Scholarlink

<p>Carbon dioxide (CO₂) fluxes in terrestrial ecosystems, particularly Net Ecosystem Exchange (NEE), Gross Primary Productivity (GPP), and Ecosystem Respiration (RE), are fundamental indicators of biosphere–atmosphere carbon exchange and play a central role in regulating the global carbon cycle. However, their accurate quantification in subtropical grasslands remains challenging due to complex nonlinear interactions among climatic drivers and increasing environmental pressures. This study addresses this limitation by applying advanced machine learning models to predict CO₂ flux components in the Southern Brazilian Pampa, an ecologically important yet underrepresented biome. Meteorological and flux data were obtained from eddy covariance towers at Aceguá (2018–2022) and Santa Maria (2014–2023) at half-hourly resolution. Six models (multiple linear regression (MLR), artificial neural network (ANN), random forest (RF), extreme gradient boosting (XGB), support vector machine (SVM), and decision tree (DT)) were evaluated, trained and tested using solar radiation, air temperature, relative humidity, wind speed, and vapour pressure deficit as predictors. Model performance was evaluated using correlation heatmaps, statistical metrics, Taylor diagrams, violin plots, and multi-temporal analyses. Results from the heatmap showed very strong physical relationship, with NEE negatively correlated with solar radiation (<i>R</i> = − 0.93 to − 0.96), while GPP exhibited strong positive correlations (<i>R</i> &gt; 0.94), and RE was closely linked to air temperature (<i>R</i> = 0.87–0.91). Ensemble models (RF and XGB) outperformed others, particularly in Santa Maria, achieving high predictive accuracy for NEE (R² = 0.98 and 0.98), GPP (R² = 0.99 and 0.98), and RE (R² = 0.95 and 0.94). Taylor diagrams indicated strong agreement between predicted and observed RE (σ = 7.3 µmol m⁻² s⁻¹; <i>R</i> &gt; 0.93), supported by close median values (actual RE: 7.32; XGB: 7.34; RF: 7.29). In Aceguá, RF and ANN provided the best RE predictions (test R² = 0.86), though with slightly higher variability. Finally, the ensemble models demonstrated strong predictive performance under the evaluated within-site chronological validation design and study limitations. This shows the potential usefulness of machine learning approaches for site-level carbon flux assessment in subtropical grassland ecosystems.</p> Graphical Abstract <p></p> <p>This graphical abstract provides a concise and visually engaging summary of the study on optimizing CO₂ flux quantification in Southern Brazilian Pampa agroecosystems using machine learning techniques. It integrates key elements of the research workflow and findings into a single visual narrative. The upper section presents the study sites (Aceguá and Santa Maria) alongside the primary meteorological drivers, including solar radiation, temperature, relative humidity, wind speed, and vapour pressure deficit. The central panel highlights the application of multiple machine learning models, such as Random Forest, XGBoost, Artificial Neural Networks, Support Vector Machines, Multiple Linear Regression, and Decision Trees for predicting carbon flux components: Net Ecosystem Exchange (NEE), Gross Primary Productivity (GPP), and Ecosystem Respiration (RE). Notably, ensemble approaches achieved the highest predictive accuracy, with coefficients of determination reaching R² ≈ 0.98 for NEE, R² ≈ 0.99 for GPP, and R² ≈ 0.95 for RE. The correlation heatmaps reveal strong relationships between variables, such as GPP exhibiting correlations up to ~ 0.96 with radiation and NEE showing inverse correlations of about − 0.96 with productivity-related variables. The lower figures present Taylor diagrams, where model performance clusters close to the reference point, indicating low standard deviation differences and high correlation (<i>R</i> &gt; 0.90) during both training and testing phases. Overall, this graphical abstract serves as a quick, visually appealing summary of the research, enabling rapid understanding of the study’s objectives, methods, and major outcomes without requiring a full reading of the manuscript.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Quantification of CO2 Fluxes in Agroecosystems: A Machine Learning Approach in the Southern Brazilian Pampa

