Cross-Dataset Band Selection and Model Generalization for Hyperspectral Plastic Classification

Abstract

Hyperspectral imaging enhances plastic classification, requiring robust bands and models across sensors, scenes, and acquisition conditions to enable scalable monitoring. However, most studies rely on single datasets or controlled conditions, and band-selection methods and classifiers often fail to generalize across heterogeneous domains. This study proposes a pipeline for identifying transferable spectral bands and assessing the limits of cross-dataset generalization across five hyperspectral datasets acquired with different sensors, environments, and object types. Using the Optimized Band Selection for Classification framework, together with multiple preprocessing strategies and leave-one-dataset-out training, band and model robustness are assessed under cross-dataset scenarios. Compact band subsets generalize substantially better than pretrained models, achieving 70%–80% accuracy in cross-sensor tests, with some cases reaching 85%, while direct model transfer typically remains below 40% except for closely matched datasets. Normalization combined with first derivative transformations improves generalization, raising some results to 70%–80%, but there are still some dataset-dependent limitations. Across experiments, the 1180–1220 nm and 1580–1680 nm regions consistently emerge as discriminative, with 1187, 1215, 1342, 1466, and 1589 nm proposed as strong multispectral candidates. The final filter centers and bandwidths should be selected according to the sensor response and atmospheric transmission. Freely available hyperspectral plastic datasets remain scarce and differ in postprocessing. Accordingly, this study provides practical guidelines for dataset harmonization and spectral-band standardization, including documenting acquisition conditions and ensuring sufficient spectral variability. The results of the study highlight that while cross-dataset model transfer remains limited, compact NIR–SWIR band subsets consistently generalize across sensors, enabling accurate and cost-effective multispectral plastic classification.