Improving the performance of an unmixing model in sediment source apportionment using synthetic sediment mixtures and an adaptive boosting algorithm

Abstract

Tracing technology is increasingly being used in fluvial and aeolian sediment provenance assessments. Using synthetic sample mixtures in validations of unmixing model performance is becoming a standard step in sediment source fingerprinting. With tracer variability fully considered in the sources and target sediment, this study explored a semiempirical modelling strategy based on virtual sediment mixtures and the adaptive boosting (AdaBoost) algorithm. The obtained integrated unmixing model contained multiple composite fingerprints, and the weight coefficient was obtained from the iterative process. The performance of the integrated unmixing model was compared with that of unmixing models applying single or equal-weighted multiple composite fingerprints. All generated virtual sediment mixtures (280) were split into a training dataset (240) and a test dataset (40) to validate the generalization ability of the models. The results showed that the integrated unmixing model achieved better performance than the unmixing models with a single composite fingerprint (basic models). The integrated unmixing model yielded an average mean absolute error (A-MAE) of 5.51% for training data and 5.72% for test data, and it achieved better accuracy than the best basic model (6.23% on the training and 6.75% on the test dataset, composite fingerprint C-1) or equal-weighted model (6.32% on the training and 6.41% on the test dataset, average of 42 basic models). The robustness of the model accuracy was also improved with the AdaBoost algorithm. The modelling approach proposed in this study has the potential to maximize the use of all tracer information and further improve the reliability of sediment fingerprinting.