USP: Uncentered SVD with post-hoc SNV for background-decoupled NIR classification of plastic in aquatic environments

Scattering interference challenges near-infrared (NIR) hyperspectral classification of plastics in aquatic backgrounds and can reduce the robustness of conventional preprocessing pipelines under distribution shifts. We propose uncentered singular value decomposition (SVD) with post-hoc standard normal variate (SNV), hereafter termed USP, in which projection first concentrates a dominant broadband direction and sample-wise normalization is then applied in principal-component (PC) score space. On the main aquatic-plastic benchmark, USP provides strong clear-background performance, shows limited degradation under BandShift perturbations, and outperforms pre-SNV on foamy backgrounds. Its advantage is condition dependent, with SNV or SG1 matching or exceeding USP in specific classifier or external-library normal-condition evaluations. The dominance of the first principal component (PC1), which accounts for more than 98% of the measured energy, together with loading-vector and ablation analyses, supports the proposed broadband-background interpretation; experiments on two independent spectral libraries further assess its cross-dataset applicability. USP is therefore positioned as a robustness-oriented framework for distribution shifts rather than as a universal accuracy-enhancing method.