Optimizing fungicide schedules for onion purple blotch with deep reinforcement learning.

Abstract

Purple Blotch disease, caused by the fungal pathogen Alternaria porri poses a significant threat to onion production in Uganda, causing yield losses of 30-50% and challenging smallholder farmers' livelihoods due to reliance on conventional, calendar-based fungicide applications. These traditional methods often result in excessive chemical use, escalating costs, environmental degradation, and increased fungicide resistance. This research introduces a new smart system using Deep Reinforcement Learning (DRL) to improve how farmers spray fungicides to fight Purple Blotch disease in onions. A custom OpenAI Gym simulation environment was developed, integrating an empirical epidemiological model, an onion growth model, and a fungicide effect model, parameterized with Ugandan meteorological and agricultural data. A Proximal Policy Optimization (PPO) DRL agent was trained within this environment to dynamically learn optimal fungicide application policies, balancing multiple objectives: maximizing onion yield, minimizing fungicide usage, reducing environmental impact, and enhancing economic returns. Extensive simulations compared the DRL system against three baseline strategies, no spray, fixed-interval spraying (every 7 or 14 days), and threshold-based spraying. Results demonstrated the DRL system's superior performance, achieving an average onion yield of 91 kg/ha (compared to 84 kg/ha for threshold-based and 38 kg/ha for no-spray strategies), reducing fungicide application by approximately 58% compared to fixed-interval methods, and significantly improving net economic returns. The system’s adaptive decision-making effectively controlled disease progression while minimizing unnecessary interventions. This study contributes a pioneering application of DRL to onion disease management, offering a scalable, data-driven framework for sustainable agriculture.