic divergence, where the artificial neural network maintained minimum errors (2.69–6973.16) while extreme gradient boosting showed maximum variability (16.18–11,751.26).
3.2. Model Validation
Under fencing and grazing management scenarios, comparative analysis of nine computational algorithms revealed distinct performance patterns in elemental ratio prediction accuracy (Table 1 and Table 2, Figure 2, Figure 3, Figure 4, Figure 5, Figure 6 and Figure 7). The RF algorithm outperformed other models, exhibiting the lowest magnitude of relative bias (|RB|) between model estimates and observational data for nitrogen content, carbon-to-phosphorus (C:P), and nitrogen-to-phosphorus (N:P) ratios across both management regimes. Conversely, the extreme gradient boosting method showed the largest deviation metrics for these parameters.
Algorithm performance exhibited management-specific variations. Under fencing conditions, random forest achieved minimal |RB| values for carbon-to-nitrogen (C:N) ratio estimation, whereas extreme gradient boosting produced maximum deviations in this parameter. In grazing systems, random forest maintained its predictive advantage for phosphorus-related metrics, showing the smallest |RB| values for both phosphorus content and phosphorus pool estimations, while extreme gradient boosting generated the largest discrepancies in these measurements. Notably, random forest’s enhanced performance extended to coupled carbon–phosphorus pool estimations under grazing management.
4. Discussion
4.1. Algorithm Performance in Ecosystem Stoichiometry Modeling
Our comparative analysis revealed fundamental differences in machine learning architectures when modeling grassland elemental ratios under contrasting management regimes. RF demonstrated structural and predictive superiority across metrics, maintaining intermediate tree densities (543–979 stems under fencing; 840–949 under grazing) that balanced model complexity with ecological generalizability (Table A1, Table A2 and Table A3). This architectural optimization translated to robust predictive accuracy (R2 = 0.53–0.91 fencing; 0.61–0.89 grazing), outperforming both SVM implementations and GBR configurations by 2.0–7.4-fold (Table A4 and Table A5). The 4.1–8.3-fold greater tree counts in RF versus SVM implementations reflect limitations in capturing grazing-induced nutrient heterogeneity, while GBR’s maximal tree counts (997–1000 stems) provided no accuracy gains over RF’s leaner architecture, challenging assumptions about ensemble size–performance relationships [29,30,31].
RF’s predictive dominance extended to elemental ratio estimation, achieving minimal relative bias (|RB|) for nitrogen content, C:P, and N:P ratios across management scenarios (Table 1 and Table 2, Figure 2, Figure 3, Figure 4, Figure 5, Figure 6 and Figure 7). This aligns with its documented resistance to outliers in ecological datasets [32] and capacity to model threshold effects governing grassland stoichiometry [33]. The algorithm’s resilience to herbivore-mediated data noise likely stems from bootstrap-aggregating mechanisms that mitigate localized nutrient redistribution artifacts [34]. In contrast, eXGB exhibited substantial validation errors despite its competitive training performance, particularly for N:P ratios. This divergence is attributable to gradient boosting architectures’ sensitivity to parameter tuning and collinear predictors in ecological contexts [12,13,35]. This contrasts with non-ecological applications, where eXGB variants often outperform RF [12], underscoring the need for ecosystem-specific algorithm validation.
Notable computational patterns emerged in auxiliary methods: CIT achieved superior precision under fencing (median error = 175.21), while ANN maintained minimum errors in dynamic grazing systems (2.69–6973.16 vs. eXGB’s 16.18–11,751.26, Table A6). These findings partially align with deep learning applications in vegetation–nutrient feedback modeling [14], yet challenge meta-analytical expectations of ANN dominance in ecological prediction [15]. The results corroborate RF’s established strengths in alpine ecosystem analysis, demonstrating consistent performance across soil properties (moisture, pH, N/P availability), plant metrics (biomass, diversity), and forage characteristics, which is critical for holistic environmental monitoring. The structural accuracy relationships identified suggest that mid-complexity models, like RF, offer optimal tradeoffs for ecological decision support systems, balancing computational demand with biological interpretability [36]. Future research should prioritize hybrid architectures combining CIT’s interpretative strengths with ANN’s dynamic adaptation capabilities, while addressing gradient-boosting methods’ generalization limitations through enhanced hyperparameter optimization in ecological datasets [35].
4.2. Random Forest Superiority in Ecosystem Stoichiometry Modeling
RF demonstrates superior accuracy in predicting plant aboveground parts C, N, and P contents in alpine grasslands (Figure 2 and Figure 3), owing to its robust capacity for modeling nonlinear ecological processes. The elemental contents in these ecosystems are shaped by responses to contrasting management regimes, such as grazing-induced P depletion and N accumulation in fenced areas [37,38]. By constructing ensembles of decision trees with recursive partitioning, RF effectively captures these nonlinear dynamics, achieving coefficients of determination (R2) of 0.96–1 for C, N, and P content predictions, significantly outperforming MLR and GBR (Figure 2 and Figure 3, Table A4). For instance, in predicting plant P content, RF reduced the mean square errors by identifying interactive effects between soil available phosphorus and root biomass [39]. These results highlight RF’s ability to integrate multisource environmental variables (e.g., satellite-based NDVI data and climatic factors) and quantify management-driven regulation of elemental contents with high precision.
RF’s algorithmic architecture provides unique advantages for predicting stoichiometric ratios (C:N, C:P, and N:P) in spatially heterogeneous alpine grasslands (Figure 4 and Figure 5). Grazing-induced vegetation patchiness generates high variability in elemental ratios [40], yet RF mitigates overfitting to localized outliers through bootstrap aggregation (bagging) and randomized feature selection [41,42]. Results also show that RF outperforms SVM in terms of overall accuracy [43]. This high precision stems from RF’s capability to detect optimal setting and enhanced robustness (Table A1). Such optimal settings are accurately quantified via multi-tree voting mechanisms (ntree: the number of trees, and mtry: the number of variables for splitting). Consequently, RF enables high-confidence diagnosis of nutrient limitations, providing critical insights for grassland management.
RF’s strengths extend to ecosystem-scale predictions of C/N/P stocks (Figure 6 and Figure 7), where it excels in integrating biomass data with elemental concentrations. Nutrient stock calculations require simultaneous consideration of plant community biomass and its elemental composition. RF addresses this complexity through feature importance analysis, automatically identifying dominant drivers such as aboveground biomass. Validation tests revealed that RF’s predictions for fenced-area C stocks deviated by only 8%, significantly lower than GBR’s 11% error (Figure 6), attributable to its precise modeling of nonlinear relationships [44]. Furthermore, RF’s generalizability across management regimes enables unified predictions of P stock dynamics in fenced and rotationally grazed areas (R2 = 0.85–0.90, Figure 6 and Figure 7), offering a scalable tool for regional nutrient cycling simulations. In summary, RF’s accuracy and stability in nutrient stock predictions establish it as a cornerstone technology for multiscale nutrient management in alpine grassland ecosystems.
