Farmer Behavior Augmented Reinforcement Learning for Smart Irrigation

Agriculture in semi-arid regions such as northern Karnataka faces persistent challenges related to water scarcity and inefficient irrigation practices. Conventional irrigation scheduling, based on fixed time intervals or simple threshold rules, fails to adapt to dynamic environmental conditions including soil moisture variability, evapotranspiration demand, and seasonal rainfall patterns. This work presents Farmer Behavior Augmented Reinforcement Learning for Smart Irrigation (FBARL) Management, an intelligent decision-support system that bridges the gap between traditional farmer knowledge and modern machine learning. The software component retrieves meteorological data for Ballari, Karnataka from the Open-Meteo archive API and simulates soil moisture dynamics using the FAO-56 water balance model to enable data-driven adaptive irrigation scheduling. Three agents are implemented and evaluated, a Rule-Based controller serving as a performance baseline, a standard Q-Learning agent, and the proposed Hybrid FBARL agent. The FBARL agent incorporates four novel mechanisms absent in conventional reinforcement learning, farmer prior Q-table initialization, Behavioral cloning warm-start from rule based decisions, constraint masking based on farmerdefined water budgets, and a farmer alignment bonus within the reward function. Comparative evaluation over 200 training episodes demonstrates that FBARL achieves a total reward of 154 units a 71.5 percent improvement over the rule-based baseline reward of 90 while consuming 255mm less water than plain Q-Learning and converging to near-optimal policy in approximately 60 episodes compared to over 150 for standard Q-Learning. A Streamlit-based graphical user interface provides adaptive irrigation dashboards, agent training controls, comparative result panels, and an interactive live demonstration mode with manual sensor override.