RL-FertNet: An intelligent agent for sustainable fertilization via sensor-driven and climate-aware optimization

Abstract

Precision agriculture demands adaptive and intelligent fertilization strategies to optimize crop yield while minimizing environmental impact. This paper presents a Reinforcement Learning (RL)-based framework for generating dynamic fertilization schedules by integrating real-time soil sensor data and weather forecasts. A Deep Q-Network (DQN) agent is trained to recommend optimal fertilizer dosages at appropriate times, with state representation derived from multi-modal inputs including soil nitrogen-phosphorus-potassium (NPK) levels, moisture content, temperature, rainfall probability, and crop growth stage. The model continuously learns from environmental interactions to minimize fertilizer waste and improve nutrient uptake efficiency. Results from simulations and field datasets indicate a 25% improvement in nitrogen use efficiency (NUE) and up to 15% yield gain compared to rule-based methods. The approach demonstrates the feasibility of deploying intelligent agents for sustainable and precision agriculture.