This study advances non-invasive screening of pesticide residues in olives by (i) extending visible–near-infrared (VNIR) hyperspectral imaging (HSI) from fruit to leaves (adaxial/abaxial) as easy-to-sample matrices; (ii) replacing hard-decision with probabilistic models, partial least squares discriminant analysis (PLS-DA) and extreme gradient boosting (XGBoost); and (iii) introducing a tunable reject-option via an indecision band to trade coverage for reliability. An intelligent pixel-selection pipeline (radiometric normalization, HSV masking of glare/defects, and mask-conditioned random sampling of \(\sim 10^4\) pixels per sample) was used. Ground-truth labels were obtained by GC–MS/MS against EU MRLs. On fruit, deltamethrin and diflufenican reached near-ceiling discrimination (AUC \(\approx 99\) –100%, F1 \(\approx 97\) –100%); on leaves, oxifluorfen and tebuconazole were best resolved (adaxial/abaxial AUC up to \(\approx 95\) –100% with F1 \(\approx 99\) %). In contrast, diflufenican shows reduced discriminability on leaf matrices, especially abaxial. Probabilistic outputs enabled an indeterminate label for low-confidence pixels identified via a model-specific indecision band, yielding high accuracy on accepted predictions while keeping the rejected share operationally manageable; an effective score (F1 \(_\text {eff}\) ), defined as a coverage-penalized F1 score, summarizes this coverage–reliability balance. Overall, coupling matrix selection (fruit vs. leaf side) with probabilistic modeling and abstention delivers robust laboratory performance and a realistic pathway to field deployment at cooperative intake, where indeterminate lots can be routed to secondary handling.