The accuracies of the RF across fenced and grazed alpine grasslands were different from those for aboveground C/N/P contents, stoichiometric ratios, and nutrient stocks (Figure 2, Figure 3, Figure 4, Figure 5, Figure 6 and Figure 7), reflecting ecosystem-specific responses to management-induced heterogeneity. In fenced areas, RF achieves exceptional accuracy for C/N/P contents (R2 = 0.97–1.00) due to stable nutrient dynamics and reduced spatial variability [45]. Conversely, under grazing regimes, RF maintains robust, but slightly lower, precision (R2 = 0.96–0.99) by dynamically adapting to herbivore-driven disturbances, such as fecal nitrogen deposition and trampling-induced soil compaction. Because grazing results in higher nitrogen retention, with 9% more nitrogen stocks observed, this indicates improved nutrient cycling in these systems [46]. For stoichiometric ratios, RF’s prediction errors for C:N in grazed areas (mean square errors =29.58) are higher than in fenced zones (7.54, Table A1), attributable to grazing-induced patchiness amplifying ratio variability at sub-meter scales [47]. The precision divergence in nutrient stock predictions between management regimes is quantitatively validated by the RF model parameters (Table A1). Under fencing conditions, RF achieved carbon stock prediction mean square errors of 149.11 with R2 = 0.65 (n = 119). In contrast, grazed phosphorus stock predictions exhibited near-zero mean square errors (0.00, Table A1) but lower explanatory power (R2 = 0.54, n = 123), reflecting wider practical error ranges due to temporal lags in nutrient redistribution. This apparent contradiction between mean square errors and field-measured deviations arises because the model’s temporal feature engineering (e.g., lagged grazing intensity variables with mtry = 2) effectively minimized arithmetic errors while struggling to capture delayed P stock fluctuations across grazing cycles. The differential performance is further evidenced by contrasting tree configurations—fencing predictions required complex ensembles (ntree = 979) to model stabilized nutrient dynamics, whereas grazing utilized simpler architectures (ntree = 891) to accommodate rapid stoichiometric shifts. These results demonstrate that while RF maintains superior precision across regimes, interpretation must consider both statistical metrics (mean square errors/R2) and the ecological reality of management-driven timescales. This precision asymmetry underscores RF’s unique capacity to balance ecological fidelity with algorithmic flexibility, enabling cross-regime comparability while accounting for management-specific drivers—a critical advancement for adaptive grassland restoration strategies.
4.3. Uncertainty Analysis
However, the current modeling framework exhibits uncertainties, primarily arising from ecological data heterogeneity and inherent methodological constraints, which should be systematically addressed in future studies. Key uncertainties stem from spatial and temporal mismatches in field data collection—for instance, grazing-induced nutrient redistribution occurs at sub-meter scales, yet sampling density limitations may inadequately capture fine-grained stoichiometric variability, potentially leading to bias in C:N ratio predictions in dynamic grazing systems. Additionally, measurement errors in aboveground biomass estimation (9–15% via NDVImax) propagate through elemental stock calculations, partially explaining the divergence between RF’s near-zero validation mean square errors and field-observed P stock fluctuations. Model generalizability across ecosystem types remains untested, as the training data exclusively represent alpine grasslands, limiting extrapolation to lowland or degraded grasslands with distinct nutrient cycling regimes. Furthermore, the exclusion of microbial mediation parameters (e.g., extracellular enzyme activities) and transient climatic variables (e.g., diurnal soil moisture shifts) introduces structural uncertainties in simulating threshold-driven stoichiometric responses, particularly under fencing scenarios where biotic interactions dominate nutrient dynamics. Future research should prioritize the following areas: (1) high-resolution spatiotemporal sampling to resolve scale mismatches, employing drone-based hyperspectral imaging and IoT soil sensors for continuous nutrient flux monitoring. (2) The integration of process-based biogeochemical models with machine learning architectures to constrain biologically implausible predictions (e.g., negative P stocks) while retaining RF’s nonlinear fitting advantages. (3) Cross-ecosystem validation using standardized protocols to assess model transferability beyond alpine grasslands. (4) Dynamic hyperparameter optimization frameworks tailored for ecological time-series data, addressing temporal autocorrelation in grazing-mediated nutrient pulses. (5) Mechanistic interpretability enhancements through SHAP value analysis coupled with stable isotope tracing, explicitly linking RF’s feature importance rankings to verified ecological pathways.
5. Conclusions
In this study, nine computational models were employed to assess and validate the characteristics (carbon, nitrogen, and phosphorus concentrations), elemental ratios (C:N, C:P, and N:P), and elemental pools (carbon, nitrogen, and phosphorus pools) in aboveground vegetation under both fenced and grazing conditions in Xizang’s grasslands. Among the nine models, the random forest algorithm was more accurate in prediction and modeling these biogeochemical parameters. Furthermore, the random forest model demonstrated the highest predictive accuracy for nitrogen content, C:P, and N:P ratios under both management regimes. It also achieved the highest accuracy in predicting the C:N ratio under fencing conditions, while under grazing conditions, it excelled in modeling the carbon pool, phosphorus content, and phosphorus pool. This research consequently establishes a novel framework for analyzing spatial patterns of grassland carbon, nitrogen, and phosphorus characteristics—including contents, elemental ratios, and pools—in aboveground vegetation across the Qinghai–Xizang Plateau, with potential applications for global-scale ecological studies.
Author Contributions
Conception and design: M.X. and G.F; acquisition, analysis, and interpretation of data: M.X., G.F., J.C., T.M., Y.M., K.Z. and Z.W; drafting the work: M.X. and G.F; revising the work critically for important intellectual content: M.X., G.F., J.C., T.M., Y.M., K.Z. and Z.W; funding acquisition: M.X., G.F. and Z.W. All authors have read and approved the final version of the manuscript and agree to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. All authors have read and agreed to the published version of the manuscript.
Funding
The study was financially supported by the Chief Scientist Program of Qinghai Province (Grant No. 2024-SF-102), the Open Project of State Key Laboratory of Plateau Ecology and Agriculture, Qinghai University [2025-ZZ-01], the Lhasa Science and Technology Plan Project [LSKJ202422], the Tibet Autonomous Region Science and Technology Project [XZ202401JD0029, XZ202501ZY0086, XZ202501ZY0056], and Construction of Zhongba County Fixed Observation and Experiment Station of First Support System for Agriculture Green Development.
Data Availability Statement
The data that support the findings of this study are available from the corresponding author upon request.
Conflicts of Interest
The authors declare no conflicts of interest.
Appendix A
Figure A1.
Sampling sites.
Table A1.
Random forest model parameters of carbon content, nitrogen content, phosphorus content, ratio of carbon to nitrogen (C:N), ratio of carbon to phosphorus (C:P), ratio of nitrogen to phosphorus (N:P), carbon pool, nitrogen pool, and phosphorus pool of aboveground parts of plant community under fencing and free-grazing conditions.
Table A1.
Random forest model parameters of carbon content, nitrogen content, phosphorus content, ratio of carbon to nitrogen (C:N), ratio of carbon to phosphorus (C:P), ratio of nitrogen to phosphorus (N:P), carbon pool, nitrogen pool, and phosphorus pool of aboveground parts of plant community under fencing and free-grazing conditions.
| Conditions | Variable | Mean Square Errors | ntree | mtry | R2 | n |
|---|---|---|---|---|---|---|
| Fencing | Carbon content | 5.56 | 833 | 1 | 0.53 | 251 |
| Nitrogen content | 0.09 | 611 | 1 | 0.67 | 261 | |
| Phosphorus content | 0.03 | 902 | 1 | 0.91 | 229 | |
| C:N | 7.54 | 922 | 2 | 0.84 | 239 | |
| C:P | 3590.68 | 703 | 1 | 0.81 | 227 | |
| N:P | 6.50 | 543 | 1 | 0.74 | 229 | |
| Carbon pool | 149.11 | 979 | 2 | 0.66 | 119 | |
| Nitrogen pool | 0.27 | 880 | 1 | 0.55 | 127 | |
| Phosphorus pool | 0.00 | 891 | 2 | 0.54 | 123 | |
| Grazing | Carbon content | 15.79 | 900 | 4 | 0.63 | 273 |
| Nitrogen content | 0.13 | 840 | 1 | 0.61 | 273 | |
| Phosphorus content | 0.02 | 940 | 1 | 0.89 | 255 | |
| C:N | 29.58 | 943 | 1 | 0.70 | 273 | |
| C:P | 4508.93 | 931 | 4 | 0.71 | 250 | |
| N:P | 5.72 | 933 | 1 | 0.62 | 250 | |
| Carbon pool | 137.86 | 949 | 4 | 0.62 | 214 | |
| Nitrogen pool | 0.14 | 910 | 1 | 0.65 | 219 | |
| Phosphorus pool | 0.00 | 901 | 1 | 0.69 | 216 |
Table A2.
