Closing the Gap to the Oracle: Benchmarking Domain-Informed Deep Reinforcement Learning for Deficit Irrigation against Perfect Foresight

Sustainable agricultural intensification necessitates precise deficit irrigation strategies to address global water scarcity. However, optimizing intra-seasonal scheduling under stochastic climatic conditions remains a complex control problem. While Deep Reinforcement Learning (DRL) offers a promising approach to flexible decision-making, existing applications often exhibit instability due to high-dimensional state spaces and an inability to enforce physical constraints. This study advances state-of-the-art irrigation control by proposing and benchmarking a tailored DRL framework based on Proximal Policy Optimization (PPO), coupled with the AquaCrop-OSPy simulation model. Moving beyond standard implementations, the research introduces specific enhancements to the learning agent: a reduced observation space limited to five causal biophysical variables, action masking to ensure strict adherence to seasonal water quotas, and a dense reward function based on transpiration efficiency. To rigorously quantify the value of information, the proposed approach is benchmarked against both a standard DRL baseline and a global Evolutionary Algorithm configured with perfect foresight of future weather events. This "oracle" defines the theoretical upper bound of achievable crop water productivity. Experimental validation on maize cultivation under deterministic and stochastic scenarios (Tunis and Nebraska) demonstrates that the proposed agent effectively navigates the trade-off between conservation and yield. The enhanced agent captures approximately 93.5% of the theoretical yield potential defined by the oracle, indicating a minimal performance penalty for the lack of future weather knowledge. Conversely, the standard reference implementation failed to converge under tight resource constraints. Economically, the proposed strategy not only stabilizes yields during extreme drought years but also increases mean net profits by up to 66% compared to the baseline. These findings confirm that integrating domain knowledge through action masking and feature selection transforms DRL into a robust tool for near-optimal irrigation scheduling without requiring extensive weather forecasting.