Application of Domain Adaptation for Diagnostics of Natural Waters by Raman Spectroscopy

Determination of ionic composition and concentration of substances in river and waste water is an important scientific and practical task of environmental monitoring. In this paper, the problem of determining the concentrations of Zn2+, Cu2+, Li+, Fe3+, Ni2+, NH4+, SO42- и NO3- ions in aqueous solutions is solved using Raman spectroscopy and artificial neural networks (NN). Raman spectroscopy allows to perform rapid and remote analysis of multicomponent liquid media in real time. However, spectral analysis faces a number of problems due to the complex chemical composition of natural waters and the presence of fluorescent background caused by dissolved organic substances. Therefore, NN must be used. In the study, the transfer learning method is implemented for the purpose of domain adaptation of the NN, based on preliminary training of models on a representative sample of spectral data that do not contain a fluorescent component (source domain), and further additional training of the NN on spectra with a fluorescent background (target domain). This approach made it possible to increase the efficiency of model training and the accuracy of determining ion concentrations in real river waters (Moscow River, Yauza, Bitsa, Setun).