Intelligent modeling of wheat water footprint for sustainable water management across Egypt’s climatic zones

Abstract Managing and modeling different water resources in arid regions is the key to an accurate estimation of water uses and achieving agricultural sustainability under limited water. In this study, four Egyptian Nile Delta Governorates namely Ad Dakahliyah, Al Gharbiyah, Kafrash shaykh and Dumyat were selected as a primary wheat-producing location for modelling water footprint for green (WF g ) and blue (WF b ) colours. Seven water footprint models were established in 2006–2016 based on monthly open access results. These models were varied in volume and structure of the independent variables. Besides, these models were compared and evaluated using five machine learning algorithms, including random forest, support vector regression, Bagging, Boosting, and Matern 5/2 Gaussian process. The results of this study revealed that Model 2 utilizing the M5/2 GPR algorithm was the best prediction model for blue and green WFs in the Ad-Daqahliyah governorate. Its characteristics were R 2 = 0.94, RMSE = 15.53 m 3 /ton, and MAE = 14.49 m 3 /ton; and R 2 = 1, RMSE = 0.32 m 3 /ton, and MAE = 0.19 m 3 /ton, respectively. As well, the best predictive model for blue WF in the Dumyat was Model 2 at the boosting algorithm, which obtained R 2 = 0.74, RMSE = 28.56 m 3 /ton, and MAE = 21.06 m 3 /ton). In contrast to the other models, Model 1 with M5/2 GPR gave the best simulation for the estimation of green WF, with high R 2 = 0.96, RMSE = 2.95 m 3 /ton, and MAE = 2.17 m 3 /ton. In addition, Model 1 at M5/2 GPR was the best model for predicting blue WF at Al Gharbiyah producing a high coefficient of determination (R 2 = 0.88) and less error (RMSE = 24.53 m 3 /ton, and MAE = 16.28 m 3 /ton). Model 6 at the M5/2 GPR algorithm had the best performance metrics (R 2 = 1.00, RMSE = 0.32 m 3 /ton, and MAE = 0.21 m 3 /ton). Model 2 with M5/2 GPR produced the best results among the models for blue and green WFs in the Kafr ash-Shaikh site, with an R 2 of 0.97, an RMSE of 11.74 m 3 /ton, and an MAE of 7.87 m 3 /ton; R 2 = 1.00, RMSE = 0.34 m 3 /ton, and MAE = 0.23 m 3 /ton, respectively. These models achieved high performance and less residual errors according to statistical analysis methods. Thus, the developed models were proven to produce satisfactory results and will be a precise tool for the process of decision making for water-managers.