Accuracy of mixing models in predicting sediment source contributions
Résumé
Determining the source of sediment using geochemical properties is now a widely used approach in catchment management. However the outcome of these studies often depends on the type of model used to determine the relative contribution from difference sources. Here we test the accuracy and robustness of four widely used sediment mixing models using artificial mixtures of three well-distinguished geologic sources. Sub-samples from these three sources were mixed to create four groups of samples, each consisting of five samples, with known source contributions, 20 samples in total. The source contributions to the individual and groups of artificial sediment mixtures were calculated using each of the four mixing models: Modified Hughes, Modified Collins, Landwehr and Distribution models. Unlike Modified Collins and Landwehr models which use calculated values from each tracer property of individual sources (e.g. mean and standard deviation), Hughes model uses the measured fingerprint property of replicated samples from each source and Distribution model incorporate distribution of tracers and correlation between tracer properties for sediment samples and sources. For the 20 individual sample mixtures the Distribution model provided the closest estimates to the known sediment source contribution values (Mean Absolute Error (MAE) = 10.8%, and standard error (SE) = 0.9%). The Modified Hughes (MAE = 13.5%, SE = 1.1%), Landwehr (MAE = 19%, SE = 1.7) and Collins models (MAE = 29%, SE = 2.1%) were the next accurate models, respectively. For the groups of the samples the Modified Hughes was the most robust source contribution predictor with 5.4% error. The Distribution model (MAE = 6.1%) and Landwehr model (MAE = 7.8%) were the second and third accurate models. Collins model with MAE of 28.3% was a significantly weaker source contribution predictor than the three other models. This study demonstrates the dependence of source attribution on model selection. The study highlight the need to test mixing model using known source and mixture samples prior to applying them to field samples. The results indicate that the Distribution and Modified Hughes models provided the most accurate source attributions using geochemical fingerprint properties.