MiNa: A dataset of Micro- and Nanoplastic particles
MiNa: A dataset of Micro- and Nanoplastic particles
Plastic pollution is an escalating global problem, with microplastics and nanoplastics (MNPs) found in water, sediment, air and food. Manual analysis of these particles in microscope images is labor-intensive and slow, which motivates automated image analysis based on artificial intelligence (AI). We present the micro- and nanoplastics dataset (MiNa), an open-source collection of 105 scanning electron microscopy (SEM) micrographs containing 24,781 particles from 200 nanometres to 100 micrometres. These particles are generated from polyethylene, polystyrene, polypropylene and polyethylene terephthalate consumer products by controlled accelerated weathering, with every particle annotated by polymer type using an AI-assisted pipeline. We used MiNa to benchmark deep learning models in three experiments. Class-agnostic detection and instance segmentation networks, including convolutional networks and a detection transformer, located particles reliably. These methods reached a box average precision of up to 82.0% at 50% overlap. For polymer identification of isolated particles, deep image classifiers reached 57-59% macro-averaged F1 score, compared with 39-44% for classical machine-learning models trained on measured shape descriptors. Input ablations and gradient-weighted attention maps showed that this signal lies in fine-scale surface texture, and that SEM appearance alone is not sufficient for reliable polymer identification. Finally, prompting a segmentation foundation model with a fine-tuned detector outperformed zero-shot prompting and transferred to external SEM datasets without retraining. MiNa provides an open benchmark for automated MNP analysis under controlled conditions.
Two dataset configurations were derived from the SEM micrographs of MiNa: image patches for particle detection and segmentation, and single-particle images for polymer classification.
The particles in the SEM images are manually annotated using the V7 platform. Due to varying MNP concentrations, some micrographs are challenging and time-consuming to annotate. To expedite this process, we employ Segment Anything Model (SAM) guided annotation and thresholding methods.
We trained four detectors (Faster R-CNN with ResNet-101 and ResNeXt-101 backbones, RT-DETR-L and YOLOv26s) and four instance segmentation models (Mask R-CNN with ResNet-101 and ResNeXt-101 backbones, YOLOv8n-seg and YOLOv26n-seg). All models were trained to localize particles irrespective of polymer type (one class), and each was trained three times with different random seeds.
While transformers excel in large-scale domains, their application to MNPs is limited by the scarcity of labeled SEM datasets. We evaluated the transferability of SAM and micro-SAM (Archit et al., 2025) using Automatic Instance Segmentation (AIS) and Automatic Mask Generation (AMG). Although AMG provided higher surface accuracy, both zero-shot methods were outperformed by prompting SAM with bounding boxes from our fine-tuned RT-DETR model.
As a baseline, five machine-learning models: logistic regression k-nearest neighbors, support vector machine, random forest and gradient boosting were trained on six measured descriptors: equivalent diameter, area, aspect ratio, form factor, roundness and convexity. They were compared with six deep image classifiers that differ in architecture and pretraining: convolutional networks (ResNet-101, ConvNeXt-T, HGNetV2-B4, YOLOv8s-cls, YOLOv26s-cls) and a vision transformer pretrained with self-supervised learning (DINOv2 ViT-S/14). All deep classifiers were fine-tuned on MiNa with an identical training recipe.