Identifying and quantifying multiple pollution sources in estuaries using fluorescence spectra and gradient-based deep learning

Abstract

This study developed an intelligent method for identifying and quantifying water pollution sources in estuarine areas. It characterized the excitation-emission matrix (EEM) fluorescence spectra from seven end-members, including seawater, rainwater, and five pollution sources typical of these areas. A deep learning model was established to identify and quantify these pollution sources in mixed water bodies. The model was fed either the original EEM or a combined EEM and gradient input. The results indicated that the combined input enhanced classification and quantification accuracy; Although model accuracy declined with an increasing number of mixed pollution sources, the combined input still improved classification accuracy by 3.1 % to 6.8 %; When the proportion of rainwater and seawater was below 70 %, the model maintained a classification accuracy of 57.4 % with original input and 61.3 % with combined input, with root mean square error values for the pollution source proportion being 12.2 % and 11.4 %, respectively.