Deep Reinforcement Learning with Weather Forecasts and Budget Pacing Improves Irrigation Scheduling Under Water Scarcity

Irrigation scheduling under seasonal water-use restrictions is a pressing challenge in water-scarce agricultural regions, where finite volumetric allocations demand careful timing and depth decisions to sustain profitability. Deep reinforcement learning (DRL) offers promise for optimizing such sequential decisions, yet existing observation designs rely on backward-looking weather statistics and omit near-term forecasts that may support the management of limited water across a growing season. This study evaluates whether augmenting the DRL agent’s observation space with seven-day precipitation and reference evapotranspiration forecasts and refactoring existing allocation information into three budget-pacing features can improve irrigation scheduling most effectively under seasonal water scarcity while retaining benefits as restrictions are relaxed. Proximal policy optimization policies were trained within the AquaCrop framework for irrigated maize in southwest Nebraska under 50, 75, 100, 125 mm and unrestricted seasonal water caps. Under the 50 mm cap, the augmented feedforward policy (FB-MLP) achieved 144.76 $/ha, which was 32.2% above the baseline policy and 22.3% above the optimized Soil Moisture Target benchmark against the validation set. Its best-run gains over the baseline across the other four scenarios averaged 2.52%, including a 1.2% improvement under unrestricted irrigation. Under the 75 mm cap, the augmented policy allocated 83.3% of its irrigation to flowering and yield formation. These findings show that the combined observation design improves irrigation scheduling most strongly where water scarcity is binding.