Retrieval of Water Parameters from Absorption Spectrum Based on Convolutional Neural Network

In the field of water quality monitoring, the traditional chemical detection methods are time-consuming and laborious, and the operation process is complicated. The technology of absorption spectrum is used to retrieve water quality parameters, which has the advantages of less time-consuming, no secondary pollution, and easy operation. Under the condition of a small dataset, we utilize the convolutional neural network (CNN) combining channel attention module (CAM) to predict three water quality parameters, including suspended solids (SS), chemical oxygen demand (COD), and chromaticity. In this paper, the water quality parameters and corresponding absorption spectra of 50 groups of samples from the mainstream of the Yangtze River were measured by the chemical approach. At the same time, to solve the problem that it is difficult for the neural network to learn effective feature representation under the condition of the small dataset, we propose a loss function combining improved L1 Loss and mean absolute percentage error (MAPE). Our scheme has achieved good results in predicting SS, COD, and chromaticity. The determination coefficients (R ² ) of SS, COD, and chromaticity of the testing data are 0.93, 0.91, and 0.93 respectively. The results show that the method proposed in this paper can alleviate the over-fitting impact of neural network regression under the condition of small datasets, and at the same time improve the retrieving performance of water quality parameters.