Mohamed Ebrahem, Mohamed I. Hosni, Adel Abdallah, Alaaeldin Mahmoud
Sea salt increasingly harbors organic contaminants from personal care products, yet current monitoring methods lack spatial resolution and require destructive sampling. This study introduces an innovative analytical framework integrating Laser-Induced Fluorescence (LIF) Hyperspectral Imaging (HSI) with machine learning for the rapid, non-destructive detection of sunscreen residues on salt crystals. To simulate contamination, seawater from the Mediterranean coast (Alexandria, Egypt) was spiked to achieve a 10 mg/L sunscreen concentration within the seawater matrix prior to crystallization; this formulation contained Ethylhexyl Methoxycinnamate, Homosalate, and Ethylhexyl Salicylate. A SOC710 HS camera (128 bands) acquired fluorescence data under 450 nm laser excitation. Raw data underwent preprocessing and dimensionality reduction via Sparse Principal Component Analysis (Sparse PCA, λ = 0.5, k = 4 components, 73.4% sparsity). A Support Vector Machine (SVM) with an RBF kernel was trained on these sparse features. Performance evaluation employed tenfold stratified cross-validation, an 80-20 holdout test on ROI-based spectra, and independent sample validation against manually annotated pixel-wise ground-truth masks. While ROI-based tests yielded near-perfect accuracy under ideal conditions, full-image evaluation achieved ≈96% pixel-wise accuracy (precision ≈ 0.99, recall ≈ 0.95, F1 ≈ 0.97), providing a realistic estimate under heterogeneous conditions. Full-image classification mapped widespread contamination (57.8% of pixels), whereas an independently prepared clean salt sample produced zero false positives. The integrated Sparse PCA-SVM framework transforms fluorescence-imaging data into spatio-chemical maps, simultaneously revealing contaminant presence and spatial distribution on salt surfaces, thereby offering a powerful paradigm for the interpretable monitoring of organic pollutants in food materials.