Generalized boosted regression parameters of carbon content, nitrogen content, phosphorus content, ratio of carbon to nitrogen (C:N), ratio of carbon to phosphorus (C:P), ratio of nitrogen to phosphorus (N:P), carbon pool, nitrogen pool, and phosphorus pool of aboveground parts of plant community under fencing and free-grazing conditions.
Table A2.
Generalized boosted regression parameters of carbon content, nitrogen content, phosphorus content, ratio of carbon to nitrogen (C:N), ratio of carbon to phosphorus (C:P), ratio of nitrogen to phosphorus (N:P), carbon pool, nitrogen pool, and phosphorus pool of aboveground parts of plant community under fencing and free-grazing conditions.
| Conditions | Variable | Tree Nos | Mean Train Error | Mean Cv Error | n |
|---|---|---|---|---|---|
| Fencing | Carbon content | 1000 | 4.59 | 6.89 | 251 |
| Nitrogen content | 1000 | 0.09 | 0.11 | 261 | |
| Phosphorus content | 1000 | 0.04 | 0.11 | 229 | |
| C:N | 1000 | 9.32 | 10.99 | 239 | |
| C:P | 1000 | 3123.08 | 4911.09 | 227 | |
| N:P | 1000 | 6.15 | 9.52 | 229 | |
| Carbon pool | 998 | 169.39 | 385.13 | 119 | |
| Nitrogen pool | 1000 | 0.24 | 0.45 | 127 | |
| Phosphorus pool | 1000 | 0.00 | 0.00 | 123 | |
| Grazing | Carbon content | 1000 | 13.79 | 28.67 | 273 |
| Nitrogen content | 1000 | 0.12 | 0.24 | 273 | |
| Phosphorus content | 1000 | 0.03 | 0.05 | 255 | |
| C:N | 1000 | 28.66 | 67.02 | 273 | |
| C:P | 1000 | 5568.85 | 10,083.84 | 250 | |
| N:P | 1000 | 5.30 | 8.78 | 250 | |
| Carbon pool | 997 | 130.11 | 309.33 | 214 | |
| Nitrogen pool | 1000 | 0.14 | 0.25 | 219 | |
| Phosphorus pool | 1000 | 0.00 | 0.00 | 216 |
Table A3.
Support vector machine parameters of carbon content, nitrogen content, phosphorus content, ratio of carbon to nitrogen (C:N), ratio of carbon to phosphorus (C:P), ratio of nitrogen to phosphorus (N:P), carbon pool, nitrogen pool, and phosphorus pool of aboveground parts of plant community under fencing and free-grazing conditions.
Table A3.
Support vector machine parameters of carbon content, nitrogen content, phosphorus content, ratio of carbon to nitrogen (C:N), ratio of carbon to phosphorus (C:P), ratio of nitrogen to phosphorus (N:P), carbon pool, nitrogen pool, and phosphorus pool of aboveground parts of plant community under fencing and free-grazing conditions.
| Conditions | Variable | Mean Residuals | Mean Decision Values | Support Vector Nos | n |
|---|---|---|---|---|---|
| Fencing | Carbon content | 0.13 | −0.04 | 208 | 251 |
| Nitrogen content | 0.07 | −0.13 | 201 | 261 | |
| Phosphorus content | 0.02 | −0.04 | 114 | 229 | |
| C:N | −0.07 | 0.01 | 179 | 239 | |
| C:P | 5.60 | −0.04 | 142 | 227 | |
| N:P | 0.49 | −0.01 | 157 | 229 | |
| Carbon pool | 3.36 | −0.16 | 101 | 119 | |
| Nitrogen pool | 0.14 | −0.18 | 106 | 127 | |
| Phosphorus pool | 0.01 | −0.17 | 105 | 123 | |
| Grazing | Carbon content | −0.41 | 0.06 | 219 | 273 |
| Nitrogen content | 0.05 | −0.08 | 222 | 273 | |
| Phosphorus content | 0.05 | −0.12 | 102 | 255 | |
| C:N | 1.17 | −0.12 | 203 | 273 | |
| C:P | 8.67 | −0.07 | 200 | 250 | |
| N:P | 0.33 | −0.08 | 207 | 250 | |
| Carbon pool | 4.03 | −0.21 | 170 | 214 | |
| Nitrogen pool | 0.11 | −0.18 | 188 | 219 | |
| Phosphorus pool | 0.01 | −0.18 | 178 | 216 |
Table A4.
Multiple linear regression parameters of carbon content, nitrogen content, phosphorus content, ratio of carbon to nitrogen (C:N), ratio of carbon to phosphorus (C:P), ratio of nitrogen to phosphorus (N:P), carbon pool, nitrogen pool, and phosphorus pool of aboveground parts of plant community under fencing and free-grazing conditions.
Table A4.
Multiple linear regression parameters of carbon content, nitrogen content, phosphorus content, ratio of carbon to nitrogen (C:N), ratio of carbon to phosphorus (C:P), ratio of nitrogen to phosphorus (N:P), carbon pool, nitrogen pool, and phosphorus pool of aboveground parts of plant community under fencing and free-grazing conditions.
| Conditions | Variable | Intercept | Temperature | Precipitation | Radiation | NDVI | R2 | n |
|---|---|---|---|---|---|---|---|---|
| Fencing | Carbon content | 42.07 | −0.12 | 0.009 | −0.0009 | 0.23 | 251 | |
| Nitrogen content | 8.27 | −0.06 | 0.0005 | −0.0009 | 0.41 | 261 | ||
| Phosphorus content | 3.25 | −0.02 | 0.0016 | −0.0005 | 0.39 | 229 | ||
| C:N | −47.54 | 0.38 | −0.01 | 0.01 | 0.53 | 239 | ||
| C:P | −2367.43 | −6.63 | −0.02 | 0.35 | 0.44 | 227 | ||
| N:P | −14.07 | 0.07 | −0.0024 | 0.0031 | 0.07 | 229 | ||
| Carbon pool | −130.44 | 2.66 | 0.02 | 0.02 | 0.08 | 119 | ||
| Nitrogen pool | −5.61 | 0.09 | 0.0004 | 0.0008 | 0.07 | 127 | ||
| Phosphorus pool | −0.39 | 0.01 | 0.0001 | 0.0001 | 0.06 | 123 | ||
| Grazing | Carbon content | −15.06 | 0.59 | 0.0007 | 0.01 | 14.10 | 0.27 | 273 |
| Nitrogen content | 3.64 | −0.02 | 0.0004 | −0.0003 | 1.50 | 0.27 | 273 | |
| Phosphorus content | 1.45 | −0.002 | −0.0003 | −0.0002 | 1.22 | 0.27 | 255 | |
| C:N | −20.86 | 1.13 | 0.0004 | 0.0067 | −7.92 | 0.14 | 273 | |
| C:P | −127.14 | 11.47 | 0.04 | 0.06 | −181.02 | 0.12 | 250 | |
| N:P | 12.74 | 0.04 | 0.004 | −0.0002 | −11.36 | 0.20 | 250 | |
| Carbon pool | −107.15 | −0.07 | 0.01 | 0.02 | 39.55 | 0.22 | 214 | |
| Nitrogen pool | −1.79 | −0.01 | 0.0001 | 0.0003 | 1.96 | 0.25 | 219 | |
| Phosphorus pool | −0.31 | −0.001 | 0.00002 | 0.00004 | 0.18 | 0.30 | 216 |
Table A5.
