Parameter Estimation to Process Optimization: Integrating a Machine Learning Framework for Sustainable Border Irrigation Management

ABSTRACT Optimizing border irrigation management is critical for sustaining agricultural productivity and mitigating irrigation‐induced land degradation, particularly by reducing water‐driven soil nutrient leaching and secondary salinization. Accurate estimation of soil infiltration and surface roughness parameters is essential for improving irrigation management and water‐use efficiency; however, existing approaches still have limited capability to jointly and reliably estimate these key parameters under field conditions. This study employs a joint inversion framework that integrates Kernel Extreme Learning Machine (KELM) with Particle Swarm Optimization (PSO) to simultaneously estimate two key soil hydraulic parameters affecting irrigation performance: infiltration parameters and roughness. Within this framework, KELM is adopted to capture nonlinear relationships between particle swarm fitness (advance‐recession simulation error) and the soil hydraulic parameters, while PSO is employed to solve optimal parameter combinations. Comparative analyses with multiple linear regression (MLR), back propagation neural networks (BP), and support vector machines (SVM) show that KELM achieves higher accuracy, with an RMSE = 1.09, MAE = 0.81 and RAE = 0.37. Based on parameter inversion, process‐based optimization of inflow rate and cutoff time through WinSRFR (a simulation tool for surface irrigation hydraulics) simulations were performed to maximize the comprehensive irrigation performance indicator while meeting infiltration depth requirements. Furthermore, validation of the KELM–PSO framework across multiple wheat and cotton irrigation events revealed spatiotemporal variability of soil hydraulic parameters, indicating the necessity of adaptive irrigation strategies in response to parameter changes. For instance, when soil infiltration rates increase, raising the inflow rate with earlier cutoff may reduce deep percolation and enhance irrigation uniformity. Such process‐based strategies may, to some extent, provide feasible guidance for sustainable irrigation in data‐scarce regions, particularly where nutrient transport monitoring is unavailable.