Modelling water - use and yield of selected irrigated subtropical crops using machine learning and hybrid models in north-eastern South Africa

Abstract

The Inkomati-Usuthu Water Management Area in Mpumalanga, South Africa, is a main producer of subtropical crops. These crops are mainly produced under irrigation, yet water resources in this catchment are nearly fully allocated. This calls for improved irrigation efficiency in the region to save water needed for supporting agricultural expansion. Accurate derivation of crop coefficients () and yield response factors () is vital for irrigation management and yield prediction. In this study, transpiration () for the crops was measured using the heat ratio method of monitoring sap flow, while evapotranspiration () was quantified using eddy covariance and surface renewal techniques. Leaf area index for the fields was derived from Landsat 8 imagery. Light gradient boosting machine (LightGBM), Random Forest (RF) and Extreme gradient boosting (XGBoost) machine learning models was investigated for predicting the crop of banana, grapefruit, litchi, mango and sugarcane. The best performing and machine learning-based models were used for developing the crop coefficients () and a hybrid model for predicting, respectively. The LightGBM achieved the highest accuracy in predicting banana, grapefruit, litchi and sugarcane. The XGBoost achieved the highest accuracy in predicting mango. The LightGBM achieved the highest accuracy in predicting the grapefruit, litchi and mango. All the models produced coefficient of determination in the range 0.83–0.96, root mean square error ranging from 0.02 to 0.10 mm/h, mean absolute error ranging from 0.01 to 0.06 mm/h and Kling-Gupta efficiency in the range 0.88–0.97. The grapefruit, litchi and mango produced values of 2.70, 2.50, and 2.90 respectively. The derived information can assist irrigation managers optimize irrigation to promote productive water use in the water-scarce regions.