Recursive regression tree parameters of carbon content, nitrogen content, phosphorus content, ratio of carbon to nitrogen (C:N), ratio of carbon to phosphorus (C:P), ratio of nitrogen to phosphorus (N:P), carbon pool, nitrogen pool, and phosphorus pool of aboveground parts of plant community under fencing and free-grazing conditions.
Table A5.
Recursive regression tree parameters of carbon content, nitrogen content, phosphorus content, ratio of carbon to nitrogen (C:N), ratio of carbon to phosphorus (C:P), ratio of nitrogen to phosphorus (N:P), carbon pool, nitrogen pool, and phosphorus pool of aboveground parts of plant community under fencing and free-grazing conditions.
| Conditions | Variable | R2 | n |
|---|---|---|---|
| Fencing | Carbon content | 0.58 | 251 |
| Nitrogen content | 0.62 | 261 | |
| Phosphorus content | 0.86 | 229 | |
| C:N | 0.77 | 239 | |
| C:P | 0.78 | 227 | |
| N:P | 0.71 | 229 | |
| Carbon pool | 0.34 | 119 | |
| Nitrogen pool | 0.34 | 127 | |
| Phosphorus pool | 0.29 | 123 | |
| Grazing | Carbon content | 0.53 | 273 |
| Nitrogen content | 0.53 | 273 | |
| Phosphorus content | 0.59 | 255 | |
| C:N | 0.39 | 273 | |
| C:P | 0.34 | 250 | |
| N:P | 0.50 | 250 | |
| Carbon pool | 0.54 | 214 | |
| Nitrogen pool | 0.57 | 219 | |
| Phosphorus pool | 0.51 | 216 |
Table A6.
Artificial neural network (ANN), generalized linear regression (GLR), conditional inference tree (CIT) and extreme gradient boosting (eXGB) parameters of carbon content, nitrogen content, phosphorus content, ratio of carbon to nitrogen (C:N), ratio of carbon to phosphorus (C:P), ratio of nitrogen to phosphorus (N:P), carbon pool, nitrogen pool, and phosphorus pool of aboveground parts of plant community under fencing and free-grazing conditions.
Table A6.
Artificial neural network (ANN), generalized linear regression (GLR), conditional inference tree (CIT) and extreme gradient boosting (eXGB) parameters of carbon content, nitrogen content, phosphorus content, ratio of carbon to nitrogen (C:N), ratio of carbon to phosphorus (C:P), ratio of nitrogen to phosphorus (N:P), carbon pool, nitrogen pool, and phosphorus pool of aboveground parts of plant community under fencing and free-grazing conditions.
| Conditions | Variable | ANN | GLR | CIT | eXGB | n |
|---|---|---|---|---|---|---|
| Fencing | Carbon content | 203.72 | 221.95 | 198.93 | 1590.91 | 251 |
| Nitrogen content | 28.86 | 32.47 | 23.44 | 48.42 | 261 | |
| Phosphorus content | 15.04 | 33.03 | 7.18 | 18.64 | 229 | |
| C:N | 294.44 | 352.40 | 245.99 | 1080.94 | 239 | |
| C:P | 6820.52 | 6938.37 | 3865.50 | 9438.07 | 227 | |
| N:P | 267.86 | 290.13 | 175.21 | 333.44 | 229 | |
| Carbon pool | 563.07 | 563.07 | 563.07 | 524.04 | 119 | |
| Nitrogen pool | 21.64 | 21.64 | 21.64 | 17.60 | 127 | |
| Phosphorus pool | 2.23 | 2.23 | 2.23 | 7.52 | 123 | |
| Grazing | Carbon content | 409.92 | 421.45 | 362.57 | 1622.17 | 273 |
| Nitrogen content | 37.01 | 38.79 | 38.25 | 53.45 | 273 | |
| Phosphorus content | 13.43 | 15.43 | 8.81 | 18.26 | 255 | |
| C:N | 587.67 | 705.95 | 588.99 | 1282.86 | 273 | |
| C:P | 6973.16 | 7312.38 | 7422.90 | 11,751.26 | 250 | |
| N:P | 224.70 | 226.60 | 233.33 | 372.42 | 250 | |
| Carbon pool | 868.04 | 1033.08 | 834.29 | 812.21 | 214 | |
| Nitrogen pool | 34.97 | 36.72 | 34.88 | 30.64 | 219 | |
| Phosphorus pool | 2.69 | 2.98 | 2.63 | 16.18 | 216 |
References
- Elser, J.J.; Sterner, R.W.; Gorokhova, E.; Fagan, W.F.; Markow, T.A.; Cotner, J.B.; Harrison, J.F.; Hobbie, S.E.; Odell, G.M.; Weider, L. Biological stoichiometry from genes to ecosystems. Ecol. Lett. 2000, 3, 540–550. [Google Scholar] [CrossRef]
- Reich, P.B.; Grigal, D.F.; Aber, J.D.; Gower, S.T. Nitrogen mineralization and productivity in 50 hardwood and conifer stands on diverse soils. Ecology 1997, 78, 335–347. [Google Scholar] [CrossRef]
- Laliberte, E.; Shipley, B.; Norton, D.A.; Scott, D. Which plant traits determine abundance under long-term shifts in soil resource availability and grazing intensity? J. Ecol. 2012, 100, 662–677. [Google Scholar] [CrossRef]
- Feng, W.-L.; Yang, J.-L.; Xu, L.-G.; Zhang, G.-L. The spatial variations and driving factors of C, N, P stoichiometric characteristics of plant and soil in the terrestrial ecosystem. Sci. Total Environ. 2024, 951, 175543. [Google Scholar] [CrossRef]
- Dibar, D.T.; Zhang, K.; Yuan, S.; Zhang, J.; Zhou, Z.; Ye, X. Ecological stoichiometric characteristics of Carbon (C), Nitrogen (N) and Phosphorus (P) in leaf, root, stem, and soil in four wetland plants communities in Shengjin Lake, China. PLoS ONE 2020, 15, e0230089. [Google Scholar] [CrossRef]
- Yang, Y.; Fang, J.; Ji, C.; Datta, A.; Li, P.; Ma, W.; Mohammat, A.; Shen, H.; Hu, H.; Knapp, B.O.; et al. Stoichiometric shifts in surface soils over broad geographical scales: Evidence from China’s grasslands. Glob. Ecol. Biogeogr. 2014, 23, 947–955. [Google Scholar] [CrossRef]
- Reichstein, M.; Camps-Valls, G.; Stevens, B.; Jung, M.; Denzler, J.; Carvalhais, N.; Prabhat, F. Deep learning and process understanding for data-driven Earth system science. Nature 2019, 566, 195–204. [Google Scholar] [CrossRef]
