Diagnosis of Rice Blast Disease Using UAV-Based Multispectral Data: A Comparative Study of Feature Selection Methods

Rice blast is one of the most destructive diseases affecting rice throughout its entire growth cycle, posing serious threats to yield and quality. Timely, accurate, and non-destructive monitoring is therefore crucial for early warning and food security. In this study, rice blast inoculation experiments were conducted in 2019 and 2020 in Wanchang Town, Jilin Province, China. UAV-based multispectral imagery was collected at three key growth stages (booting, heading, and milk stages), and real-time kinematic (RTK) positioning was used to precisely locate ground sampling points where disease observations were recorded simultaneously. Spectral reflectance, vegetation indices, and texture features were extracted; four machine learning algorithms were applied to construct rice blast diagnostic models, and multiple feature selection methods were used to optimize the original high-dimensional feature sets. The main contributions of this study include establishing a multi-source feature system combining spectral, vegetation index, and texture information from multi-stage UAV imagery, proposing three new red edge indices tailored for rice blast detection, and systematically evaluating interactions between machine learning models and feature selection strategies. Results showed that the random forest model based on texture features performed best before feature selection (R 2 = 0.795, RMSE = 0.873). After feature selection, all models improved, with the gradient boosting decision tree model using the selected spectral - texture features achieving the highest accuracy (R 2 = 0.850, RMSE = 0.748). These findings demonstrate that integrating machine learning with appropriate feature selection can significantly enhance rice blast diagnosis and contribute to the development of precision agriculture.