Integrating machine learning and change detection for enhanced crop disease forecasting in rice farming: A multi-regional study

Crop diseases are increasingly causing devastating yield losses each year, posing a significant threat to global food security. Currently, fungicide treatment is one of the most effective measures to mitigate these disease-induced yield losses. Accurately predicting disease occurrence, onset timing, and development is critical for optimizing fungicide application schedules to achieve high efficacy. In this study, we compiled weekly disease observations, crop management, and weather data from 207 locations across China, India, and Japan. We compared six machine learning (ML) models for their performance in simulating disease severity. By mimicking disease development curves and trends at early stages, we developed a novel change detection (CD) model, named rolling linear regression (RLR), and combined it with the best-performing ML model to predict the occurrence, severity, and onset dates of four major rice diseases: leaf blast (Magnaporthe oryzae), panicle-neck blast (Magnaporthe oryzae), sheath blight (Rhizoctonia solani), and false smut (Ustilaginoidea virens). Our findings showed that the RandomForestRegressor demonstrated the highest performance in simulating disease severity, with mean absolute errors (MAE) below 0.5 % and root mean square errors (RMSE) below 2.5 %. The RLR model showed a distinct advantage over the other four widely used CD models based on five evaluation metrics. Additionally, the paired RandomForestRegressor+RLR model achieved the highest F1-score, ranging from 0.7 to 0.8, in predicting disease occurrence, outperforming 29 other ML+CD model pairs. Furthermore, the RandomForestRegressor+RLR model predicted onset dates with fewer than 6 error days and accuracies ranging from 74 % to 87 %. Combining ML with CD models not only shows robust generalization across diverse environmental conditions but also proves highly effective for large-scale disease risk forecasting in rice farming regions. The adaptability of ML techniques, when sufficient training data are available, particularly enhances decision support systems aimed at optimizing rice disease management practices for growers across various regions. Our hybrid model thus presents a compelling advancement in the precision agriculture domain, with significant implications for improving disease management strategies for crops beyond rice through timely intervention. This approach can contribute to safeguarding global food security by reducing crop losses due to disease damage.