Application of a modified set of GoogLeNet and ResNet-18 convolutional neural networks towards the identification of environmentally derived-MPLs in the Yadkin-pee dee river basin

Abstract

Abstract Microplastic (MPL) abundance is a well-elucidated problem in the marine environment but not so much in the terrestrial environment. In order to contribute to this research gap, a field study was performed in the Yadkin-Pee Dee River Basin. Due to their heterogenous nature and difficulty in characterization, a diverse set of pictorial training data from µ-Raman was used to perform transfer learning on 2 CNNs of interest: GoogLeNet (GN) and ResNet-18 (RN). In the first trial, using 10% of the initial training dataset, the CNNs exhibited high levels of accuracy rates, generally above 90%. Irrespective of spectroscopic mode, marginal improvements in accuracy rates were seen, with the best improvements occurring in the Raman-based models (U[GN(FTIR), RN(Raman), GN(FTIR), and RN(Raman)]: 39, 42, 38.5, and 34.5; p-value: 1, .6753, .9719, and .4978). However, for the external trial, pictorial data from Primpke (FTIR) and DongMiller (Raman) was predicted less accurately, with the largest loss occurring across the following sets: U[GN(Raman) and RN(FTIR)], 45.5 and 35; p-value:, .3268 and .5476. However, set RN fared marginally better, and due to the usage of µ-Raman, and its performance in the 10% trial, RN18_ADAM_.0011 was selected as the champion model for the field study data. In the unknown microparticle (MP) trial, generally, the most ID’d polymer type was CA, PET, and PE representing a relative concentration range for a given water source and area (MPL/MP) of 4.17–37.5%, 4.17–8.33%, and 4.17–8.33% for CNN and OpenSpecy (OS). A FEDS algorithm, equipped with natural and synthetic polymer standards and biological material, used to compare the strength of each model determined similar frequency in ascertaining positive MPL results across both models with corroboration between the CNN and OS around 1/3 of the time. Results indicate the models detect MPLs with similar frequency elucidating comparable strength of the CNN as well as a focus on particle type distribution rather than individual identification. Moreover, the largest influential factor in this study appears to be either laundry wastewater effluent or atmospheric deposition, which is stressed as a primary focus of remediating MPLs in similar freshwater environments. Lastly, it appears that these MPL are of primary origin as opposed to secondary in the oceanic and coastal environments.

Keywords

• Microplastics
• Primary microplastics
• Neural networks
• Artificial intelligence
• μ-Raman