Synchronous Measurement Method of Seawater Temperature and Salinity Based on Raman Spectrum
Dong Rui,Ye Wangquan,Gui Bin,Chen Yu,Lu Yuan,Guo Jinjia,Zheng Ronger
DOI: https://doi.org/10.3788/aos231952
2024-01-01
Acta Optica Sinica
Abstract:Objective The monitoring of temperature and salinity(electrical conductivity)in seawater is of significant importance for understanding and predicting the responses of marine ecosystems,hydrological cycles,climate change,and the sustainable utilization of marine resources.The high spatial gradient characteristics of extreme environmental regions,such as hydrothermal or cold seep areas,pose new requirements for in situ measurements of temperature and salinity.Traditional conductivity,temperature,and depth(CTD)equipment,based on contact measurement,cannot achieve high spatial resolution simultaneous measurement of temperature and salinity,nor perform simultaneous measurement of temperature and salinity at a single point.It has been proven that the Raman spectrum of water exhibits a clear linear relationship between temperature and salinity.Raman spectrum can offer non-contact measurements and simultaneous detection of various water parameters.These capabilities provide the potential for measuring temperature and salinity in extreme submarine environments.In this study,we aim to achieve fast,accurate,and real-time in situ detection of seawater temperature and salinity using the Raman spectrum. Methods A 532 nm excitation optical setup(Fig.2)is established in the laboratory to acquire Raman spectra of OH bonds at different temperatures and salinities.Simulated seawater samples are prepared with varying concentrations of NaCl,and their salinities are measured using a conductivity meter.Temperature control is achieved using a Peltier-based cuvette holder for precise temperature regulation.A total of 170 sets of Raman spectra are obtained(Tables 1 and 2),divided into training and prediction sets at a ratio of 7∶3.The acquired Raman spectra are baseline subtracted and normalized for consistency.The Levenberg-Marquardt(L-M)method is employed to decompose the Raman spectra into five Gaussian peaks(Figs.3 and 4).The extracted peak heights,widths,and positions of these Gaussian peaks are used as training features,in conjunction with machine learning methods including partial least squares regression(PLSR),minimum absolute shrinkage and selection operator(LASSO)regression,support vector regression(SVR),and long short-term memory network with an integrated attention mechanism(LSTM+AM).To enhance predictive performance,a Stacking ensemble learning model is constructed using PLSR,SVR,LASSO,and multiple linear regression(MLR)as primary learners,with MLR serving as the secondary learner to simultaneously predict temperature and salinity.Evaluation metrics are utilized such as mean squared error(EMS),mean absolute error(EMA),and coefficient of determination(R2). Results and Discussions The OH stretching vibration peak spectra of water are compared at different temperatures under the same salinity and at different salinities under the same temperature.Figure 5 illustrates that the Raman shifts at 3170 and 3536 cm-1 exhibit the highest sensitivity to temperature,whereas 3195 cm-1 shows the highest sensitivity to salinity.Changes in spectral intensity demonstrate a clear linear relationship with both temperature and salinity.The OH stretching vibration peak is resolved into five sub-peaks,each of which also displays a robust linear relationship with temperature and salinity.Quantitative analysis is independently conducted using PLSR,LASSO,SVR,and LSTM+AM.LSTM+AM yields the best simultaneous predictions for temperature and electrical conductivity,with mean squared errors below 0.28 ℃ and 1.89 mS/cm,respectively(Fig.8).A subsequent Stacking model incorporating PLSR,LASSO,SVR,and MLR achieves even better quantitative results(Fig.9),with mean squared errors of 0.23 ℃ for temperature prediction and 1.63 mS/cm for electrical conductivity prediction. Conclusions The Raman OH stretching vibration peak of water molecules consists of multiple sub-peaks due to the local hydrogen bond network effect.Changes in temperature and salinity influence the hydrogen bond composition,thereby altering the spectral shape.Analysis of simulated seawater Raman spectra across varying temperatures and salinities reveals that OH sub-peak intensities exhibit clear linear relationships with temperature and salinity respectively.With the help of this linear relationship,the Raman spectrum proves capable of accurately measuring seawater temperature and salinity.The L-M algorithm decomposes the water peak into five sub-peaks corresponding to different hydrogen bonds,utilizing peak intensity,width,and other sub-peak characteristics for simultaneous calibration of temperature and salinity.While several traditional multivariate calibration methods are employed for simultaneous prediction,the LSTM+AM model outperforms them.To further strengthen accuracy and robustness,we use a Stacking ensemble learning model to integrate multiple base models such as PLSR,LASSO,and SVR during training.Quantitative results demonstrate the proposed method's effectiveness in simultaneously measuring water temperature and salinity.Mean squared errors for temperature and electrical conductivity are 0.23 ℃ and 1.63 mS/cm respectively.This method of using Raman spectrum for simultaneous prediction of seawater temperature and salinity holds promise for in situ Raman spectrum research,particularly in extreme deep-sea environments.