基于分类梯度提升算法和人工神经网络的食松和樱核圆柏的气孔导度模拟  

Stomatal Conductance Simulation of Pinus Edulis and Juniperus Monosperma Based on CatBoost and Artificial Neural Network

作　　者：邵欣欣 雷国庆 曾文治[1] 黄介生[1] SHAO Xin-xin;LEI Guo-qing;ZENG Wen-zhi;HUANG Jie-sheng(School of Water Resources and Hydropower Engineering,Wuhan University,Wuhan 430072,Hubei Province,China)

机构地区：[1]武汉大学水利水电学院,湖北武汉430072

出　　处：《中国农村水利水电》2023年第6期168-174,共7页China Rural Water and Hydropower

基　　金：国家自然科学基金项目(52179039);中国博士后科学基金资助项目(2020M682475);中央高校基本科研业务费专项资金资助(2042021kf0051)。

摘　　要：气孔导度是表征植物蒸腾状态的重要指标,气孔导度的准确量化对于地表水文循环研究具有重要意义。为探索提高气孔导度模拟准确性的方法,本研究利用分类梯度提升算法(CatBoost,CAT)和人工神经网络(Artificial Neural Network,ANN)两种机器学习模型对食松(Pinus edulis)与樱核圆柏(Juniperus monosperma)的气孔导度进行了模拟,并将二者的模拟结果与Ball-Berry模型和Medlyn模型进行了比较。机器学习模型以净光合速率A_(n)、叶表二氧化碳浓度C_(s)、相对湿度RH、饱和水汽压差VPD、叶片温度TL和黎明前叶水势LWP为输入变量,设计了3种建模策略:策略①输入变量为A_(n)、C_(s)和RH;策略②输入变量为A_(n)、C_(s)和VPD;策略③输入变量为A_(n)、C_(s)、RH、VPD、TL和LWP。结果表明①Ball-Berry模型和Medlyn模型模拟效果相近,RMSE分别0.0138和0.0139 mol/(m^(2)·s);②机器学习模型对气孔导度的模拟效果明显优于Ball-Berry模型和Medlyn模型,不同输入策略下CAT和ANN模型的RMSE相比于Ball-Berry模型分别降低了19.35%~45.65%和26.90%~55.07%;③机器学习模型中策略③模拟效果优于策略①和②,且ANN优于CAT,其中策略③中ANN模型的RMSE比策略①和②分别提高了36.70%和38.54%;④各模型和策略下对两种植物组成的整个数据集的气孔导度模拟与分别对食松和樱核圆柏的气孔导度模拟规律是一致的,其中对食松的气孔导度模拟结果优于樱核圆柏。研究表明,机器学习模型(特别是ANN模型)更适用于植物气孔导度的精准模拟,可为植物蒸腾能力估算和农业水文模拟提供实用工具。Stomatal conductance is an important indicator of plant transpiration,and accurate quantification of stomatal conductance is significant for the study of the surface hydrological cycle.To improve the accuracy of stomatal conductance simulations,this paper uses two machine learning models,CatBoost(CAT)and Artificial Neural Network(ANN),and a dataset consisting of Pinus edulis and Juniperus monosperma,to simulate stomatal conductance,and the simulation results are compared with the Ball-Berry model and Medlyn model.The input variables of net photosynthetic rate(A_(n)),leaf surface carbon dioxide concentration(C_(s)),relative humidity(RH),saturated vapor pressure difference(VPD),leaf temperature(TL),and predawn leaf water potential(LWP)are designed as three modeling strategies in machine learning models.The input variables in strategy①are A_(n),C_(s),and RH;in Strategy②are A_(n),C_(s),and VPD;in Strategy③are A_(n),C_(s),RH,VPD,TL,and LWP.The results show that①Ball-Berry model and Medlyn model have similar simulation effects,with RMSE of 0.0138 and 0.0139 mol/(m^(2)·s),respectively.②Compared with the Ball-Berry model and Medlyn model,the RMSE of the CAT model and ANN model under different input strategies decrease by 19.35%~45.65%and 26.90%~55.07%,respectively.③In the machine learning models,the simulation effect of strategy③is better than strategy①and②,and ANN is better than CAT.In strategy③,the RMSE of the ANN model is 36.70%and 38.54%higher than that of strategy①and②,respectively.④Stomatal conductance simulations for the entire dataset consisting of both plants under each model and strategy are consistent with the simulated patterns of stomatal conductance for Pinus edulis and Juniperus monosperma,respectively,with better results for Pinus edulis and then for Juniperus monosperma.These results indicate that the machine learning model(ANN model in particular)is more suitable for accurate simulation of the stomatal conductance of plants,and can provide a practical tool for estimating plant

关 键 词：气孔导度 Ball-Berry模型 Medlyn模型 分类梯度提升算法 人工神经网络模型 

分 类 号：TV93[水利工程—水利水电工程] S271[农业科学—农业水土工程]