|
| 基于不同机器学习方法的典型冬小麦-夏玉米农田CO2通量模拟研究 |
| Simulation of CO2 fluxes in a typical winter wheat-summer maize cropland based on different machine learning methods |
| 投稿时间：2024-05-23 |
| DOI：10.13254/j.jare.2024.0347 |
| 中文关键词: CO2通量，机器学习模型，影响因素，模拟能力，冬小麦，夏玉米，农田 |
| 英文关键词: CO2 flux, machine learning model, influencing factor, simulation ability, winter wheat, summer maize, cropland |
| 基金项目:国家自然科学基金项目（41801013）；江苏省科协青年科技人才托举工程项目（TJ-2023-032）；扬州大学“青蓝工程”资助项目（2023）；宿迁市级指导性科技计划项目（Z2023144）；江苏省气象局青年基金项目（KQ202420）；“宿迁英才”群英计划培养资助项目（2024） |
|
| 摘要点击次数: 209 |
| 全文下载次数: 261 |
| 中文摘要: |
| 为探究典型冬小麦-夏玉米农田CO2通量的影响因素及不同机器学习模型对其模拟能力，本研究选取中国生态系统研究网络禹城综合试验站冬小麦-夏玉米农田为研究对象，基于气象、土壤和作物生长等数据，通过类别型特征提升（CATBoost）、极限梯度提升（XGBoost）、支持向量机（SVM）、人工神经网络（ANN）、随机森林（RF）和K近邻法（KNN）六种机器学习算法，分析其对逐日净生态系统碳交换量（NEE）、总初级生产力（GPP）和生态系统呼吸（Re）模拟的准确性。结果表明：对不同机器学习模型而言，CATBoost模型对农田NEE、GPP 和Re模拟的准确性都优于其他模型，XGBoost模型次之。对不同输入变量组合而言，当包含气象变量、土壤变量和作物变量时各机器学习模型的模拟效果最佳，其次是包含气象变量和作物变量，而包含气象变量和土壤变量的模拟效果相对较弱。对不同农田类型而言，六种机器学习模型在不同输入变量组合下表现出冬小麦田NEE、GPP 和Re的模拟效果优于对应的夏玉米田，其中冬小麦田GPP 和Re的模拟准确性相近，且优于NEE；而夏玉米田NEE 和GPP 的模拟准确性相近，且优于Re。 |
| 英文摘要: |
| In order to investigate the influencing factors of CO2 fluxes in the typical winter wheat-summer maize cropland and assess the simulation ability of different machine learning models, a winter wheat-summer maize cropland at Yucheng Comprehensive Experimental Station of China Ecosystem Research Network was selected as a case study. We obtained the data of meteorology, soil, and crop growth and used six machine learning algorithms including the Categorical Boosting（CATBoost）, Extreme Gradient Boosting（XGBoost）, Support Vector Machines（SVM）, Artificial Neural Network（ANN）, Random Forest（RF）and K Nearest Neighbor Method（KNN）, to analyze the accuracy of their simulations of daily net ecosystem exchange（NEE）, gross primary productivity（GPP）and ecosystem respiration（Re）. The results showed that for different machine learning models, the CATBoost model had better accuracy than the other models in simulating cropland NEE, GPP and Re and had the strongest generalization ability; the XGBoost model ranked second. For different combinations of input variables, the simulation effect of each machine learning model was best when it included meteorological variables, soil variables and crop variables, followed by the inclusion of meteorological variables and crop variables, while the simulation effect of the inclusion of meteorological variables and soil variables was relatively weak. For different cropland types, the performance of the six machine learning models with different combinations of input variables was better for NEE, GPP and Re for the winter wheat field than for the corresponding summer maize field, respectively. Among them, the simulation accuracy of GPP and Re for the winter wheat field was similar and better than NEE, while the simulation readiness of NEE and GPP for summer maize field was similar and better than Re. |
| HTML
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
