Developing and evaluating automated deep learning and human-in-the-loop vision-language systems for microplastic characterization

Microplastic (MP) pollution poses escalating environmental risks, demanding efficient and reproducible tools for morphological characterization of plastic particles. Traditional manual microscopy is labour-intensive, operator-dependent, and poorly suited to large-scale monitoring. This study presents a comparative evaluation of two distinct artificial intelligence paradigms for the analysis of optical microscope images of microplastics. The first paradigm is a domain-specific, multi-task deep learning (DL) class…

Developing and evaluating automated deep learning and human-in-the-loop vision-language systems for microplastic characterization
Statistical profiles of particles’ size and shape for each plastic polymer found in plastic-bottled water by Authors of the study: Naixin Qian https://orcid.org/0000-0001-6433-063X , Xin Gao https://orcid.org/0000-0002-0911-3656 , Xiaoqi Lang, Huiping Deng, Teodora Maria Bratu, Qixuan Chen, Phoebe Stapleton https://orcid.org/0000-0003-3049-4868 , Beizhan Yan https://orcid.org/0000-0002-1321-779X yanbz@ldeo.columbia.edu, and Wei Min. CC BY 4.0 · https://creativecommons.org/licenses/by/4.0

In brief

Researchers compared a domain-specific EfficientNet-B0 multi-task deep learning classifier trained on about 700 annotated optical microscope images against a zero-shot Claude Vision API augmented with expert human-in-the-loop guidance. Both were tested on the same independent test set for predicting microplastic shape/type, color, and surface texture, with the DL model reaching F1-scores of 91.2%, 88.5% and 85.1% and the VLM improving from 72-81% to 84-89% after refinement.

Automated morphological characterization could reduce labor and operator dependence in microplastic pollution monitoring, enabling more reproducible large-scale environmental assessment. As of the August 2026 publication date, results are limited to a small in-house image set and a single test set, leaving open how well either approach generalizes to diverse field samples, imaging conditions, and broader monitoring workflows.

Main points

  1. Study compared two AI paradigms on identical independent test set using accuracy, macro-averaged precision, recall, and F1-score.
  2. Domain-specific multi-task deep learning classifier based on EfficientNet-B0 with transfer learning trained on approximately 700 annotated images to predict five shape/type classes, 10 color classes, and two texture classes.
  3. Zero-shot Claude Vision API augmented with structured human-in-the-loop mechanism allowing domain experts to provide targeted guidance for ambiguous particles improved from 72-81% raw F1 to approximately 84-89% after refinement.
  4. Comparative framework deployed as freely accessible web application via Hugging Face Spaces.

The gain

Domain-specific EfficientNet-B0 classifier and Claude Vision API with human-in-the-loop guidance automated morphological characterization of microplastics from optical microscope images, achieving high F1-scores for shape/type, color and texture.

The rundown

Traditional manual microscopy described as labour-intensive, operator-dependent, and poorly suited to large-scale monitoring, motivating automated alternatives. Study evaluated both systems on identical independent test set.

DL classifier excelled in high-throughput reproducible screening while VLM-HITL offered interpretability and flexibility for ambiguous cases. System made available via Hugging Face Spaces as practical deployment example.

Sources

  1. Peer-reviewedScientific Reports2026-08-25

The debate