Predicting rice blast disease: machine learning versus process-based models.

Area of Science:

• Agricultural Science
• Plant Pathology
• Computational Science

Background:

• Rice blast disease poses a significant threat to crop yield.
• Predictive modeling is crucial for timely disease management and reducing fungicide application.
• This study compares process-based and machine learning models for rice blast prediction.

Purpose of the Study:

• To compare the performance of two process-based models (Yoshino, WARM) and two machine learning models (M5Rules, RNN) for predicting rice blast.
• To evaluate the effectiveness of these models in providing early warnings for disease management.
• To assess the potential of machine learning as an alternative to process-based models in plant disease management.

Main Methods:

• Utilized in situ telemetry data for model development and validation.
• Implemented two process-based models: Yoshino and Water Accounting Rice Model (WARM).
• Developed two machine learning models: M5Rules (rule-based) and Recurrent Neural Networks (RNN) (neural network).

Main Results:

• All models successfully provided early warnings for rice blast onset and presence.
• Machine learning models (M5Rules, RNN) demonstrated competitive performance against process-based models (Yoshino, WARM).
• M5Rules achieved a maximum average normalized score of 0.80, closely followed by WARM at 0.77.

Conclusions:

• Both process-based and data-driven (machine learning) models can effectively provide early warnings for rice blast.
• Machine learning models present a viable and potentially more adaptable alternative to process-based approaches, especially with sufficient training data.
• These predictive models support informed fungicide applications, mitigating yield losses and reducing chemical use.