Deep convolutional autoencoders for the spectral reconstruction of filter-interfered FTIR spectra in microplastic...

The widespread contamination of aquatic environments by microplastics demands robust monitoring strategies, with Fourier transform infrared spectroscopy serving as a widely adopted technique for particle identification. However, the reliability of acquired spectra is frequently compromised by interference from filter substrates, especially when analyzing particles of very small sizes. This interference requires the reconstruction of high-fidelity spectra from distorted spectra to maintain the analytical integrity of environmental measurements. While traditional shallow convolutional autoencoders (CAEs) have been employed for this reconstruction task, their limited depth often fails to extract the complex latent features required to resolve high-dimensional spectroscopic data. In this work, we propose a deep CAE framework designed for high-fidelity spectral reconstruction to enhance the accuracy of microplastic assessment. We integrated the feature-extraction hierarchies of six state-of-the-art deep convolutional neural network architectures, including VGG16, Inception, ResNet50, Inception-ResNet, Xception, and MobileNet, as specialized encoders to facilitate multi-scale feature extraction. The performance of these models was benchmarked against a shallow CAE using Pearson correlation coefficients, root mean square error, and spectral angular mapping across varying signal-to-interference ratio (SIR) levels. Results demonstrate that deep encoder architectures substantially outperform shallow configurations. At an extreme SIR of -30 dB, the Xception-CAE emerged as the top-performing model, achieving a Pearson correlation of 0.9403 compared to 0.8531 for the shallow CAE. Principal component analysis and high-resolution class activation mapping confirmed the superior performance of the proposed models in preserving class separability and prioritizing key characteristic bands. Furthermore, we show that the reconstructed spectra markedly improved downstream classification performance. Finally, we validate the performance of the proposed deep CAE approach on real-world filter-interfered spectra. Overall, these findings establish that deep feature-extraction hierarchies provide a definitive guideline for CAE-based architectures tailored for spectral reconstruction and offer a tool for high-precision environmental monitoring and pollution assessment.