Affiliation:
1. a Faculty of Architecture, ThuDauMot University, ThuDauMot 820000, Vietnam
2. b Faculty of Geology and Geography of Sciences University, Hue University, Hue 49118, Vietnam
3. c Department of International Business, National Kaohsiung University of Science and Technology, Kaohsiung 82445, Taiwan
Abstract
Abstract
This paper presents to study the performance of machine learning techniques consisting of multivariate adaptive regression spline (MARS), feed forward neural network-back propagation (FFNN-BP), and decision tree regression (DTR) for estimating the physico-chemical properties of groundwater in the coastal plain area in Vinh Linh and Gio Linh districts of Quang Tri province of Vietnam. With 290 groundwater samples collected in two districts, this study has identified three main elements CO2, Ca, CaCO3 for simulation. Quantitative analysis results have shown that these three components are such as CaCO3 with from 0 to 25.8 mg/l, Ca from 0 to 87.55 mg/l and CO2 from 0 to 12 mg/l. In the present examination, groundwater quality index (GQI) values and their representative categories have been referred by the Vietnam Groundwater Standard (QCVN01). Furthermore, the statistical accuracy parameters were used to compare among models. To deploy FFNN-BP and DTR, different types of transfer and kernel functions were tested, respectively. Determining the results of MARS, FFNN-BP and DTR showed that three models have suitable carrying out for forecasting water quality components. Comparison of outcomes of MARS model with the FFNN-BP and DTR models indicated that this model has good performance for forecasting the elements of water quality, its level of accuracy was slightly more than the other. To assess the accurate values of the models according to the measurement parameters for training phase illustrated that the order of the models was MARS to give the best result, followed by DTR and finally FFNN-BP, respectively.
Subject
Water Science and Technology
Cited by
1 articles.
订阅此论文施引文献
订阅此论文施引文献,注册后可以免费订阅5篇论文的施引文献,订阅后可以查看论文全部施引文献