Deep learning-based prediction of trace silver concentrations in marine sediments from mid-infrared spectroscopy

Silver (Ag) is a biologically toxic element, and the rapid detection of its concentrations in environmental samples is thus of great importance. This study investigates mid-infrared (MIR) spectra of low-Ag sediments from the South China Sea (SCS), and successfully predicts Ag concentrations by three machine learning models, including Kennard Stone-competitive adaptive reweighted sampling-partial least squares regression (KS-CARS-PLSR), Kennard Stone-random forest (KS-RF), and random selecting-squeeze excitation attention-residual network (RS-SE-ResNet-1D). Among the three techniques, the RS-SE-ResNet-1D model performs best in terms of Ag prediction, achieving an R² of 0.82, an RMSE of 0.03, and an RPD of 2.38. A common feature wavenumber band in the range of 1408.3-1408.7 cm^-1, corresponding to the v₃ symmetric stretching absorption band of carbonate ions (CO₃^2-), is identified from all three models. This suggests that marine productivity is possibly the dominant factor controlling Ag levels in marine sediments of the SCS. The machine learning models based on MIR spectral data developed in this study provide a new analytical approach for determining Ag concentrations in low-Ag marine sediments, and offer new insights into the study of biogeochemical cycling of Ag.