Evaluation of the bias and precision of regression techniques and machine learning approaches in total dissolved solids modeling of an urban aquifer

• Published in: Environmental science and pollution research international Vol. 26; no. 2; pp. 1821 - 1833
• Main Authors: Pan, Conglian, Ng, Kelvin Tsun Wai, Fallah, Bahareh, Richter, Amy
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.01.2019; Springer Nature B.V
• Subjects: Aquatic Pollution; Aquifers; Artificial intelligence; Artificial neural networks; Atmospheric Protection/Air Quality Control/Air Pollution; Back propagation; Bias; Canada; data collection; Drinking water; Earth and Environmental Science; Ecotoxicology; Environment; Environmental Chemistry; Environmental Health; Environmental monitoring; Environmental Monitoring - methods; Environmental science; Extreme values; Groundwater; Groundwater - chemistry; Groundwater quality; guidelines; Irrigation water; Landfill; Landfills; Learning algorithms; Machine Learning; Mathematical models; Modelling; Models, Statistical; monitoring; Neural networks; Performance evaluation; polymerase chain reaction; Quality assessment; Quality control; Regression analysis; Research Article; Statistical analysis; Total dissolved solids; Waste disposal sites; Waste Water Technology; Water Management; Water Pollutants - analysis; Water Pollution - statistics & numerical data; Water Pollution Control; Water quality; wells; Canada; Principal component regression; Bias and precision; Total dissolved solids; Artificial neural network; Multivariate statistical analysis; Machine learning methods
• ISSN: 0944-1344; 1614-7499; 1614-7499
• DOI: 10.1007/s11356-018-3751-y

TDS is modeled for an aquifer near an unlined landfill in Canada. Canadian Drinking Water Guidelines and other indices are used to evaluate TDS concentrations in 27 monitoring wells surrounding the landfill. This study aims to predict TDS concentrations using three different modeling approaches: dual-step multiple linear regression (MLR), hybrid principal component regression (PCR), and backpropagation neural networks (BPNN). An analysis of the bias and precision of each models follows, using performance evaluation metrics and statistical indices. TDS is one of the most important parameters in assessing suitability of water for irrigation, and for overall groundwater quality assessment. Good agreement was observed between the MLR1 model and field data, although multicollinearity issues exist. Percentage errors of hybrid PCR were comparable to the dual-step MLR method. Percentage error for hybrid PCR was found to be inversely proportional to TDS concentrations, which was not observed for dual-step MLR. Larger errors were obtained from the BPNN models, and higher percentage errors were observed in monitoring wells with lower TDS concentrations. All models in this study adequately describe the data in testing stage ( R 2  > 0.86). Generally, the dual-step MLR and hybrid PCR models fared better ( R 2 avg  = 0.981 and 0.974, respectively), while BPNN models performed worse ( R 2 avg  = 0.904). For this dataset, both regression and machine learning models are more suited to predict mid-range data compared to extreme values. Advanced regression methods (hybrid PCR and dual-step MLR) are more advantageous compared to BPNN.