Monitoring of salinity of water on the THA CHIN River basin using portable Vis-NIR spectrometer combined with machine learning algorithms

Abstract

• VIS-NIR spectroscopy has reliability in monitoring salinity of water on river. • The portable NIR instrument can be used to monitor salinity of water. • Machine learning is advantageous in monitoring salinity of water. • Extremely randomized trees algorithm displayed higher prediction performance. The goal of this work is to study the alternative practices for monitoring the salinity of water using a combination of portable vis-NIR spectrometer and machine learning approachs. Along 80 km of the Tha Chin River basin from the Gulf of Thailand, the data of salinity and NIR spectrum were collected during winter and summer seasons of Thailand. Salinity of water samples was measured by using a handheld electrical conductivity meter and NIR spectra was recorded with portable FQA-NIR GUN in the wavelength range of 600 to 1100 nm. The 10 machine learning models including partial least square regression (PLS), support vector machine (SVR), decision tree (DT), random forest (RF), adaptive boosting (AB), gradient boosting (GB), bagging meta-estimator (BME), extremely randomized trees (ERT), backpropagation neural networks (BPNN) and hybrid principal component analysis-neural network (PC NN) were applied to train the NIRs models for predicting salinity. All machine learning algorithms showed good prediction results which R p 2 values were higher than 0.84. The models built by tree-based algorithms (DT, RF, AB, GB, BME and ERT) displayed higher performances of calibration set and prediction set than those of PLS, SVM, BPNN and PC NN. Among these, the ERT algorithm showed the best performance R p 2 of 0.97, RMSEP of 0.41 g/L and RPD of 6.00. It was shown that NIR spectroscopy coupled with machine learning could be an alternative simpler way for predicting salinity of water.