NeuraMFs: A deep learning model for airborne microfiber identification in plant biomonitors

Airborne anthropogenic microfibers (A-MFs), including synthetic and industrially modified cellulosic fibers, are an emerging class of atmospheric contaminants of growing concern due to their persistence, inhalation exposure, and potential to transport associated chemicals. These particles are intercepted by plant biomonitors as mosses, lichens, and vascular plant leaves, but their reliable quantification remains a major methodological bottleneck because it relies on visual stereomicroscopic inspection of heterogeneous and organic-rich substrates. This paper introduces NeuraMFs, a deep-learning architecture based on a two-stage Cascade R-CNN (Region-based Convolutional Neural Network) with a ResNet-101 backbone, optimized for the identification, segmentation, and classification of microfibers extracted from plant biomonitors. A heterogeneous dataset of high-resolution stereomicroscope images, obtained from mosses, lichens, and vascular plants processed with multiple extraction protocols, was constructed to preserve fiber morphology and improve robustness across complex backgrounds. NeuraMFs showed high analytical performance, correctly identifying 109 out of the 110 microfibers in the adopted test set and achieving an AUC (Area Under Curve) of 0.978 with an color-classification accuracy of 0.895. By integrating test-time augmentation, the model enhanced recall for thin and low-contrast fibers and automatically extracted quantitative descriptors including color and length, enabling rapid screening and prioritization of samples for expert validation. In practical terms, NeuraMFs processes a typical sample in approximately 3 min, allowing analysts to quickly identify highly contaminated samples. These results demonstrate that deep learning can provide a rapid and scalable support tool for biomonitoring of airborne A-MFs, facilitating large-scale environmental exposure and risk assessment of hazardous atmospheric microfibers.