- Moore, C.E.; Beringer, J.; Donohue, R.J.; Evans, B.; Exbrayat, J.-F.; Hutley, L.B.; Tapper, N.J. Seasonal, interannual and decadal drivers of tree and grass productivity in an Australian tropical savanna. Glob. Change Biol. 2018, 24, 2530–2544. [Google Scholar] [CrossRef]
- Barrena-González, J.; Antoneli, V.; Contador, J.F.L.; Fernández, M.P. Assessing how Grazing Intensity Affects the Spatial Distribution of Soil Properties. Earth Syst. Environ. 2024. [Google Scholar] [CrossRef]
- Schimel, D.; Pavlick, R.; Fisher, J.B.; Asner, G.P.; Saatchi, S.; Townsend, P.; Miller, C.; Frankenberg, C.; Hibbard, K.; Cox, P. Observing terrestrial ecosystems and the carbon cycle from space. Glob. Chang. Biol. 2015, 21, 1762–1776. [Google Scholar] [CrossRef]
- Klein, J.A.; Harte, J.; Zhao, X.Q. Experimental warming, not grazing, decreases rangeland quality on the Tibetan Plateau. Ecol. Appl. 2007, 17, 541–557. [Google Scholar] [CrossRef] [PubMed]
- He, W.; Yao, Y.; Li, Y.; Yu, L.; Ni, L.; Fu, B.; Huang, J.; Li, D. Accurately estimating and revealing the patterns of leaf C:N:P stoichiometry with remote sensing and neural network methods in a karst area. Comput. Electron. Agric. 2025, 231, 110006. [Google Scholar] [CrossRef]
- Gao, J.; Liang, T.; Liu, J.; Yin, J.; Ge, J.; Hou, M.; Feng, Q.; Wu, C.; Xie, H. Potential of hyperspectral data and machine learning algorithms to estimate the forage carbon-nitrogen ratio in an alpine grassland ecosystem of the Tibetan Plateau. ISPRS J. Photogramm. Remote Sens. 2020, 163, 362–374. [Google Scholar] [CrossRef]
- Karpatne, A.; Atluri, G.; Faghmous, J.H.; Steinbach, M.; Banerjee, A.; Ganguly, A.; Shekhar, S.; Samatova, N.; Kumar, V. Theory-Guided Data Science: A New Paradigm for Scientific Discovery from Data. IEEE Trans. Knowl. Data Eng. 2017, 29, 2318–2331. [Google Scholar] [CrossRef]
- Liu, Z.; Peng, C.; Xiang, W.; Tian, D.; Deng, X.; Zhao, M. Application of artificial neural networks in global climate change and ecological research: An overview. Chin. Sci. Bull. 2010, 55, 3853–3863. [Google Scholar] [CrossRef]
- Zhang, G.; Fu, G. Changes in soil organic carbon, total nitrogen and total phosphorus in 2000–2020 and their driving mechanisms in Tibetan alpine grasslands. Glob. Planet. Change 2024, 239, 104484. [Google Scholar] [CrossRef]
- Fu, G.; Shen, Z.-X.; Zhang, X.-Z. Increased precipitation has stronger effects on plant production of an alpine meadow than does experimental warming in the Northern Tibetan Plateau. Agric. For. Meteorol. 2018, 249, 11–21. [Google Scholar] [CrossRef]
- Gao, J.; Liang, T.; Zhang, D.; Liu, J.; Feng, Q.; Wu, C.; Wang, Z.; Zhang, X. Hyperspectral remote sensing of forage stoichiometric ratios in the senescent stage of alpine grasslands. Field Crops Res. 2024, 313, 109427. [Google Scholar] [CrossRef]
- Ding, J.; Yang, T.; Zhao, Y.; Liu, D.; Wang, X.; Yao, Y.; Peng, S.; Wang, T.; Piao, S. Increasingly important role of atmospheric aridity on Tibetan alpine grasslands. Geophys. Res. Lett. 2018, 45, 2852–2859. [Google Scholar] [CrossRef]
- Ran, Q.; Hao, Y.; Xia, A.; Liu, W.; Hu, R.; Cui, X.; Xue, K.; Song, X.; Xu, C.; Ding, B.; et al. Quantitative Assessment of the Impact of Physical and Anthropogenic Factors on Vegetation Spatial-Temporal Variation in Northern Tibet. Remote Sens. 2019, 11, 1183. [Google Scholar] [CrossRef]
- Schickhoff, U.; Bobrowski, M.; Mal, S.; Schwab, N.; Singh, R. The world’s mountains in the Anthropocene. Mt. Landsc. Transit. Eff. Land Use Clim. Change 2022, 1–144. [Google Scholar] [CrossRef]
- Yang, J.; Dong, J.; Xiao, X.; Dai, J.; Wu, C.; Xia, J.; Zhao, G.; Zhao, M.; Li, Z.; Zhang, Y.; et al. Divergent shifts in peak photosynthesis timing of temperate and alpine grasslands in China. Remote Sens. Environ. 2019, 233, 111395. [Google Scholar] [CrossRef]
- R Core Team. A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria. Available online: https://www.Rproject.org/ (accessed on 24 April 2024).
- Garge, N.R.; Bobashev, G.; Eggleston, B. Random forest methodology for model-based recursive partitioning: The mobForest package for R. BMC Bioinform. 2013, 14, 125. [Google Scholar] [CrossRef] [PubMed]
- Greenwell, B.; Boehmke, B.; Cunningham, J.; Developers, G.; Greenwell, M.B.J.R.p.v. Package ‘gbm’. Available online: https://cran.r-project.org/web/packages/gbm/index.html (accessed on 28 June 2024).
- Meyer, D.; Dimitriadou, E.; Hornik, K.; Weingessel, A.; Leisch, F.; Chang, C.-C.; Lin, C.-C.; Meyer, M.D.J.T.R.J. Package ‘e1071’. Available online: http://r.meteo.uni.wroc.pl/web/packages/e1071/e1071.pdf (accessed on 16 September 2021).
- Therneau, T.; Atkinson, B.; Ripley, B.; Ripley, M.B.J.A.o.c.m.i.a.u.w.p.r.r.p. Package ‘Rpart’. Available online: https://cran.r-project.org/web/packages/rpart/index.html (accessed on 7 January 2025).
- Cortez, P.; Cortez, M.P. Package ‘rminer’. Available online: https://cran.r-project.org/package=rminer (accessed on 4 June 2025).
- Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar] [CrossRef]
- Cutler, D.R.; Edwards Jr, T.C.; Beard, K.H.; Cutler, A.; Hess, K.T.; Gibson, J.; Lawler, J.J. Random forests for classification in ecology. Ecology 2007, 88, 2783–2792. [Google Scholar] [CrossRef]
- Probst, P.; Boulesteix, A.-L. To tune or not to tune the number of trees in random forest. J. Mach. Learn. Res. 2018, 18, 6673–6690. [Google Scholar]
- Díaz-Uriarte, R.; Alvarez de Andrés, S. Gene selection and classification of microarray data using random forest. BMC Bioinform. 2006, 7, 3. [Google Scholar] [CrossRef]
- Sterner, R.W.; Elser, J.J. Ecological Stoichiometry; Princeton University Press: Princeton, NJ, USA, 2003. [Google Scholar]
- Bloor, J.M.G.; Tardif, A.; Pottier, J. Spatial Heterogeneity of Vegetation Structure, Plant N Pools and Soil N Content in Relation to Grassland Management. Agronomy 2020, 10, 716. [Google Scholar] [CrossRef]
- Zhu, J.-J.; Yang, M.; Ren, Z.J. Machine Learning in Environmental Research: Common Pitfalls and Best Practices. Environ. Sci. Technol. 2023, 57, 17671–17689. [Google Scholar] [CrossRef]
- Tao, W.-K.; Moncrieff, M.W. Multiscale cloud system modeling. Rev. Geophys. 2009, 47, 1–41. [Google Scholar] [CrossRef]
- Craine, J.M.; Ballantyne, F.; Peel, M.; Zambatis, N.; Morrow, C.; Stock, W.D. Grazing and landscape controls on nitrogen availability across 330 South African savanna sites. Austral Ecol. 2009, 34, 731–740. [Google Scholar] [CrossRef]
- Riesch, F.; Wichelhaus, A.; Tonn, B.; Meißner, M.; Rosenthal, G.; Isselstein, J. Grazing by wild red deer can mitigate nutrient enrichment in protected semi-natural open habitats. Oecologia 2022, 199, 471–485. [Google Scholar] [CrossRef] [PubMed]
- Gao, J.; Meng, B.; Liang, T.; Feng, Q.; Ge, J.; Yin, J.; Wu, C.; Cui, X.; Hou, M.; Liu, J. Modeling alpine grassland forage phosphorus based on hyperspectral remote sensing and a multi-factor machine learning algorithm in the east of Tibetan Plateau, China. ISPRS J. Photogramm. Remote Sens. 2019, 147, 104–117. [Google Scholar] [CrossRef]
- Tonn, B.; Densing, E.M.; Gabler, J.; Isselstein, J. Grazing-induced patchiness, not grazing intensity, drives plant diversity in European low-input pastures. J. Appl. Ecol. 2019, 56, 1624–1636. [Google Scholar] [CrossRef]
- Rodriguez-Galiano, V.; Sanchez-Castillo, M.; Chica-Olmo, M.; Chica-Rivas, M. Machine learning predictive models for mineral prospectivity: An evaluation of neural networks, random forest, regression trees and support vector machines. Ore Geol. Rev. 2015, 71, 804–818. [Google Scholar] [CrossRef]
- Wang, L.a.; Zhou, X.; Zhu, X.; Dong, Z.; Guo, W. Estimation of biomass in wheat using random forest regression algorithm and remote sensing data. Crop J. 2016, 4, 212–219. [Google Scholar] [CrossRef]
- Avcı, C.; Budak, M.; Yağmur, N.; Balçık, F. Comparison between random forest and support vector machine algorithms for LULC classification. Int. J. Eng. Geosci. 2023, 8, 1–10. [Google Scholar] [CrossRef]
- Wu, Z.J.; Yao, F.M.; Zhang, J.H.; Liu, H.Y. Estimating Forest Aboveground Biomass Using a Combination of Geographical Random Forest and Empirical Bayesian Kriging Models. Remote Sens. 2024, 16, 1859. [Google Scholar] [CrossRef]
- Daufresne, T.; Loreau, M. Plant-herbivore interactions and ecological stoichiometry: When do herbivores determine plant nutrient limitation? Ecol. Lett. 2001, 4, 196–206. [Google Scholar] [CrossRef]
- Mosier, S.; Apfelbaum, S.; Byck, P.; Calderon, F.; Teague, R.; Thompson, R.; Cotrufo, M.F. Adaptive multi-paddock grazing enhances soil carbon and nitrogen stocks and stabilization through mineral association in southeastern US grazing lands. J. Environ. Manag. 2021, 288, 112409. [Google Scholar] [CrossRef]
- Wang, F.; Apfelbaum, S.I.; Thompson, R.L.; Teague, R.; Byck, P. Effects of adaptive multiple paddock and continuous grazing on fine-scale spatial patterns of vegetation species and biomass in commercial ranches. Landsc. Ecol. 2021, 36, 2725–2741. [Google Scholar] [CrossRef]
Figure 1.
Integrated research framework for alpine grassland stoichiometry dynamics.
Figure 2.
Comparison of simulated and observed carbon, nitrogen, and phosphorus contents of aboveground parts of plants for (a–c) RF, (d–f) GBR, (g–i) SVM, (j–l) MLR, (m–o) RRT, (p–r) GLR, (s–u) ANN, (v–x) CIT, and (y–aa) eXGB under fencing conditions. Each red circle represents a single paired observation of observed and simulated values for C, N, or P content in the samples analyzed. The solid lines are the linear regression between the estimated and observed values. RF, random forest; GBR, generalized boosted regression; MLR, multiple linear regression; SVM, support vector machine; RRT, recursive regression tree; ANN, artificial neural network; GLR, generalized linear regression; CIT, conditional inference tree; eXGB, extreme gradient boosting.
Figure 2.
Comparison of simulated and observed carbon, nitrogen, and phosphorus contents of aboveground parts of plants for (a–c) RF, (d–f) GBR, (g–i) SVM, (j–l) MLR, (m–o) RRT, (p–r) GLR, (s–u) ANN, (v–x) CIT, and (y–aa) eXGB under fencing conditions. Each red circle represents a single paired observation of observed and simulated values for C, N, or P content in the samples analyzed. The solid lines are the linear regression between the estimated and observed values. RF, random forest; GBR, generalized boosted regression; MLR, multiple linear regression; SVM, support vector machine; RRT, recursive regression tree; ANN, artificial neural network; GLR, generalized linear regression; CIT, conditional inference tree; eXGB, extreme gradient boosting.
Figure 3.
Comparison of simulated and observed carbon, nitrogen, and phosphorus contents of aboveground parts of plants for (a–c) RF, (d–f) GBR, (g–i) SVM, (j–l) MLR, (m–o) RRT, (p–r) GLR, (s–u) ANN, (v–x) CIT, and (y–aa) eXGB under grazing conditions. Each red circle represents a single paired observation of observed and simulated values for C, N, or P content in the samples analyzed. The solid lines are the linear regression between the estimated and observed values. RF, random forest; GBR, generalized boosted regression; MLR, multiple linear regression; SVM, support vector machine; RRT, recursive regression tree; ANN, artificial neural network; GLR, generalized linear regression; CIT, conditional inference tree; eXGB, extreme gradient boosting.
Figure 3.
Comparison of simulated and observed carbon, nitrogen, and phosphorus contents of aboveground parts of plants for (a–c) RF, (d–f) GBR, (g–i) SVM, (j–l) MLR, (m–o) RRT, (p–r) GLR, (s–u) ANN, (v–x) CIT, and (y–aa) eXGB under grazing conditions. Each red circle represents a single paired observation of observed and simulated values for C, N, or P content in the samples analyzed. The solid lines are the linear regression between the estimated and observed values. RF, random forest; GBR, generalized boosted regression; MLR, multiple linear regression; SVM, support vector machine; RRT, recursive regression tree; ANN, artificial neural network; GLR, generalized linear regression; CIT, conditional inference tree; eXGB, extreme gradient boosting.
Figure 4.
Comparison of simulated and observed ratio of carbon to nitrogen, ratio of carbon to phosphorus, and ratio of nitrogen to phosphorus in aboveground parts of plants for (a–c) RF, (d–f) GBR, (g–i) SVM, (j–l) MLR, (m–o) RRT, (p–r) GLR, (s–u) ANN, (v–x) CIT, and (y–aa) eXGB under fencing conditions. Each red circle represents a single paired observation of observed and simulated values for ratio of carbon to nitrogen, ratio of carbon to phosphorus, and ratio of nitrogen to phosphorus in the samples analyzed. The solid lines are the linear regression between the estimated and observed values. RF, random forest; GBR, generalized boosted regression; MLR, multiple linear regression; SVM, support vector machine; RRT, recursive regression tree; ANN, artificial neural network; GLR, generalized linear regression; CIT, conditional inference tree; eXGB, extreme gradient boosting.
Figure 4.
