AgriScience Journal of Sustainable Agriculture and Agroecology
Reinforcement Learning and Deep Q-Networks for Optimized Irrigation Decision-Making in Dynamic Agricultural Environments
Abstract
Priyanka Balley and Kanchan K Doke
Efficient irrigation planning is essential in meeting the rising demand for water resources in agriculture with sustainability. Traditional irrigation systems are usually based on static heuristics or manual decision making that cannot respond to changing dynamic environmental conditions, including variability in soil moisture and weather uncertainties. This is an argument for the development of intelligent adaptive frameworks that would help optimize irrigation strategies in real timestamp scenarios. This paper discusses RL-based irrigation planning with a structured synthetic dataset mimicking realistic agricultural scenarios. The parameters included are soil moisture, crop type, weather forecasts, and irrigation needs for controlled experimentation and reproducibility. A customized RL environment is developed by including normalized soil moisture levels, weather probabilities, and irrigation demands as state representations and irrigation or no irrigation sets as actions. This dynamic irrigation model simulates the real-world situation with reward. It makes use of an agent rewarded for effectiveness of irrigation using a DQN-based feedforward neural network architecture consisting of two layers with 24 neurons, where each has ReLU activation. Training was done through epsilon-greedy to balance exploration and exploitation of Q Values iteratively with the use of the Bellman Process. The stability of training is attained using experience replay and gradient-based optimization with a mean squared error loss function. The performance metrics are cumulative rewards and epsilon decay to ascertain convergence and sets of learning efficacy for the model process. This paper shows that the proposed system, based on RL, greatly improves irrigation decision making and introduces policy stability with scalability levels.
