| 摘要: |
| 【目的】针对气象数据缺失问题，研究机器学习方法在计算参考作物蒸散量（ET0）中的应用，为ET0的估算提供支持。【方法】基于已有的本地气象站与邻站数据，利用极限梯度提升法（XGBoost）模型和支持向量机（SVM）模型2种机器学习算法，结合江西省吉安和鄱阳2个气象站及对应邻站1966－2015年逐月气象资料，使用K折交叉验证法及4种统计指标（决定系数（R2）、均方根误差（RMSE）、平均偏置误差（MBE）和归一化均方根误差（NRMSE））评估2种输入模式（本地输入或与邻站数据相融合输入）下估算逐月ET0的适用性。【结果】2种输入模式下，XGBoost模型的性能整体优于SVM模型。只使用本地资料作为输入时，以最高温度（Tmax）、最低温度（Tmin）、地表总辐射量（Rs）为参数的模型性价比最高。使用邻站结合本地资料作为输入时，XGBoost模型对应的最佳输入参数为邻站ET0数据（ET0-ex），其平均R2为0.986，RMSE和MBE分别为0.195和-0.106 mm/d，NRMSE为0.079。【结论】综合精度和稳定性等因素，当存在部分气象资料缺失时，使用本地数据或与邻站数据相结合，可成功估算出目标站点的ET0值。推荐使用XGBoost模型且2种输入模式下最实用的输入组合分别为Tmax、Tmin、Rs和ET0-ex，可用于类似江西鄱阳湖地区气象资料缺乏条件下ET0的估算。 |
| 关键词: 参考作物蒸散量 机器学习算法 极限梯度提升 支持向量机 K折交叉验证法 鄱阳湖地区 |
| DOI： |
| 分类号: |
| 基金项目:国家自然科学基金项目（51709143，51769010,51979133） |
|
| Cross-station adaptability of ET0 based on machine learning |
|
DONG Jianhua,LIU Xiaogang,WU Lifeng,et al
|
| Abstract: |
| 【Objective】This study aimed to evaluate the application of machine learning methods on estimating reference crop evapotranspiration (ET0) in the scenario of lacking full meteorological data.【Method】Two machine learning algorithms,the extreme gradient boosting (XGBoost) method and the support vector machine (SVM) model,were used in this study to evaluate the applicability of monthly ET0 estimation under two input patterns of local data only and local data in combination with cross station data.Monthly meteorological data in 1966-2015 from two weather stations (Ji’an and Poyang) and their corresponding adjacent stations were used as inputs.The K fold cross validation and four statistical indicators including coefficient of determination (R2),root mean square error (RMSE),mean bias error (MBE) and normalized root mean square error (NRMSE) were applied for model evaluation.【Result】The XGBoost model overall performed better than the SVM model.Using local data only,the model with maximum air temperature (Tmax),minimum air temperature (Tmin) and global solar radiation (Rs) as parameters had the highest cost performance.When data from the adjacent station was combined with local data as inputs,the best input parameter for the XGBoost model was ET0 data from the adjacent station (ET0-ex).The average values of R2,RMSE,MBE,and NRMSE were 0.986,0.195 mm/d,-0.106 mm/d and 0.079,respectively.【Conclusion】In the scenario of lacking full meteorological data,the ET0 value of a target station can be successfully estimated by using local data or combined data from local and adjacent stations.The XGBoost model is recommended under both input patterns with most practical inputs of Tmax,Tmin and Rs combination and ET0-ex,respectively.The model can be used to estimate ET0 in regions similar to the Poyang Lake area of Jiangxi when full meteorological data is lacking. |
| Key words: reference crop evapotranspiration machine learning algorithm extreme gradient boosting support vector machines K-fold cross validation Poyang Lake area |
