Distributed crop disease detection with deep Q-network agent decision-making in a compressed state space

Crop disease detection is vital for reducing agricultural losses. Traditional centralized methods face challenges with data privacy and model performance. Distributed learning offers a promising alternative, but data imbalance across nodes in agricultural settings often limits its effectiveness. We propose a distributed crop disease detection method based on deep reinforcement learning. We design a node selection agent using a deep Q-network, which is able to make node selections through the current environment and historical experience, and we compress the agent state space in order to improve efficiency. During the interaction between the agent and the environment, the agent is rewarded when the global performance is improved, and vice versa, it is penalised. Through this continuous iterative interaction, the agent learns to select the strategy that can obtain the maximum future reward in different states, and such a strategy can alleviate the problem of slow convergence of the model and low performance caused by the data imbalance. With the continuous iterative Optimization of the strategy, the agent's state-action value function will gradually approach the optimal value, thus obtaining a better strategy. We simulate various real-world agricultural scenarios by applying multiple data augmentation methods to the PlantVillage dataset. Experiments show that by compressing the state space for the agent, the compression ratio ranges from 1.82 × 10 6 to 1.05 × 10 7 , which improves the computational efficiency. Compared to the baselines, our method improves accuracy and recall in 9 out of 10 crops, and enhances precision and F1-score across all 10 crops. On average, it achieves performance gains of 1.58% in highest accuracy, 1.42% in final accuracy, 1.72% in highest precision, 1.38% in final precision, 1.58% in highest recall, 1.42% in final recall, 1.67% in highest F1-score, and 1.49% in final F1-score. This work provides a new solution for crop disease detection in distributed scenarios, which plays a positive role in promoting the development of distributed smart agriculture. • Proposed a distributed crop disease detection method with deep reinforcement learning. • Designed a deep Q-network based agent for node selection strategy decision-making. • Designed a compression method for agent's high-dimensional state space to improve agent's decision-making efficiency. • Validated method effectiveness on nine single-crop and one mixed-crop datasets, including augmented datasets simulating real-world agricultural conditions.