Abstract
Heavy metals, possessing high toxicity and the ability to bioaccumulate, pose a serious threat to aquatic ecosystems and human health. Therefore, the development of effective methods for monitoring the concentrations of heavy metals in natural waters is one of the urgent tasks of modern ecology. This work addresses the problem of determining the concentrations of Zn \({}^{2+}\) , Cu \({}^{2+}\) , Li \({}^{+}\) , Fe \({}^{3+}\) , Ni \({}^{2+}\) , NH \({}_{4}^{+}\) , SO \({}_{4}^{2-}\) , and NO \({}_{3}^{-}\) ions in aqueous solutions using Raman spectroscopy and artificial neural networks (NN). To adapt the NN to the specifics of the spectra of real natural waters from the Moscow River, Yauza, Bitza, and Setun Rivers (presence of luminescent background, suspended large particles in the medium, etc.), transfer learning and domain-adversarial learning algorithms were implemented. The proposed approach using domain adaptation significantly improved the accuracy of determining the concentrations of the studied ions in all river water solutions, reducing the mean absolute error by an average of 50 \(\%\) . Testing the method on Raman spectra of real natural waters showed that the developed technology allows determining ion concentrations with errors satisfactory for environmental monitoring.
