Estimating virtual water content and yield of wheat using machine learning tools

The global escalation of water demand has led to significant depletion of water resources, making virtual water content (VWC) and yield assessment crucial for agricultural water management. Traditional calculations heavily rely on pre-determined crop coefficient (Kc) values and extensive datasets, presenting three major challenges: limited data availability in many regions, inaccurate reflection of local conditions through standardized values, and inability to capture spatial–temporal variations in water use patterns. Our study addresses these challenges by developing an innovative machine learning (ML) framework that eliminates Kc dependency while maintaining high prediction accuracy. Our approach combines climate, soil and agronomic variables collected from 81 Turkish provinces (2008–2019) to develop Linear Regression (LR) and Random Forest (RFR) models. Comparative analysis revealed LR's superior performance, achieving high accuracy (R2 = 0.98) and consistently low error rates (<5 %) in predicting both VWC and wheat yield across diverse geographical regions. Feature importance analysis identified April rainfall and early-season temperatures as the key predictive variables. We also derived simplified equations for VWC and yield predictions, incorporating key inputs. This approach provides a more accessible and precise tool for estimating water consumption in wheat production, particularly valuable in regions with limited data availability.