Non-destructive detection of spotted wing Drosophila infestation in blueberry fruit using hyperspectral imaging technology

Abstract

Spotted Wing Drosophila (SWD) infestation in blueberries presents a significant threat to blueberry industries due to yield loss and quality safety issues during the postharvest process, where infested fruit is zero-tolerated. Current detection methods require destructive sampling, which is time-consuming and labor-intensive. Therefore, they are not suitable for high-volume inspection of individual products during postharvest handling. This study presents an innovative hyperspectral imaging-based approach to detect SWD infestation in highbush blueberry fruit. Two benchtop hyperspectral imaging systems in reflectance mode, operating in the visible-near-infrared (Vis-NIR, 400–1000 ​nm) and short-wavelength infrared (SWIR, 900–1700 ​nm) ranges, respectively, were in-house assembled for acquiring images of 945 (including 706 healthy and 235 infested) blueberry samples hand-picked from orchards. Hyperspectral imagery was processed to segment blueberries and extract mean spectra from individual samples. Infested blueberries showed lower spectral reflectance in the region of 750–1350 ​nm than normal samples. Baseline models were built using six different classifiers for sample classification, and the models based on partial least squares discriminant analysis (PLS-DA) yielded the best overall accuracy of 90.2 ​% and 92.5 ​% for the Vis-NIR and SWIR systems, respectively, with the corresponding recall rates of 74.2 ​% and 80.6 ​% for infested fruit. Three alternative modeling pipelines were proposed by implementing oversampling of the minority infested fruit class and waveband selection, through an exhaustive search for optimal methods, resulting in improved detection performance. Among the optimization strategies, oversampling proved more effective than waveband selection for enhancing model performance, and their combination (oversampling followed by waveband selection) yielded the best classification, with PLS-DA remaining the best classifier. The Vis-NIR and SWIR systems achieved the best overall accuracies of 93.7 ​% and 97.2 ​%, respectively, with the corresponding recall rates of 85.9 ​% and 95.7 ​% for infested fruit. This research showed that hyperspectral imaging, especially in the SWIR range, was useful for rapid, non-destructive detection of SWD infestation in blueberry fruit.