Temporal convolution-network-based models for modeling maize evapotranspiration under mulched drip irrigation

Abstract Accurate prediction of crop actual evapotranspiration (ETc) has great significance in designing irrigation plans and improving the water-resource use efficiency. However, existing experiment methods are either expensive or labor-costly, and the crop-coefficient (Kc) approach always results in high errors in calculating ETc, especially for nonstandard conditions like drip irrigation under plastic-film mulch. In this study, a Temporal Convolution Network (TCN) with two engineering methods (Principal Component Analysis (PCA) and Maximal Information Coefficient (MIC)) was developed to predict ETc using a two-year dataset from lysimeters for maize under drip irrigation with film mulch. The TCN models comprised of Long Short-Term Memory Networks (LSTM) and Deep Neural Networks (DNN). To further test the results of the TCN models, they were compared in predicting Kc values with FAO-56 Kc values in the literature. The results suggested that plant height, mean temperature, maximal temperature, relative humidity, solar radiation, leaf-area index, and soil temperature are the seven most important features affecting maize evapotranspiration. TCN models all performed well in predicting ETc, with R2 in the range of 0.91–0.95, MSE 0.144–0.296 mm/d, and MAE 0.309–0.434 mm/d. Compared with the LSTM and DNN models, TCN with all input features (TCN-all) improved R2 by 0.13 and 0.06, respectively, and decreased MSE and MAE by 0.402 and 0.233 mm/d, and 0.187 and 0.153 mm/d, respectively. TCNs with features selected by the PCA and MIC methods both outperformed the PCA-based LSTM and DNN models, and the MIC-based LSTM and DNN models. Kc values predicted by the TCN-all model were closer to the actual Kc value than those modified by FAO-56.