太阳成8722(中国)有限公司-GREEN NO.1

标题: Exploring the potential of deep factorization machine and various gradient boosting models in modeling daily reference evapotranspiration in China

作者: Zhou, ZG (Zhou, Zhigao); Zhao, L (Zhao, Lin); Lin, AW (Lin, Aiwen); Qin, WM (Qin, Wenmin); Lu, YB (Lu, Yunbo); Li, JY (Li, Jingyi); Zhong, Y (Zhong, Yang); He, LJ (He, Lijie)

来源出版物: ARABIAN JOURNAL OF GEOSCIENCES  卷: 13  期: 24  文献号: 1287  DOI: 10.1007/s12517-020-06293-8  出版年: DEC 2020  

摘要: Reference evapotranspiration (ETo) data are the great important information for the irrigation scheduling design and regional water resources planning. This study evaluated the potential of a newly proposed artificial intelligence (i.e., DeepFM) and three newly developed gradient boosting models (i.e., CatBoost, LightGBM, and XGBoost) for estimating daily ETo using limited meteorological parameters in contrasting climates of China. The three commonly used tree-based models (i.e., GBDT, RF, and ET) and one kernel-based model (i.e., SVM-RBF) were also assessed for comparison. Eight various input combinations of meteorological parameters (temperature, humidity, wind speed, and sunshine duration) from twelve stations during 1961-2016 in contrasting climates of China were employed for model training and testing. The results showed that all models (except ET) could obtain satisfactory prediction accuracy (R2 >= 0.9906, RMSE <= 0.1773 mm d-1, and MAE <= 0.1226 mm d-1) for ETo estimation in contrasting climates of China when complete meteorological parameters were input. Moreover, all models can be sorted from the best to the poorest as follows based on the model accuracy: CatBoost > LightGBM > XGboost > GBDT > SVM-RBF > DeepFM > RF > ET; regarding their stability, the eight algorithms can be ranked from the best to the poorest as follows: ET > SVM-RBF > DeepFM > CatBoost > XGboost > GBDT > LightGBM > RF. Therefore, considering the prediction accuracy and model stability comprehensively, the CatBoost and LightGBM models were highly recommended for predicting daily ETo in this study. The XGBoost and GBDT models were promising alternative models for estimating daily ETo. The DeepFM model also had the potential to estimating daily ETo under single variable combinations. The sunshine duration played a more important role than temperature, humidity, and wind speed in estimating ETo, and it was a promising alternative parameter under the circumstances of lacking global radiation observation.

入藏号: WOS:000606496500009

语言: English

文献类型: Article

作者关键词: Reference evapotranspiration; Deep factorization machine; Gradient boosting models; Prediction accuracy; Model stability

地址: [Zhou, Zhigao; Zhao, Lin; Lin, Aiwen; Zhong, Yang] Wuhan Univ, Sch Resource & Environm Sci, Wuhan 430079, Peoples R China.

[Qin, Wenmin; Lu, Yunbo] China Univ Geosci, Sch Earth Sci, Lab Crit Zone Evolut, Wuhan 430074, Peoples R China.

[Li, Jingyi] Capital Univ Econ & Business, Sch Stat, Beijing 100026, Peoples R China.

[He, Lijie] Chinese Univ Hong Kong, Dept Geog & Resource Management, Shatin, Hong Kong 999077, Peoples R China.

通讯作者地址: Zhou, ZG (通讯作者)，Wuhan Univ, Sch Resource & Environm Sci, Wuhan 430079, Peoples R China.

电子邮件地址: leehong@whu.edu.cn

影响因子：1.327