Exploring interpretable and non-interpretable machine learning models for estimating winter wheat evapotranspiration using particle swarm optimization with limited climatic data

• Published in: Computers and electronics in agriculture Vol. 212; p. 108140
• Main Authors: Zhao, Xin, Zhang, Lei, Zhu, Ge, Cheng, Chenguang, He, Jun, Traore, Seydou, Singh, Vijay P.
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.09.2023
• Subjects: agriculture; agroecosystems; air; crop coefficient; Crop evapotranspiration; electronics; evapotranspiration; Interpretability; irrigation; Machine learning; meteorological data; Particle swarm optimization algorithm; solar radiation; support vector machines; temperature; water management; water use efficiency; wind speed; Winter wheat; Interpretability; Crop evapotranspiration; Particle swarm optimization algorithm; Machine learning; Winter wheat
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2023.108140

•Kc and solar radiation are important for obtaining reliable winter wheat ETc.•The non-interpretable ML showed better performance in winter wheat ETc estimation.•PSO-based support vector machine model obtained the best results for estimating ETc.•LIME analysis detected the inflection points of climatic parameters related to ETc. Accurate estimation of crop evapotranspiration (ETc) is crucial for improving the water use efficiency and designing and operating irrigation systems. To accurately calculate winter wheat ETc with limited meteorological data, the present study proposed two interpretable machine learning (ML) models (random forest (RF) and extreme gradient boosting (XGBoost)) as well as non-interpretable ML models (support vector machine (SVM) and deep neural network (DNN)) based on the particle swarm optimization (PSO) algorithm using observed winter wheat ETc data during the period from 2007 to 2013 at Luan Cheng Agro-ecosystem Experimental Station. Mean absolute error (MAE), root mean square error (RMSE), Nash-Sutcliffe efficiency (NSE), coefficient of determination (R2), and global performance indicator (GPI) were used to assess the performance of models. This demonstrated that the ML models based on the crop coefficient (Kc) and solar radiation (Rn) were accurate and offered a workaround for calculating winter wheat ETc in the absence of meteorological data. In four ML models, the ninth input combination, consisting of Kc, Rn, daily air maximum temperature (Tmax), daily air minimum temperature (Tmin), sunshine hours (n), and wind speed with a height of 2 m (U2), produced the best estimate of ETc. Among them, the PSO-based SVM (PSO-SVM) model obtained the best results for estimating ETc with MAE, RMSE, NSE, R2, and GPI values of 0.389 mm·d−1, 0.562 mm·d−1 0.910, 0.911, and 0.975, respectively, showing the advantages of the non-interpretable ML model in ETc forecasting. Accurate descriptions of actual hydrological and climatic processes were given by local interpretable model-agnostic explanations (LIME). The inflection points of daily climatic parameters (Tmin, Tmax, Rn, n) related to ETc were determined to be 3.80 °C, 5.50 °C, 1.62 MJ·m−2·d−1, 1.37 h, respectively. This work has potential to overcome the difficulty of measuring winter wheat ETc properly due to the lack of meteorological data and accomplish appropriate water management to conserve water and increase water productivity.