Single-wavelength near-infrared imaging and machine learning for detecting Queensland fruit fly damage in cherries

Abstract

Efficient detection and removal of infested fruits can be a valuable tool for reducing the spread of quarantine pests through trade. Automated grading technologies offer non-destructive solutions for detecting fruit fly infestations, though current optical methods face challenges due to either high computational demands (hyperspectral) or low specificity (multi- and single-spectral). In this study, we introduced a novel imaging method and machine learning approach to detect Queensland fruit fly (Qfly) infestations in fresh cherries, at both the image and fruit levels. Using hyperspectral imaging (HSI), we identified a wavelength of 730 nm within the visible to near-infrared (NIR) spectrum as most effective for distinguishing Qfly oviposition damage from natural pigmentation and mechanical damage. A library of 1771 high-resolution, single-wavelength NIR images was created, with Qfly oviposition sites manually labelled for model training. We proposed a novel machine learning approach called the Bounding Box Histogram Fusion Classifier (BBHFC). This method transforms spot-level predictions of Qfly oviposition damage, generated by a trained object detection model, into histogram-based feature vectors. These vectors are then used for efficient and accurate image-level infestation classification. BBHFC achieved high precision, recall, and F1 scores (all > 0.93), demonstrating the effectiveness of the approach. The proposed BBHFC outperformed traditional visual inspection, achieving over 89 % accuracy, compared to 60 % for manual detection. Integrating advanced imaging techniques into grading systems can significantly enhance biosecurity in horticultural industries by detecting and removing infested fruit. This technology could also supplement existing, costly, visual inspections of traded fruit that governments are required to undertake.