Development of fine-tuned deep learning models for physical classification of microplastics extracted from different environmental matrices

Microplastics (plastics < 5 mm) pose significant analytical challenges due to low-resolution imaging and labour-intensive classification susceptible to human errors. Henceforth, in this study, fine-tuned deep learning (DL) models (MobileNetV2, ResNet50, and InceptionV3) are applied to physically classify MPs extracted from diverse environmental matrices: personal care and daily-use products, wastewater, and rainwater samples. Very-Low-Magnification (VLM) MPs images were generated into three shapes (microbeads, fragments, and fibers) and physically classified using Convolutional Neural Networks (CNNs). Transfer learning (TL) methods were employed using a 7:1.5:1.5 split of the augmented dataset (training, validation, and testing), and their effectiveness was evaluated using validation dataset. The original set of 364 MPs images was augmented by applying horizontal and vertical flips along with zooming (1.1× and 1.2×) to yield 4368 samples and assess how larger training and validation sets affected CNNs performance. All the three DL models exhibited exceptional performance, achieving average accuracy, precision, recall, and F1 score values >99.8 %. MobileNetV2 obtained perfect scores (100 %) for microbeads but showed a slightly lower recall (99.6 %) for fragments. ResNet50 and InceptionV3 both demonstrated consistently high performance across all MP shapes, each achieving 100 % accuracy for fibers, with ResNet50 attaining a perfect F1 score and InceptionV3 showing high metrics overall. These results underscore the models' strong potential for accurate, efficient MP classification across diverse environmental samples.