Developing spectral indices and models for monitoring rice planthoppers (Hemiptera: Delphacidae) with hyperspectral data and machine learning

Insect pests pose a significant threat to crop health including yield and quality, making population monitoring essential for effective pest management. Reflectance spectroscopy is a powerful tool for assessing crop health. Spectral characteristics of crops are closely linked to pest damage, yet it has not been widely used in pest monitoring. The rice planthoppers, Nilaparvata lugens (Stål), Sogatella furcifera (Horváth), and Laodelphax striatellus (Fallén) are serious pests of rice in China. This study focuses on developing spectral indices and models for monitoring these pests using hyperspectral remote sensing and machine learning. Reflectance from rice plants infested with planthoppers was examined and transformed into the relative reflectance to healthy plants. Three overlapping sensitive spectral bands (420 to 509 nm, 600 to 698 nm, and 728 to 986 nm) were identified across different planthopper species and rice growth stages, and the spectral indices, average relative reflectance in a successively sensitive band range, were developed. The infestation duration of planthoppers significantly influenced the average relative reflectance. Modeling methods including linear regression and machine learning, such as backpropagation neural networks (BPNN), support vector regression, categorical boosting, and adaptive boosting based on 3 average relative reflectance indices and infestation duration day, were developed to estimate planthopper density at tillering and booting stages. The BPNN model demonstrated a powerful ability to monitor planthoppers with the highest coefficient of determination and the lowest root mean square error for training and test datasets. A promising application of the novel spectral indices and BPNN model in intelligent monitoring systems for rice planthoppers was designed.