<p>GPS coordinates of sampling stations.</p>

A U-Net model was trained on composite images to segment fluorescent particles. The workflow was applied to giant clams from three French Polynesian islands (Makemo, Hao, Tubuai), and NR-based estimates were validated against µFTIR spectroscopy. The model achieved F1-scores of 0.741 for giant clam samples and 0.657 for controls, comparable to human annotation (F1 = 0.680).

MPs were detected across all islands, with highest concentrations in gills (16.9–52.7 particles·g⁻¹ wet weight) compared to viscera (2.5–11.0 particles·g⁻¹ ww). µFTIR validation revealed that NR overestimates MP counts (µFTIR: 0.80 ± 0.16 particles·g⁻¹ ww at Tubuai), primarily due to false positives from proteins, cellulose, and stearates. In Tubuai, polyamide (28.9%), PVC (12.6%), and polystyrene (10.7%) were the dominant polymers, suggesting contributions from fishing gear, agriculture, and household waste.

While NR-based quantification overestimates absolute MP counts, the automated pipeline demonstrates potential for high-throughput image processing, reproducible sample analysis, and methodological standardization. This workflow represents a first step toward scalable, low-cost approaches for MPs monitoring in insular systems, highlighting areas for further calibration and optimization. Future work should refine fluorescence thresholds and expand validation across species and locations.</p></div>