Estimation of nonlinear water-quality trends in high-frequency monitoring data

Recent advances in high-frequency water-quality sensors have enabled direct measurements of physical and chemical attributes in rivers and streams nearly continuously. Water-quality trends can be used to identify important watershed-scale changes driven by natural and anthropogenic influences. Statistical methods to estimate trends using high-frequency data are lacking. To address this gap, an evaluation of the generalized additive model (GAM) approach to test for trends in high-frequency data was conducted. Our proposed framework includes methods for handling serial correlation, trend estimation and slope-change detection, and trend interpretation at arithmetic scale for log-transformed variables. Water-temperature and turbidity data, representing two analytes with different temporal patterns, collected from the James River at Cartersville, Virginia. USA, were chosen for this analysis. Results indicated that the model, including flow, season, time covariates, and interaction between flow and season performed well for both analytes. The same model structure was applied to specific conductance data, collected from a small highly urbanized watershed, with satisfactory model performance. The water temperature GAM results indicated that the significant decreasing-then-increasing patterns after 2012 were mainly driven by air temperature changes. The turbidity trend was not significant over time. The specific conductance results showed a consistently upward trend over the last decade due to ever-increasing urbanization in the small watershed. This study suggests that the GAM method has great potential as a useful tool for trend analysis on high-frequency data, and for informing watershed managers of hydro-climatic and human influences on water quality by detecting crucial signal variation over time. (C) 2020 The Authors. Published by Elsevier B.V.