ABSTRACT To address the issues of irrigation decisions in water spinach production being highly dependent on experience and struggling to cope with rainfall uncertainty and crop growth stage differences, this study proposes an environmentally enhanced proximal policy optimization (EN‐PPO) precision irrigation control method. On the basis of the traditional proximal policy optimization (PPO) reinforcement learning framework, the environmental information enhancement mechanism is introduced to decompose the state space into policy and environmental states. The EN‐PPO is not simply the direct application of PPO but rather incorporates two key improvements addressing the strong uncertainties and safety constraints of agricultural irrigation. First, it introduces a dynamic shearing strategy in the near‐end policy update, allowing the policy update magnitude to adaptively adjust with training phases and environmental fluctuations, which mitigates the training oscillations caused by rainfall randomness and sample scarcity; second, it designs a negative incentive mechanism oriented towards production safety and resource efficiency, penalizing behaviors that violate soil moisture safety ranges, redundant irrigation, and failure to meet rotation and switching constraints, which guides the agent to avoid high‐risk decisions during the exploration phase and maintain stable and water‐saving irrigation strategy outputs during the utilization phase. The experimental results show that the EN‐PPO algorithm exhibits superior comprehensive performance in terms of policy convergence stability, water‐saving effect, and rainfall utilization efficiency, which achieves more reasonable irrigation timing and water volume regulation without affecting normal crop growth and yield. It provides a feasible approach for the engineering application of reinforcement learning in agricultural precision irrigation.
Tang et al. (Mon,) studied this question.