Four-band narrowband CMOS imaging system optimized for robust marine litter classification.

Marine plastic debris seriously threatens ocean ecosystems; however, monitoring is hindered by the tradeoff between material diagnostic accuracy and cost. Although shortwave infrared sensors (1100-1700 nm) provide reliable polymer identification, their high costs limit their wide-scale deployment. Conversely, affordable RGB systems lack the spectral resolution required to distinguish plastics from natural materials such as wood. This study presents a four-band multispectral framework for optimized complementary metal-oxide-semiconductor (CMOS) sensors operating in the visible-to-near-infrared range (<1000 nm). Using a laboratory-grade hyperspectral library of 190 samples including weathered beach debris, virgin plastics, and coastal driftwood, we identified 798 nm as a robust reference wavelength for a novel data-driven normalization strategy. By maximizing the interclass separability, we identified three discriminative wavelengths (754, 784, and 826 nm) that effectively separate plastic from wood. The data acquired using this optimized four-band configuration enabled reliable material discrimination using conventional supervised classifiers, confirming that high performance can be achieved through spectral shape optimization rather than through complex model architectures. Practical applicability was validated using a custom-built CMOS prototype controlled by a Raspberry Pi, which successfully transferred laboratory-derived band selection to real-world debris collection. This framework demonstrates that plastic and wood can be accurately distinguished below 1000 nm, thereby providing a scalable and cost-effective solution for field-deployable marine plastic monitoring.