Comparison of simulated and observed ratio of carbon to nitrogen, ratio of carbon to phosphorus, and ratio of nitrogen to phosphorus in aboveground parts of plants for (a–c) RF, (d–f) GBR, (g–i) SVM, (j–l) MLR, (m–o) RRT, (p–r) GLR, (s–u) ANN, (v–x) CIT, and (y–aa) eXGB under fencing conditions. Each red circle represents a single paired observation of observed and simulated values for ratio of carbon to nitrogen, ratio of carbon to phosphorus, and ratio of nitrogen to phosphorus in the samples analyzed. The solid lines are the linear regression between the estimated and observed values. RF, random forest; GBR, generalized boosted regression; MLR, multiple linear regression; SVM, support vector machine; RRT, recursive regression tree; ANN, artificial neural network; GLR, generalized linear regression; CIT, conditional inference tree; eXGB, extreme gradient boosting.
Figure 5.
Comparison of simulated and observed ratio of carbon to nitrogen, ratio of carbon to phosphorus, and ratio of nitrogen to phosphorus in aboveground parts of plants for (a–c) RF, (d–f) GBR, (g–i) SVM, (j–l) MLR, (m–o) RRT, (p–r) GLR, (s–u) ANN, (v–x) CIT, and (y–aa) eXGB under grazing conditions. Each red circle represents a single paired observation of observed and simulated values for ratio of carbon to nitrogen, ratio of carbon to phosphorus, and ratio of nitrogen to phosphorus in the samples analyzed. The solid lines are the linear regression between the estimated and observed values. RF, random forest; GBR, generalized boosted regression; MLR, multiple linear regression; SVM, support vector machine; RRT, recursive regression tree; ANN, artificial neural network; GLR, generalized linear regression; CIT, conditional inference tree; eXGB, extreme gradient boosting.
Figure 5.
Comparison of simulated and observed ratio of carbon to nitrogen, ratio of carbon to phosphorus, and ratio of nitrogen to phosphorus in aboveground parts of plants for (a–c) RF, (d–f) GBR, (g–i) SVM, (j–l) MLR, (m–o) RRT, (p–r) GLR, (s–u) ANN, (v–x) CIT, and (y–aa) eXGB under grazing conditions. Each red circle represents a single paired observation of observed and simulated values for ratio of carbon to nitrogen, ratio of carbon to phosphorus, and ratio of nitrogen to phosphorus in the samples analyzed. The solid lines are the linear regression between the estimated and observed values. RF, random forest; GBR, generalized boosted regression; MLR, multiple linear regression; SVM, support vector machine; RRT, recursive regression tree; ANN, artificial neural network; GLR, generalized linear regression; CIT, conditional inference tree; eXGB, extreme gradient boosting.
Figure 6.
Comparison of simulated and observed carbon, nitrogen, and phosphorus pools of aboveground parts of plants for (a–c) RF, (d–f) GBR, (g–i) SVM, (j–l) MLR, (m–o) RRT, (p–r) GLR, (s–u) ANN, (v–x) CIT, and (y–aa) eXGB under fencing conditions. Each red circle represents a single paired observation of observed and simulated values for C, N, or P pool in the samples analyzed. The solid lines are the linear regression be-tween the estimated and observed values. RF, random forest; GBR, generalized boosted regression; MLR, multiple linear regression; SVM, support vector machine; RRT, recursive regression tree; ANN, artificial neural network; GLR, generalized linear regression; CIT, conditional inference tree; eXGB, extreme gradient boosting.
Figure 6.
Comparison of simulated and observed carbon, nitrogen, and phosphorus pools of aboveground parts of plants for (a–c) RF, (d–f) GBR, (g–i) SVM, (j–l) MLR, (m–o) RRT, (p–r) GLR, (s–u) ANN, (v–x) CIT, and (y–aa) eXGB under fencing conditions. Each red circle represents a single paired observation of observed and simulated values for C, N, or P pool in the samples analyzed. The solid lines are the linear regression be-tween the estimated and observed values. RF, random forest; GBR, generalized boosted regression; MLR, multiple linear regression; SVM, support vector machine; RRT, recursive regression tree; ANN, artificial neural network; GLR, generalized linear regression; CIT, conditional inference tree; eXGB, extreme gradient boosting.
Figure 7.
Comparison of simulated and observed carbon, nitrogen, and phosphorus pools of aboveground parts of plants for (a–c) RF, (d–f) GBR, (g–i) SVM, (j–l) MLR, (m–o) RRT, (p–r) GLR, (s–u) ANN, (v–x) CIT, and (y–aa) eXGB under grazing conditions. Each red circle represents a single paired observation of observed and simulated values for C, N, or P pool in the samples analyzed. The solid lines are the linear regression be-tween the estimated and observed values. RF, random forest; GBR, generalized boosted regression; MLR, multiple linear regression; SVM, support vector machine; RRT, recursive regression tree; ANN, artificial neural network; GLR, generalized linear regression; CIT, conditional inference tree; eXGB, extreme gradient boosting.
Figure 7.
Comparison of simulated and observed carbon, nitrogen, and phosphorus pools of aboveground parts of plants for (a–c) RF, (d–f) GBR, (g–i) SVM, (j–l) MLR, (m–o) RRT, (p–r) GLR, (s–u) ANN, (v–x) CIT, and (y–aa) eXGB under grazing conditions. Each red circle represents a single paired observation of observed and simulated values for C, N, or P pool in the samples analyzed. The solid lines are the linear regression be-tween the estimated and observed values. RF, random forest; GBR, generalized boosted regression; MLR, multiple linear regression; SVM, support vector machine; RRT, recursive regression tree; ANN, artificial neural network; GLR, generalized linear regression; CIT, conditional inference tree; eXGB, extreme gradient boosting.
Table 1.
The relative bias (%) between the simulated and observed carbon content, nitrogen content, phosphorus content, ratio of carbon to nitrogen (C:N), ratio of carbon to phosphorus (C:P), ratio of nitrogen to phosphorus (N:P), carbon pool, nitrogen pool, and phosphorus pool of aboveground parts of plant community.
Table 1.
The relative bias (%) between the simulated and observed carbon content, nitrogen content, phosphorus content, ratio of carbon to nitrogen (C:N), ratio of carbon to phosphorus (C:P), ratio of nitrogen to phosphorus (N:P), carbon pool, nitrogen pool, and phosphorus pool of aboveground parts of plant community.
