VNIR-NIR-derived multispectral imaging and explainable machine learning for sorting postharvest blueberries with asymptomatic spotted wing drosophila infestation

Spotted wing drosophila (SWD) is a zero-tolerance postharvest pest of blueberries that oviposits inside intact, ripening fruit, leaving no visible external symptoms during early infestation and creating a hidden risk to quality that current visual sortation cannot address. We characterized the evolution of the visible-near-infrared (VNIR, 400–1000 nm) and near-infrared (NIR, 900–1700 nm) hyperspectral signatures of SWD-infested blueberries over six storage days and translated them into a deployable multispectral configuration. Mid-level fusion of Successive Projections Algorithm (SPA)-selected wavelengths was evaluated using three classical classifiers, namely SVM, LDA, and PLS-DA, and a 1D-CNN. LDA achieved the highest fused test accuracy on every day, with a mean accuracy of 91.07% and 93.3% accuracy on Day 1. The 1D-CNN remained within 2.3 percentage points of LDA, which suggests that wavelength selection had a stronger influence than classifier architecture for the sample size used in this study. Cross-day generalization revealed a structural spectral transition between Days 4 and 5. This transition coincided with the larval-to-pupal developmental transition of SWD. Cross-day SHAP recurrence identified a 12-wavelength filter-wheel configuration at 770, 780, 920, 925, 950, 960, 1090, 1095, 1110, 1285, 1355, and 1370 nm. These wavelengths correspond to water O-H overtones or pigment red-edge absorbers. Two bands were VNIR-exclusive red-edge wavelengths at 770 and 780 nm, four were located in the VNIR-NIR overlap region from 920 to 960 nm, and six were NIR-exclusive water overtone bands from 1090 to 1370 nm. This configuration retained test accuracy between 80% and 91% across storage. Wavelength rankings showed agreement across three attribution methods on the LDA anchor, between SHAP-LDA and SHAP-CNN, with mean Jaccard@10 = 0.47, and across bootstrap seeds and LDA/CNN anchors.