Using bivariate linear mixed models to monitor the change in spatial distribution of heavy metals at the site of a historic landfill

摘要

To improve accuracy and efficiency of monitoring remediated sites, the current study proposed the use of bivariate linear mixed modelling and subsequent hypothesis testing to determine significant change in contaminant concentrations over time. The modelling method integrated soil heavy metal (arsenic-As, lead-Pb and zinc-Zn) concentrations obtained from Bicentennial Park, Sydney, Australia, in the years 1990 (n = 144) and 2015 (n = 60), alongside potential influencing factors as predictor variables. Following variable selection, significant predictors included As (1990)plan curvature, land cover change; As (2015)-multi-resolution ridge top flatness (MRRTF); Pb (1990)-elevation, MRRTF, type of nearest road; Pb (2015)-land cover change; Zn (1990)-distance to the nearest road and road type; and for Zn (2015)-aspect and land cover change. Model quality statistics (standardised squared prediction error; SSPE) indicated relatively good estimates of the prediction variance (mean similar to 1.0 for all metals, median = 0.512 for As (1990), 0.420 for As (2015), 0.417 for Pb (1990), 0.388 for Pb (2015), 0.342 for Zn (1990) and 0.263 for Zn (2015)), however Lin's concordance correlation coefficient indicated poor prediction of point estimates (LCCC = 0.263 for As (1990), 0.414 for As (2015), 0.250 for Pb (1990), 0.166 for Pb (2015), 0.233 for Zn (1990) and 0.408 for Zn (2015)). Pb in 1990 exceeded the Australian guide value of 600 mgkg(-1) in small, isolated areas of the park, and by 2015, these 'hotspots' had significantly diminished (P < 0.05). Concentrations of As were low in both 1990 and 2015, not exceeding the 300 mg kg(-1) guide; yet, in 2015, As had significantly increased in the south of the study area (P < 0.2). Zn concentrations in 1990 were elevated but did not exceed the guide value of 30,000 mg kg(-1). Overall, the models exhibited good estimation of prediction variance and therefore are suitable for hypothesis testing; however, they exhibited poor prediction quality at times. Despite this, bivariate linear mixed modelling is worth exploring as it provides an advantage over modelling single time points and can assist with tracking potential contaminant sources before they cause harm.