Application of a convolutional neural network for automated multiclass identification of field-collected microplastics and diatom algae from optical microscopy images

Microplastics are present all around the globe, and they are a major threat to the environment because of the challenges they pose. Their sampling, isolation, and analysis processes are laborious and difficult due to their size, shape, and spreading dynamics. Furthermore, the lack of standardized protocols in microplastic research makes it difficult to compare results and unify the progress of the field. In this context, this work proposes and evaluates a model architecture based on deep learning to classify microplastic images using a dataset of real microplastics sampled from a freshwater reservoir, with convolutional neural network and transfer learning. Moreover, the model identifies diatom algae frustules, which can persist in the hydrogen peroxide degradation during the process of microplastic isolation due to their biosilica composition. The model was developed in Python using the Google Colab environment. A total of 1,140 images were used, and to ensure a robust and generalized evaluation, 5-fold cross-validation was applied. The model achieved 93% accuracy, with a recall of 97, 95, 92, and 90% for algae, microplastic filaments, fragments, and pellets, respectively. Overall, the accuracy of the model is encouraging considering the dataset size and all the challenges that involve the automatic identification of microplastics, with all their shape variations and nuances; thus the results are promising. To our knowledge, this is the first work addressing diatom presence after one of the most common microplastic isolation techniques and their automated classification among microplastics as well.