Abstract: As a highly complex agricultural system, the irrigation district involves multi-scale processes ranging from farmland irrigation and canal system water conveyance to irrigation district water supply in its management and planning, with the coexistence of multi-level decision-makers such as farm managers and irrigation district administrators. Current research on irrigation district water resources allocation inadequately considers the cross-scale and cross-level water quantity coordination as well as dynamic interaction and feedback relationships, which to a certain extent separates the integrity of the optimal allocation of irrigation district water resources systems. To address this, we integrate large-scale system decomposition-coordination theory with bi-level programming to develop a novel scale-coupled and hierarchical game-based optimization model. This model establishes a cross-scale dynamic feedback mechanism linking field irrigation schedules, multi-level canal water allocation, and irrigation district water supply decisions, thus achieving daily-scale refined water management in the irrigation district. An intelligent cyclic algorithm automates the cross-scale ‘optimization-feedback’ process, ensuring consistency between single-scale control and system-wide optimization. This method not only improves irrigation water use efficiency but also enhances the coordination among different management entities. Using the Qinglongshan Irrigation District in Heilongjiang Province as a case study, we validated the model's effectiveness. Results show that the model achieves multi-scale coordination and efficient water allocation: compared with conventional irrigation practices, the optimized scheme increases water productivity by 18%. The performance of this model is significantly superior to that of isolated single-scale optimization models—using only the field-scale model increases irrigation water use by 9%; adopting only the canal-scale bi-level programming model results in a 4.1% rise in water conveyance losses, which highlights the advantage of the proposed model in balancing cross-scale benefit trade-offs. In comparison with the integrated model developed in this study, the standalone district-scale model increases water diversion by 24% and reduces economic water productivity by 19%, indicating that the established unified framework can effectively avoid resource waste and suboptimal solutions caused by local-scale optimization. The bi-level programming module in the model can effectively capture the hierarchical decision-making attributes among water users, improving the coordination index of water allocation schemes by 9.1%. By establishing a closed-loop regulatory mechanism of "demand-allocation-supply", this study bridges the gap between system integrity and scale adaptability, providing a theoretical basis and practical tool for achieving resilient, efficient, and coordinated agricultural water management in large-scale irrigation districts with multiple water sources and administrative levels.