SHAP-guided dimensionality reduction for hyperspectral imaging in plastic waste sorting.

Abstract

Hyperspectral imaging captures rich spectral information for polymer identification, but its dimensionality is a barrier to practical optical sensing. We propose a SHapley Additive exPlanations (SHAP)-based band-selection framework that identifies a minimal, physically interpretable set of wavelengths for compact multispectral sensors built from discrete bandpass filters. From 146 bands, 10 wavelengths were reproducibly identified across five image-wise cross-validation folds derived from 10 independently acquired scenes (173,355 labeled pixels; PET, HDPE, PP, and PS), at 88% mean pairwise selection overlap against ∼30% for sequential forward selection. The selected bands cluster in three C-H absorption windows near 1148-1218, 1368, and 1648-1683 nm. Random forest and a 1D-CNN reached 99.55±0.30% and 99.70±0.14% on the 10-band subset, retaining 99.9% of full-spectrum accuracy; because the 60 fragments, cut from approximately 20 parent objects, were re-imaged across scenes, these estimates reflect generalization to new acquisitions rather than to unseen specimens. Performing selection on spectrally aggregated channels yields a directly realizable specification of five 20 nm channels at 1146, 1205, 1365, 1665, and 1681 nm, achieving 99.24±0.32% with perfectly reproducible selection across folds, and grouping pixel predictions into objects by majority vote gives 100% object-level accuracy. Accuracy remains above 99% under 5% per-band Gaussian noise. These results indicate a pathway to hardware-efficient multispectral NIR sensing for plastic sorting, bound by the four polymer classes examined, the reused-sample protocol, laboratory illumination, and the single sensor platform.