Quantitative assessment of emulsified oil concentration based on near-infrared spectroscopy and multiple machine learning algorithms

Marine emulsified oil is formed from oil wastewater discharged by ships or through wind and wave action following marine oil spills. A rapid quantitative analysis method for emulsified oil concentration is therefore crucial for effective pollution cleanup and disaster assessment. A quantitative assessment method using near-infrared spectroscopy with kernel density estimation (KDE) combined with multiple machine learning algorithms is developed for measuring emulsified oil concentration. Five machine learning models are applied in this study: random forest (RF), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), support vector regression (SVR), and deep neural network (DNN). Laboratory measurements of near-infrared spectra are conducted on six emulsified oil samples using a mini-spectrometer. The characteristic band of the measured spectra is identified and selected, with results consistent with previous studies. The results demonstrate that KDE preprocessing significantly improves the predictive accuracy of all models, resulting in correlation coefficient (R²) values above 0.95 and relative improvements ranging from 5% to 35.1%. Notably, the RF model showed the most substantial improvement from 0.706 to 0.954. Moreover, the DNN model is able to achieve the most accurate prediction among the five machine learning models at the cost of more computation time. The XGBoost model, or the LightGBM model may be a favorable choice when training time is limited. The RF model and the SVR model are limited in prediction accuracy and computation time, respectively.