Implementation of Prioritized Experience Replay and Noisy Network on DQN for Fertilization Management in Soybeans

Smart agriculture using Machine Learning (ML) offers solutions by optimizing fertilizer use, improving crop productivity, and reducing environmental impact. However, traditional methods like Linear Regression and Neural Networks struggle with dynamic environmental changes, leading to suboptimal fertilization strategies. Reinforcement Learning (RL) has emerged as a promising alternative, enabling adaptive decision-making for agricultural management. Prior RL research has primarily focused on nitrogen fertilization in crops like corn and wheat, neglecting the balanced nutrient needs of soybeans. This paper introduces an RL-based approach using a Deep Q-Network (DQN) enhanced with Prioritized Experience Replay (PER) and Noisy Networks (NN) to provide adaptive fertilizer recommendations for urea, SP36, and KCL. The proposed model dynamically adjusts fertilization strategies based on real-time soil conditions, optimizing nutrient balance for soybean crops. Simulation results demonstrate that DQN with PER and Noisy Networks achieves an average reward 5% higher than standard DQN, outperforming other reinforcement learning methods like Advantage Actor-Critic (A2C) and Proximal Policy Optimization (PPO). Additionally, the action distribution obtained is more even and diverse, aligning with varying nutrient conditions. This research contributes to precision agriculture by enhancing fertilization efficiency and sustainability, potentially increasing soybean yield while minimizing excessive fertilizer use.