High-resolution global modeling of wheat’s water footprint using a machine learning ensemble approach

Abstract Background The agricultural sector accounts for more than 80% of global freshwater consumption, making accurate water demand modeling crucial for preserving these scarce resources, particularly in arid and semi-arid regions. Traditional water footprint (WF) analyses present significant challenges, as they require extensive datasets and complex modeling of soil-crop-water interactions. Machine learning (ML) methods offer advantages through their ability to process complex data relationships efficiently while maintaining high prediction accuracy. Previous ML applications in WF estimation have focused primarily on regional scales. This study extends the application to a global scale for wheat WF prediction, exploring the potential of ML in large-scale agricultural water management. Results This study enhances WF modeling for wheat through the implementation of the AdaBoost algorithm, which offers reduced computation time, handles diverse geographical conditions effectively, and achieves high prediction accuracy with minimal calibration requirements. The model achieved a mean absolute error (MAE) of 108.5 m 3 /t, mean squared error (MSE) of 239.9 m 3 /t, and mean absolute percentage error (MAPE) of 1.51, along with a high prediction accuracy evidenced by a test score of 98.49% and an R 2 value of 0.87. The study revealed distinct outcomes for different clustering methods, demonstrating the model's robustness across varying spatial scales. Conclusions Our findings demonstrate that high-accuracy WF analysis can be achieved with fewer datasets and in a shorter time compared to traditional methods. The ML approach enhances both precision and efficiency of WF estimation for wheat cultivation, offering a practical tool for agricultural water management. This methodology provides valuable insights for researchers and policymakers working towards sustainable water resource management.