Water is a major contributing factor for plant growth and development. Agricultural water management is a major concern for agriculturists, as fresh water resources are being depleted and research in the area of water optimization for agricultural irrigation is in its initial stages. The main goal of the agricultural systems is to supply water spatially according to the soil/land conditions such as water holding capacity of the soil, texture, drainage, texture and topography. While currently existing irrigation systems provide constant irrigation throughout the field, this may result in over irrigation in some areas and under irrigation in the other areas. The main goal of my research is to minimize the excess application of water in a specific location according to the daily conditions like temperature, solar radiation, rainfall from weather data website and soil water content reported by the sensors. The precision techniques that are presently being used are solely based on sensor data obtained from different sources and leverage supervised learning technique i.e. the system is provided with the solutions to all different environments initially. The main disadvantage of this approach is that, the actual scenario might differ widely from the programmed/provided cases. So the system needs to adapt for variable weather, soil and plant conditions and learn from the past experience as well as try new methods. In this thesis a novel technique to use reinforcement learning (an adaptive learning system) on crop system models to make the irrigation system adaptive is discussed. The main goal is to optimize the water consumption as much as possible without affecting the crop yield by using reinforcement learning algorithm on maize crop simulation model. This is done by using the soil parameters, crop parameters, weather data (temperature, probability of rainfall and solar radiation) on DSSAT (Decision Support System for Agro-technology Transfer) maize crop simulation model. Then the water consumption is minimized adaptively by using Q-learning to optimize the daily irrigation. Using simple modular approach of DSSAT to calculate daily yield and leaf area index along with the proposed reinforcement learning controller, almost 40% decrease in water consumption is achieved in comparison to constant irrigation method of the DSSAT model without any significant decrease in yield and leaf area index.