| Conditions | Variable | RF | GBR | SVM | MLR | RRT | GLR | ANN | CIT | eXGB |
|---|---|---|---|---|---|---|---|---|---|---|
| Fencing | Carbon content | 0.73 | 0.71 | −0.33 | 0.00 | 0.79 | −0.55 | 0.01 | 0.14 | −48.81 |
| Nitrogen content | 0.39 | 1.13 | −3.97 | 0.39 | 0.57 | 0.63 | 0.40 | 0.96 | −33.50 | |
| Phosphorus content | 2.46 | −0.58 | −2.26 | −7.67 | −3.40 | 0.07 | 7.04 | −3.70 | 5.78 | |
| C:N | −0.14 | 1.40 | −0.85 | −3.96 | −2.51 | −5.48 | −3.97 | −2.04 | −49.95 | |
| C:P | 2.37 | 2.13 | 7.47 | 19.97 | 10.69 | 15.54 | 19.97 | 10.59 | −45.12 | |
| N:P | 4.63 | 4.50 | 6.25 | 31.79 | 9.07 | 32.60 | 31.78 | 9.50 | −43.09 | |
| Carbon pool | −2.94 | −1.61 | −22.53 | −2.94 | −13.87 | −3.78 | −3.78 | −12.11 | −54.11 | |
| Nitrogen pool | 0.03 | −1.79 | 5.29 | 31.43 | 14.93 | 32.40 | 32.40 | 32.40 | −14.18 | |
| Phosphorus pool | 4.54 | 5.14 | −23.65 | −1.43 | −6.72 | −4.88 | −4.88 | −4.88 | 223.63 | |
| Grazing | Carbon content | −2.86 | −2.13 | −0.97 | −3.28 | −2.63 | −2.96 | −2.88 | −1.84 | −50.47 |
| Nitrogen content | 1.31 | 1.97 | −1.51 | 31.97 | 3.73 | 5.59 | 7.01 | 4.76 | −32.45 | |
| Phosphorus content | 3.25 | 4.29 | −9.96 | −36.93 | 8.25 | 29.14 | 3.38 | 2.47 | 89.16 | |
| C:N | 1.84 | 4.60 | −0.16 | 6.65 | 5.71 | 8.18 | 6.65 | 8.16 | −47.71 | |
| C:P | −0.56 | 1.10 | −0.74 | 7.56 | 7.65 | 8.48 | 7.67 | 3.95 | −51.14 | |
| N:P | −1.02 | 0.41 | −5.06 | −3.41 | −1.43 | −4.79 | −5.28 | −2.23 | −49.69 | |
| Carbon pool | −0.26 | −2.14 | −25.48 | −5.86 | 1.58 | −16.40 | −7.28 | −1.50 | −52.25 | |
| Nitrogen pool | 4.77 | 3.52 | −9.41 | 48.43 | 4.96 | 0.63 | 7.03 | 3.31 | −10.55 | |
| Phosphorus pool | 0.93 | 1.13 | −23.14 | 1.54 | −0.74 | −1.40 | −2.46 | −5.05 | 380.70 |
RF, random forest; GBR, generalized boosted regression; SVM, support vector machine; MLR, multiple linear regression; RRT, recursive regression tree; ANN, artificial neural network; GLR, generalized linear regression; CIT, conditional inference tree; eXGB, extreme gradient boosting.
Table 2.
The RMSE between the simulated and observed carbon content, nitrogen content, phosphorus content, ratio of carbon to nitrogen (C:N), ratio of carbon to phosphorus (C:P), ratio of nitrogen to phosphorus (N:P), carbon pool, nitrogen pool, and phosphorus pool of aboveground parts of plant community.
Table 2.
The RMSE between the simulated and observed carbon content, nitrogen content, phosphorus content, ratio of carbon to nitrogen (C:N), ratio of carbon to phosphorus (C:P), ratio of nitrogen to phosphorus (N:P), carbon pool, nitrogen pool, and phosphorus pool of aboveground parts of plant community.
| Conditions | Variable | RF | GBR | SVM | MLR | RRT | GLR | ANN | CIT | eXGB |
|---|---|---|---|---|---|---|---|---|---|---|
| Fencing | Carbon content (%) | 1.83 | 2.05 | 2.39 | 2.69 | 2.07 | 2.86 | 2.69 | 2.13 | 19.71 |
| Nitrogen content (%) | 0.31 | 0.33 | 0.38 | 0.51 | 0.29 | 0.49 | 0.51 | 0.33 | 0.65 | |
| Phosphorus content (%) | 0.13 | 0.14 | 0.15 | 0.41 | 0.16 | 0.45 | 0.56 | 0.15 | 0.30 | |
| C:N | 1.14 | 1.75 | 3.04 | 4.64 | 2.91 | 5.11 | 4.64 | 3.12 | 15.72 | |
| C:P | 40.88 | 47.23 | 46.75 | 86.00 | 58.33 | 86.21 | 86.00 | 58.76 | 116.52 | |
| N:P | 1.09 | 1.15 | 1.67 | 3.94 | 1.97 | 4.71 | 3.94 | 1.63 | 3.60 | |
| Carbon pool (g C m−2) | 10.13 | 11.86 | 20.59 | 22.74 | 18.56 | 22.70 | 22.70 | 21.57 | 22.20 | |
| Nitrogen pool (g N m−2) | 0.29 | 0.29 | 0.44 | 0.52 | 0.45 | 0.49 | 0.49 | 0.49 | 0.33 | |
| Phosphorus pool (g P m−2) | 0.04 | 0.04 | 0.06 | 0.07 | 0.06 | 0.07 | 0.07 | 0.07 | 0.61 | |
| Grazing | Carbon content (%) | 3.92 | 4.22 | 4.36 | 4.88 | 4.37 | 4.64 | 4.85 | 4.60 | 19.38 |
| Nitrogen content (%) | 0.30 | 0.31 | 0.48 | 0.69 | 0.46 | 0.52 | 0.53 | 0.55 | 0.64 | |
| Phosphorus content (%) | 0.03 | 0.06 | 0.12 | 0.21 | 0.16 | 0.24 | 0.20 | 0.16 | 0.21 | |
| C:N | 6.51 | 6.95 | 7.85 | 8.61 | 8.28 | 9.52 | 8.61 | 9.28 | 14.27 | |
| C:P | 65.10 | 73.40 | 71.50 | 95.03 | 85.71 | 115.06 | 95.09 | 90.66 | 155.49 | |
| N:P | 2.22 | 2.47 | 3.07 | 3.95 | 2.83 | 3.90 | 3.98 | 3.58 | 5.95 | |
| Carbon pool (g C m−2) | 8.26 | 11.48 | 17.99 | 18.98 | 13.75 | 21.48 | 19.00 | 14.74 | 18.08 | |
| Nitrogen pool (g N m−2) | 0.33 | 0.37 | 0.44 | 0.62 | 0.47 | 0.56 | 0.54 | 0.57 | 0.45 | |
| Phosphorus pool (g P m−2) | 0.02 | 0.03 | 0.04 | 0.04 | 0.04 | 0.04 | 0.04 | 0.04 | 0.22 |
RF, random forest; GBR, generalized boosted regression; SVM, support vector machine; MLR, multiple linear regression; RRT, recursive regression tree; ANN, artificial neural network; GLR, generalized linear regression; CIT, conditional inference tree; eXGB, extreme gradient boosting.
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. |
© 2025 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).
Share and Cite
MDPI and ACS Style
Xiang, M.; Fu, G.; Cheng, J.; Ma, T.; Ma, Y.; Zheng, K.; Wang, Z. Simulating the Carbon, Nitrogen, and Phosphorus of Plant Above-Ground Parts in Alpine Grasslands of Xizang, China. Agronomy 2025, 15, 1413. https://doi.org/10.3390/agronomy15061413
AMA Style
Xiang M, Fu G, Cheng J, Ma T, Ma Y, Zheng K, Wang Z. Simulating the Carbon, Nitrogen, and Phosphorus of Plant Above-Ground Parts in Alpine Grasslands of Xizang, China. Agronomy. 2025; 15(6):1413. https://doi.org/10.3390/agronomy15061413
Chicago/Turabian StyleXiang, Mingxue, Gang Fu, Jianghao Cheng, Tao Ma, Yunqiao Ma, Kai Zheng, and Zhaoqi Wang. 2025. "Simulating the Carbon, Nitrogen, and Phosphorus of Plant Above-Ground Parts in Alpine Grasslands of Xizang, China" Agronomy 15, no. 6: 1413. https://doi.org/10.3390/agronomy15061413
APA StyleXiang, M., Fu, G., Cheng, J., Ma, T., Ma, Y., Zheng, K., & Wang, Z. (2025). Simulating the Carbon, Nitrogen, and Phosphorus of Plant Above-Ground Parts in Alpine Grasslands of Xizang, China. Agronomy, 15(6), 1413. https://doi.org/10.3390/agronomy15061413
Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.
