Abstrak

Microplastics have become pervasive pollutants that pose risks to biodiversity, ecosystem integrity, food safety, and human health. Most existing approaches for microplastic detection still rely heavily on advanced instrumentation, highlighting the need for simple, rapid, and accurate alternatives. In this study, we present a photoluminescence-based approach for quantifying polyethylene terephthalate (PET) microplastics using carbon dots (CDs) as fluorescent probes, coupled with image analysis, machine learning, and deep feature embedding. The green-channel photoluminescence intensity was identified as the most sensitive and robust descriptor, yielding a sensitivity of 3.647 ± 0.156 a.u. mg⁻¹ L, an excellent coefficient of determination (R² = 0.937), and favorable detection limits (LOD = 0.771 ± 0.030 mg L⁻¹; LOQ = 2.336 ± 0.099 mg L⁻¹). Machine learning models using color-intensity features further improved predictive accuracy, achieving R² values of 0.959 and 0.949 for linear regression (LR) and artificial neural network (ANN) models, respectively. SHAP analysis confirmed the dominance of green and grayscale channel intensities in microplastics quantification. Incorporating deep feature embeddings further enhanced performance, attaining excellent prediction (R² = 1.000 for LR; R² = 0.999 for ANN), underscoring the strong correlation between image-derived features and microplastic concentration. This work establishes a simple yet powerful optical–computational framework for microplastic quantification and demonstrates the potential of integrating photoluminescence imaging with artificial intelligence to enable fully automated, end-to-end detection systems.