Enhancing microplastic classification through filter-interfered FTIR spectra using dimensionality reduction and deep learning in low-dimensional spaces

The increasing threat of microplastic pollution in aquatic environments underscores the urgent need for effective detection and classification methods. A widely used technique for identifying microplastics is Fourier-transform infrared (FTIR) spectroscopy, which characterizes their optical absorbance profiles, shown as spectra. However, the classification accuracy is often compromised by filter-interfered FTIR spectra, which arise from the sample's small size and the spectral interference introduced by membrane filters used during sample preparation. To address this challenge, we propose a framework that combines dimensionality reduction (DR) with deep learning (DL) classification to enhance microplastic classification based on filter-interfered spectra. The framework first applies a DR technique to convert high-dimensional spectra into low-dimensional representations that retain essential features while mitigating interference. These representations are then input into a one-dimensional convolutional neural network (CNN) based on the LeNet5 architecture to predict the microplastic type. We evaluate the proposed framework by applying five different DR techniques and assessing their impact on classification performance using a dataset comprising 22 microplastic types. Experimental results demonstrate that the proposed framework achieves classification accuracies ranging from 96.64% to 98.83%, which outperform a baseline approach directly using high-dimensional spectra (94.95%). Moreover, the number of trainable parameters in the LeNet5 model is reduced by over 98% when using low-dimensional inputs. These findings demonstrate the effectiveness of combining DR with DL for microplastic classification, which offers an efficient and enhanced approach for analyzing filter-interfered spectra in aquatic environmental monitoring.