Deep learning–based turbidity compensation for ultraviolet-visible spectrum correction in monitoring water parameters

Abstract

Ultraviolet-visible spectroscopy is an effective tool for reagent-free qualitative analysis and quantitative detection of water parameters. Suspended particles in water cause turbidity that interferes with the ultraviolet-visible spectrum and ultimately affects the accuracy of water parameter calculations. This paper proposes a deep learning method to compensate for turbidity interference and obtain water parameters using a partial least squares regression approach. Compared with orthogonal signal correction and extended multiplicative signal correction methods, the deep learning method specifically utilizes an accurate one-dimensional U-shape neural network (1D U-Net) and represents the first method enabling turbidity compensation in sampling real river water of agricultural catchments. After turbidity compensation, the R2 between the predicted and true values increased from 0.918 to 0.965, and the RMSE (Root Mean Square Error) value decreased from 0.526 to 0.343 mg. Experimental analyses showed that the 1D U-Net is suitable for turbidity compensation and provides accurate results.