A reinforcement learning approach to irrigation decision-making for rice using weather forecasts

• Published in: Agricultural water management 2021-05, Vol.250, p.106838, Article 106838
• Main Authors: Chen, Mengting, Cui, Yuanlai, Wang, Xiaonan, Xie, Hengwang, Liu, Fangping, Luo, Tongyuan, Zheng, Shizong, Luo, Yufeng
• Format: Article
• Language: English
• Subjects: Irrigation decision-making; Paddy rice irrigation; Reinforcement learning; Weather forecasts
• ISSN: 0378-3774; 1873-2283
• DOI: 10.1016/j.agwat.2021.106838

Improving efficiency with the use of rainfall is one of the effective ways to conserve water in agriculture. At present, weather forecasting can be used to potentially conserve irrigation water, but the risks of unnecessary irrigation and the yield loss due to the uncertainty of weather forecasts should be avoided. Thus, a deep Q-learning (DQN) irrigation decision-making strategy based on short-term weather forecasts was proposed to determine the optimal irrigation decision. The utility of the method is demonstrated for paddy rice grown in Nanchang, China. The short-term weather forecasts and observed meteorological data of the paddy rice growth period from 2012 to 2019 were collected from stations near Nanchang. Irrigation was decided for two irrigation decision-making strategies, namely, conventional irrigation (i.e., flooded irrigation commonly used by local farmers) and DQN irrigation, and their performance in water conservation was evaluated. The results showed that the daily rainfall forecasting performance was acceptable, with potential space for learning and exploitation. The DQN irrigation strategy had strong generalization ability after training and can be used to make irrigation decisions using weather forecasts. In our case, simulation results indicated that compared with conventional irrigation decisions, DQN irrigation took advantage of water conservation from unnecessary irrigation, resulting in irrigation water savings of 23 mm and reducing drainage by 21 mm and irrigation timing by 1.0 times on average, without significant yield reduction. The DQN irrigation strategy of learning from past irrigation experiences and the uncertainties in weather forecasts avoided the risks of imperfect weather forecasting. ●A reinforcement learning approach for irrigation decision-making is proposed and tested.●Past irrigation experiences and uncertainties of weather forecasts are intelligently learned.●The proposed method can conserve irrigation water and reduce irrigation time without yield loss.●The proposed reinforcement learning approach for irrigation is promising for smart irrigation practices.