Assessing the value of deep reinforcement learning for irrigation scheduling

Due to increasing global water scarcity pressure, researchers, policy makers and industry are looking for innovative solutions to increasing agricultural water productivity. Motivated by recent success within complex decision-making environments, Deep Reinforcement Learning (DRL) is being proposed as a method for optimizing irrigation strategies. Early research has hinted towards increased profits with DRL compared to heuristic approaches such as soil-moisture thresholds or fixed schedules. However, an assessment of the value of DRL for irrigation scheduling that incorporates local climate variability and water-use restrictions has yet to be performed. To address this gap in the literature, we created aquacrop-gym, an open-source Python framework for researchers to train and evaluate customized irrigation strategies within the crop-water model AquaCrop-OSPy. In this analysis, aquacrop-gym was used to quantify the value of DRL in comparison to conventional irrigation scheduling techniques (e.g., optimized soil-moisture heuristic) for maize production in an intensively irrigated region of the central United States. The DRL and heuristic approaches were both trained on 70 years of weather data produced from the weather generator LARS-WG, and evaluated on 30 unseen validation years of generated weather data. Findings from this analysis show that in the presence of high rainfall variability, DRL does not outperform conventional optimized heuristics. However, in the scenario where rainfall is set to zero, DRL approaches achieve higher profits on the unseen validation years. Similarly, DRL approaches also outperform optimized heuristics when severe water-use restrictions are introduced. Our analysis demonstrates that DRL approaches are a promising method of irrigation scheduling, notably in regions where farmers are faced with significant physical or regulatory water scarcity.