Toward reliable rapid morphological screening: A systematic benchmark of deep learning models for microplastic detection in mangrove filter images.

Microplastics are a global concern, but reliable high-throughput morphological candidate screening remains constrained by labor-intensive visual inspection and the limited availability of chemically supported image benchmarks. Here, we constructed an FTIR-supported image dataset from mangrove water and sediment samples and benchmarked 11 deep-learning detector configurations. The dataset contained 2865 unique FTIR-confirmed particles assigned to fragment, fiber, and film categories, together with background-only image tiles representing heterogeneous membrane residues and non-plastic interference. Among the evaluated configurations, YOLOv5m achieved the highest performance (F1 = 0.906; mAP@0.5 = 0.910). Model-derived particle counts showed a strong linear association with FTIR-confirmed counts (R2 = 0.92). Across 45 membranes, automated complete-image screening and counting with YOLOv5m required an average of 6.57 s per membrane. Performance was higher for in-distribution mangrove samples than for out-of-distribution samples, indicating that environmental background and particle heterogeneity remain important determinants of model transferability. Class-specific analysis further showed that fiber-related errors were dominated by missed detections rather than confusion with other morphological categories. The framework is intended as a morphological screening and triage tool before FTIR or Raman confirmation, providing a quick solution for microplastic monitoring in complex field samples.