Gross Primary Production (GPP) of Vegetation Calculated by Machine Learning

Based on flux tower observations, this study comprehensively evaluated five mainstream GPP products, including MODIS (Moderate Resolution Imaging Spectroradiometer), PML-V2 (Penman-Monteith-Leuning Version 2), GOSIF (Global Orbiting Carbon Observatory-2 based Solar Induced chlorophyll Fluorescence), CEDAR (sCaling Ecosystem Dynamics with ARtifical intelligence), and TL-LUE, and quantified their uncertainties using the Bayesian Three-Cornered Hat method. On this basis, by integrating multi-source data, a high-fidelity GPP dataset was generated using five machine learning methods: Categorical Boosting, Support Vector Machine Regression, Light Gradient Boosting, Extreme Gradient Boosting, and Random Forest. Validated against data from 15 flux tower sites in mainland China, CatBoost exhibited the best performance, with the lowest RMSE and MAE and the highest R².

本研究基于通量塔观测数据，对5款主流总初级生产力（Gross Primary Productivity, GPP）产品开展了全面评估，涵盖MODIS（中分辨率成像光谱仪，Moderate Resolution Imaging Spectroradiometer）、PML-V2（彭曼-蒙特斯-刘恩版本2，Penman-Monteith-Leuning Version 2）、GOSIF（基于全球轨道碳观测卫星-2的太阳诱导叶绿素荧光产品，Global Orbiting Carbon Observatory-2 based Solar Induced chlorophyll Fluorescence）、CEDAR（人工智能驱动生态系统动力学尺度化模型，sCaling Ecosystem Dynamics with ARtifical intelligence）以及TL-LUE，并采用贝叶斯三角帽法（Bayesian Three-Cornered Hat method）量化了各产品的不确定性。在此基础上，本研究通过整合多源数据，运用分类提升算法（Categorical Boosting）、支持向量机回归（Support Vector Machine Regression）、轻量梯度提升（Light Gradient Boosting）、极端梯度提升树（Extreme Gradient Boosting）与随机森林（Random Forest）共5种机器学习方法，构建了高保真总初级生产力数据集。经中国大陆地区15个通量塔站点的观测数据验证，分类提升算法表现最优，其均方根误差（Root Mean Square Error, RMSE）、平均绝对误差（Mean Absolute Error, MAE）最低，决定系数（Coefficient of Determination, R²）最高。