Abstract The Colorado potato beetle ( Leptinotarsa decemlineata ) is a major pest of potato crops, causing yield losses of up to 20% and challenging sustainable production due to insecticide resistance and environmental risks associated with chemical control. Current pest detection methods, like RGB-based imaging, are limited by scalability, large dataset requirements, and reduced accuracy in detecting camouflaged pests. This study investigates hyperspectral imaging (937–1718 nm) coupled with machine learning for precise CPB detection on potato leaves. Hyperspectral images of 327 beetles were captured and a band selection method was developed to identify a minimal set of informative spectral bands based on spectral slope differences between Colorado potato beetle and potato leaves. This method was compared with established feature selection techniques (forward sequential feature selection, importance-correlation, and recursive feature elimination) using support vector machine, random forest, and k-nearest neighbour classifiers. Our approach achieved over 97.3% pixel-level F1-score using only four bands, outperforming the computationally intensive recursive feature elimination (3 days) in just 0.28 s, with superior accuracy across all classifiers. This level of precision was achieved using only four bands, which is crucial for real-time field applications. The identified key informative bands were 972, 1125, 1188, and 1286 nm. The findings suggest that a narrow band multispectral camera, designed using the informative bands, could operate in snapshot mode, facilitating real-time insect detection. This study shows the possibility of developing faster, more efficient hyperspectral systems for targeted insect detection, with the potential to revolutionize pest management by enabling more precise and environmentally friendly approaches.
Gebremeskel et al. (Tue,) studied this question.