Quantitative analysis of microplastics in soil based on two-trace two-dimensional infrared correlation spectroscopy and deep learning

A novel method for the quantitative analysis of microplastics (MPs) in soil was proposed by combining two-trace two-dimensional (2T2D) correlation spectroscopy with deep learning. The infrared (IR) spectra of soil-MPs mixtures were acquired and preprocessed (data augmentation, baseline correction, Savitzky-Golay filtering, standard normal variate, and multiplicative scatter correction), then transformed into five types of 2T2D spectra (auto-synchronous, auto-asynchronous, sy-asynchronous, cross-synchronous, and cross-asynchronous). To better capture the strip spatial correlation pattern features in 2T2D spectra and adaptively enhance information-rich fingerprint bands, a deep residual shrinkage network embedded with channel attention and strip attention (DRSN-SpCA) was developed for concentration prediction. The results indicate that the prediction accuracy of DRSN-SpCA (PE: R² = 0.9992, RMSE = 59.2 g kg^-1; PET: R² = 0.9977, RMSE = 90.9 g kg^-1; PP: R² = 0.9985, RMSE = 71.5 g kg^-1; PS: R² = 0.9945, RMSE = 160.3 g kg^-1) is significantly superior to that of conventional models. Regarding the 2T2D spectra, the sy-asynchronous spectra demonstrate the best overall predictive performance. This study demonstrates that the integration of 2T2D spectroscopy and the deep learning framework enables rapid and accurate detection of MPs concentrations in soil.