GANsemble for Small and Imbalanced Data Sets: A Baseline for Synthetic Microplastics Data

TL;DR

This paper tackles the data scarcity and class-imbalance problem in microplastics research by introducing GANsemble, which automates augmentation strategy selection and leverages a class-conditioned GAN (MPcGAN) to generate synthetic, class-conditioned microplastics spectra data. A data chooser performs a factorial search over base augmentations to produce Aug^*, which then trains MPcGAN; a SYMP-Filter post-processes generated samples to improve quality. The authors establish baseline SYMP data quality using Fréchet Inception Distance and Inception Score, demonstrate the optimal augmentation size (50 samples per class) for best generalization, and show that Aug^* plus SYMP-Filter yields the most realistic and diverse synthetic data, enabling better model learning in small-sample, imbalanced settings. This framework provides a general approach to augmenting and generating synthetic data in data-scarce domains and sets baselines for future synthetic microplastics data generation.

Abstract

Microplastic particle ingestion or inhalation by humans is a problem of growing concern. Unfortunately, current research methods that use machine learning to understand their potential harms are obstructed by a lack of available data. Deep learning techniques in particular are challenged by such domains where only small or imbalanced data sets are available. Overcoming this challenge often involves oversampling underrepresented classes or augmenting the existing data to improve model performance. This paper proposes GANsemble: a two-module framework connecting data augmentation with conditional generative adversarial networks (cGANs) to generate class-conditioned synthetic data. First, the data chooser module automates augmentation strategy selection by searching for the best data augmentation strategy. Next, the cGAN module uses this strategy to train a cGAN for generating enhanced synthetic data. We experiment with the GANsemble framework on a small and imbalanced microplastics data set. A Microplastic-cGAN (MPcGAN) algorithm is introduced, and baselines for synthetic microplastics (SYMP) data are established in terms of Frechet Inception Distance (FID) and Inception Scores (IS). We also provide a synthetic microplastics filter (SYMP-Filter) algorithm to increase the quality of generated SYMP. Additionally, we show the best amount of oversampling with augmentation to fix class imbalance in small microplastics data sets. To our knowledge, this study is the first application of generative AI to synthetically create microplastics data.

Figures (4)

• Figure 1: Example of the polar conversion that allows DNNs to treat polymer spectra as images source_dataset.
• Figure 2: Determining the most effective size for purely augmented small microplastics data sets. The black dotted line represents averaged algorithm run-time after min-max normalization. The best performance is achieved with a pre-trained ResNet50 model using 50 augmented samples from each class. Augmentation strategy 1 was used to generate figure \ref{['PLATNICK.fig02.png']}.
• Figure 3: Qualitative results of SYMP data generated from MPcGAN instances trained on the top 3 strategies and baselines. We include SYMP samples from the first 4 classes. Boxes are drawn to highlight significant class-dependent features. The 20 samples were generated in a batch and not cherry picked. MPcGAN trained on data oversampled with $Aug^*$ results in the best SYMP in terms of ground truth resemblance.
• Figure 4: Qualitative results of our SYMP-Filter algorithm. Three synthetic images of Cellulosic were generated using a cGAN trained on $Aug^*$. The two images on the left contain unprocessed cGAN output, while the rightmost image shows filtered output.