Quantifying the sources of suspended sediments in the Nianchu River Basin, southern Tibetan Plateau: Evaluation and application of a semiempirical fingerprinting framework

Quantitative information on suspended sediment sources is an important aspect of clarifying soil erosion and sediment transport processes. Sediment source fingerprinting has been a principal method used to estimate such information, and the semiempirical fingerprinting strategy is a new research direction that requires more exploration. Taking the Nianchu River Basin as a case study, this study collected suspended sediment and source soil samples. Virtual sediment mixtures were utilized as samples to train the integrated un-mixing model under the framework of the AdaBoost.RT algorithm. The effectiveness and suitability of this semiempirical fingerprinting framework in improving model accuracy were evaluated. One main finding was that if a potential base learner, i.e., the mass conservative un-mixing model with a composite fingerprint, had a correlation coefficient of the mean absolute error (MAE) with the optimal base learner less than 0.7 (high diversity) and its average MAE ranking was before the first quartile among all base learners on the training dataset (high accuracy), the model accuracy could be effectively improved through the training process. Moreover, cropland close to the river channel with an area ratio of only 8.6% consistently contributed more than 50% of the suspended sediment in the Nianchu River Basin during the wet period. Moreover, the relative contributions from the cropland and grassland showed a negative and positive linear relationship with rainfall intensity, respectively, indicating increased sediment connectivity from grasslands to river channels under rainfall conditions. This study confirmed that the error-ambiguity decomposition theory is valid in the context of the semiempirical fingerprinting framework. The diversity and accuracy of the potential base learners should be the key aspects to simultaneously consider in determining whether to use a theoretical or semiempirical fingerprinting strategy.