Effect of data splitting scenarios on machine-learning-driven prediction of broccoli actual evapotranspiration under semi-arid condition

Abstract Actual crop evapotranspiration estimation is needed for development of irrigation schedules to enhance crop yield and water use efficiency. Machine learning (ML) algorithms are sensitive to proportion of training to testing datasets and thus it is essential to assess how data partitioning scenarios influences ML models performance for prediction of variables efficiently. Therefore, in this study, we have developed and assessed nine data partitioning scenarios for predicting Broccoli evapotranspiration (ET c ) utilizing eight machine learning algorithms. The developed scenarios were 50:50 (training: testing) scenario (S1), 55:45 (S2), 60:40 (S3), 65:35(S4),70:30(S5), 75:25(S6), 80:20(S7), 85:15(S8) and 90:10 (S9). Eight ML algorithms were Random Forest (RF), Linear regression (LR), Decision Tree (DT), Support Vector Machine (SVM), gradient boosting (GB), stochastic gradient boosting (SGD), Adaptive Boosting (AdaBoost) and artificial neural network (ANN). The predicted broccoli ET c was compared with the actual broccoli ET c obtained using the weighing type field lysimeters available at the Water Technology Center Farm, ICAR-IARI, New Delhi. Additionally, SHAP (SHapley Additive exPlanations) analysis was done to assess the impacts of predictive variables on the Broccoli ET c . Results revealed that the RF and LR algorithms found best for predicting daily broccoli actual ET c during training and testing phases. The partitioning of the dataset has influenced the algorithms performance. The best results were found when training and testing dataset were equal (scenario S1). Under other scenarios, the performance of the developed models had reduced during both training and testing phases. The RF algorithm was best during training phase and yielded value of coefficient of determination (R 2 ), mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) as 0.956, 0.038, 0.195, 0.143, and 6.4%, respectively. While, the LR found best during testing phase and the values of R 2 , MSE, RMSE, MAE, and MAPE were 0.77, 0.375, 0.613, 0.514, and 21.8%. The SHAP analysis revealed that the sunshine hour (SSH) was the most influencing input variable affecting broccoli ET c prediction for all the developed models. The developed models should be tested in other climatic conditions to check their suitability in predicting the broccoli ET c . The findings of the current study could be useful for planning and designing of efficient irrigation system.