Testing the sensitivity of a multivariate mixing model using geochemical fingerprints with artificial mixtures

Sediment source fingerprinting is increasingly used to provide insight into the dynamics of catchment sediment transfer processes, yet relatively few studies seek to validate source apportionments obtained from unmixing models. Our work focuses on simulating natural processes to test the accuracy of source apportionments obtained using a multivariate unmixing model called FingerPro. A relevant laboratory experiment is proposed to test the sensitivity of the model, using as experimental sediments 14 artificial mixtures composed of different proportions and numbers of sources selected from five soils as experimental sources. Twelve artificial mixtures were created by mixing a known proportion of source soils sieved to < 63 mu m in different proportions obtaining experimental sediments with three or four sources (experiment 1), while two additional artificial mixtures were prepared by combining mixing and sieving to obtain experimental sediments sieved to < 40 and < 15 lambda m (experiment 2). This research aims to test the sensitivity of the model by comparing the estimated source contributions for three sets of selected tracers (experiment 1) and for variations in particle size of the sources and mixtures (experiment 2). Experiment 1 show that source apportionments estimated by the FingerPro model for the same mixture reached maximum differences of 10% by using different tracers, with significantly different GOF and RMSE values between tracer sets (GOF means: 90% set A, 94% set B and 96% set C; RMSE means: 1.9% set A, 3% set B and 2.7% set C). Experiment 2 showed the inconsistency of model outputs when sources and mixtures had different particle size fractions. The accuracy of the model declined as the sediment become finer, and the mean RMSE increased from 2% to 4% up to 12% for mixtures at < 63, < 20 and < 15 mu m, respectively. The source apportionments estimated using a particle size correction factor improved slightly but not in all cases, with a maximum improvement of around one-third of the RMSE (mixture 10-B). Our results highlight the usefulness of employing artificial mixtures to test the accuracy of model simulations based on different tracer selections, source combinations and particle size fractions.