Effects of different LED light wavelengths on deep learning-based insect detection in arable crops

Approaches for automated image-based insect detection are rapidly maturing and becoming a reality for commercial use. However, complex image backgrounds, lighting variations and high variability between different insects’ physical appearance can reduce detection accuracy. Integrating artificial LED lighting into the image acquisition process can enhance the visibility of key insect morphological features, particularly in field applications, thereby improving feature extraction and overall detection performance of machine learning models. Achieving this requires the identification of light wavelengths that best reveal important insect characteristics, such as pigmentation, body patterns, and nanoscale-structural reflections, for more effective discrimination analysis and model training. We investigated the effect of five different LED wavelengths (red (600–700 nm), blue (400–500 nm), white (400–700 nm), UVA (315–400 nm) and a composite mixture of Blue, White, and red (BWR)) on insect identification and on the performance of deep learning-based detection methods in comparison with images taken in natural field conditions. Images of Colorado potato beetle on potato plants and ladybirds and aphids on potato, faba bean and sugar beet were collected with images comprising: (i) simple scenes containing a single insect and (ii) complex multi-class image scenes, with multiple insect individuals on a single image. Using the pre-processed datasets, two deep learning models, YOLOv11 and Swin transformer, were trained and evaluated. Results demonstrate promising potential on multi-class insect datasets, with the blue LED achieving the best detection performance (mAP@50 of 94.5 %) for adult ladybirds. Meanwhile, the composite mixture (BWR) produced the highest performance for both adult and larval Colorado potato beetles, with mAP@50 scores of 96.7 % and 94.4 %, respectively. The red LED performed best for pre-adult stages of ladybirds (84.2 % for eggs, 87.9 % for larvae and 95.3 % for pupae) and for eggs of Colorado potato beetles (88.0 %). For aphid detection, white LED light showed promising performance, achieving mAP@50 of 91.2 %, while the baseline outdoor imaging achieved an mAP@50 of 50.7 %. These results demonstrate that LED lighting enables better model learning and discrimination between different insect species and life stages. The use of LED lights demonstrates strong potential to improve model performance without increasing its architectural computational complexity, thereby facilitating future on-farm deployment of targeted pest management solutions.