RICE-GUARD

RICE-GUARD predicts rice blast onset by integrating real-time in-field environmental measurements and applying process-based (Yoshino, Water Accounting Rice Model - WARM) and machine learning (M5Rules, Recurrent Neural Networks) models to provide early-warning signals for management of the disease caused by the fungus Magnaporthe grisea (Pyricularia oryzae Cav.).

Key Features:

• Wireless sensor network: Low-cost measurement nodes capture site-specific environmental variables and transmit them in real time to a central server for analysis.
• Process-based models: Implements the Yoshino model and the Water Accounting Rice Model (WARM) for mechanistic prediction of rice blast risk.
• Machine learning models: Implements M5Rules and Recurrent Neural Networks (RNN) for data-driven prediction of disease onset.
• Comparative benchmarking: Conducts head-to-head comparisons between process-based and machine learning approaches across sites and seasons.
• Performance metrics: Reports normalized scores and quantitative metrics including correlation coefficient (r), determination coefficient (r²), and Mean Absolute Error (%MAE).
• Early-warning capability: Provides predictions focused on the critical early-warning period for rice blast onset to enable preventive actions.
• Site-specific adaptability: Demonstrates site-level variability in model performance, including notable results for Yoshino at Kalochori 2015.
• Management impact assessment: Supports evaluation of potential to mitigate yield losses and reduce fungicide use through early detection.

Scientific Applications:

• Early warning of rice blast: Predicts onset of disease caused by Magnaporthe grisea (Pyricularia oryzae Cav.) to inform timely interventions.
• Model selection and deployment: Compares process-based and machine learning models to guide operational choice of prediction methods.
• Site- and context-specific modeling: Enables adaptation of predictive models to specific sites and climatic conditions when local data are available.
• Decision support for disease management: Informs preventive actions aimed at reducing yield loss and fungicide applications.

Methodology:

In-field environmental variables are captured by a wireless sensor network of low-cost nodes and transmitted in real time to a central server; predictive models applied include Yoshino and WARM process-based models and M5Rules and Recurrent Neural Networks, with evaluation using normalized scores and metrics r, r², and %MAE and comparisons across sites including Kalochori 2015.