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月23日

查看全文
  • Olusola Samuel Ojo,
  • Michel Baptistella Stefanello,
  • Alecsander Mergen,
  • Débora Regina Roberti

摘要

Carbon dioxide (CO₂) fluxes in terrestrial ecosystems, particularly Net Ecosystem Exchange (NEE), Gross Primary Productivity (GPP), and Ecosystem Respiration (RE), are fundamental indicators of biosphere–atmosphere carbon exchange and play a central role in regulating the global carbon cycle. However, their accurate quantification in subtropical grasslands remains challenging due to complex nonlinear interactions among climatic drivers and increasing environmental pressures. This study addresses this limitation by applying advanced machine learning models to predict CO₂ flux components in the Southern Brazilian Pampa, an ecologically important yet underrepresented biome. Meteorological and flux data were obtained from eddy covariance towers at Aceguá (2018–2022) and Santa Maria (2014–2023) at half-hourly resolution. Six models (multiple linear regression (MLR), artificial neural network (ANN), random forest (RF), extreme gradient boosting (XGB), support vector machine (SVM), and decision tree (DT)) were evaluated, trained and tested using solar radiation, air temperature, relative humidity, wind speed, and vapour pressure deficit as predictors. Model performance was evaluated using correlation heatmaps, statistical metrics, Taylor diagrams, violin plots, and multi-temporal analyses. Results from the heatmap showed very strong physical relationship, with NEE negatively correlated with solar radiation (R = − 0.93 to − 0.96), while GPP exhibited strong positive correlations (R > 0.94), and RE was closely linked to air temperature (R = 0.87–0.91). Ensemble models (RF and XGB) outperformed others, particularly in Santa Maria, achieving high predictive accuracy for NEE (R² = 0.98 and 0.98), GPP (R² = 0.99 and 0.98), and RE (R² = 0.95 and 0.94). Taylor diagrams indicated strong agreement between predicted and observed RE (σ = 7.3 µmol m⁻² s⁻¹; R > 0.93), supported by close median values (actual RE: 7.32; XGB: 7.34; RF: 7.29). In Aceguá, RF and ANN provided the best RE predictions (test R² = 0.86), though with slightly higher variability. Finally, the ensemble models demonstrated strong predictive performance under the evaluated within-site chronological validation design and study limitations. This shows the potential usefulness of machine learning approaches for site-level carbon flux assessment in subtropical grassland ecosystems.

Graphical Abstract

This graphical abstract provides a concise and visually engaging summary of the study on optimizing CO₂ flux quantification in Southern Brazilian Pampa agroecosystems using machine learning techniques. It integrates key elements of the research workflow and findings into a single visual narrative. The upper section presents the study sites (Aceguá and Santa Maria) alongside the primary meteorological drivers, including solar radiation, temperature, relative humidity, wind speed, and vapour pressure deficit. The central panel highlights the application of multiple machine learning models, such as Random Forest, XGBoost, Artificial Neural Networks, Support Vector Machines, Multiple Linear Regression, and Decision Trees for predicting carbon flux components: Net Ecosystem Exchange (NEE), Gross Primary Productivity (GPP), and Ecosystem Respiration (RE). Notably, ensemble approaches achieved the highest predictive accuracy, with coefficients of determination reaching R² ≈ 0.98 for NEE, R² ≈ 0.99 for GPP, and R² ≈ 0.95 for RE. The correlation heatmaps reveal strong relationships between variables, such as GPP exhibiting correlations up to ~ 0.96 with radiation and NEE showing inverse correlations of about − 0.96 with productivity-related variables. The lower figures present Taylor diagrams, where model performance clusters close to the reference point, indicating low standard deviation differences and high correlation (R > 0.90) during both training and testing phases. Overall, this graphical abstract serves as a quick, visually appealing summary of the research, enabling rapid understanding of the study’s objectives, methods, and major outcomes without requiring a full reading of the manuscript.