Modeling the Relationship between Catchment Attributes and In-stream Water Quality

Abstract

The physical attributes of catchments have a significant influence on the chemistry and physical features of in-stream water quality. Consequently, modeling this relationship is important for informing management strategies aimed at improving regional water quality. This study used a machine learning approach (Artificial Neural Networks or ANNs) to model the relationship between land use/cover, associated with other physical attributes of the catchment such as geological permeability and hydrologic soil groups, and in-stream water quality parameters (e.g., K+, Na+, Mg2+, Ca2+, SO4 2−, Cl−, HCO3 −, SAR, pH, EC, TDS). Eighty-eight catchments in the southern basins of the Caspian Sea were explored. To enhance the architecture of ANNs, the study applied backward elimination-based multiple linear regression, through which the optimum input nodes of ANNs can be determined amongst the most relevant variables. A transformation approach was also applied to qualify the performance of ANNs in four quality classes, ranging from unsatisfactory to very good. According to the findings, ANN based TDS model performance improved from unsatisfactory to very good. However, the linear regression-based pH model resulted in a decrease in performance, from “very good” to satisfactory. Moreover, among all catchment attributes, urban areas had the greatest impact on K+, Na+, Mg2+, Cl−, SO4 2−, EC and SAR concentration values. K+, TDS and EC were influenced by agricultural area (%). Bare land areas (%) had the largest impact on Na+, Ca2+ and HCO3 −. Assessing the performance of the ANN-based models developed in this study indicates that 10 out of 11 models had “very good” quality ratings and can be reliably used in practice.