A Framework for Quantitative Riverine Sediment Provenance Modeling

Thesis Abstract

Riverine sediment provenance studies are often constrained by fragmented methods, insufficient spatial coverage, and uncertain source attribution, limiting the ability to diagnose sediment routing, catchment denudation rates, and anthropogenic impacts on sediment budgets. This study delivers a framework for quantitative riverine sediment provenance modeling that integrates geochemical, mineralogical, and isotopic tracers with probabilistic and machine learning approaches to improve source attribution and sediment-flow prediction. The aim is to develop a coherent methodology that yields transferable, multi-scale provenance estimates for mixed-sediment regimes typical of temperate to semi-arid river networks. Specific objectives are (i) to synthesize a tracer suite comprising major and trace elements, Sr-Nd-Pb isotopes, mineralogical indices, and particle-size distributions; (ii) to construct a hierarchical Bayesian mixing model calibrated with field- and laboratory-derived endmember signatures; (iii) to implement a data-fusion framework that reconciles intrinsic tracer variability with measurement uncertainties; (iv) to evaluate the framework against independent sediment-transport events to test predictive performance; and (v) to explore scenario-based futures to assess how land-use change and climate variability alter provenance signals. The methodology adopts a multiphase research design combining field sampling, laboratory analysis, and computational modeling. The population consists of riverine fine-grained sediments collected from ten river basins representing diverse lithologies and hydrologic regimes. A stratified sampling strategy yields 300 grain-size–fractionated samples, complemented by sources from potential tributary and bank-derived endmembers (n ? 120). Analytical techniques include inductively coupled plasma mass spectrometry (ICP-MS) for trace elements, X-ray fluorescence (XRF) for major elements, isotope ratio mass spectrometry (IRMS) for Sr-Nd-Pb isotopes, X-ray diffraction (XRD) for mineralogy, and laser diffraction for particle-size distribution. The core data-collection instruments are field portable XRF for rapid in-situ screening, high-precision mass spectrometers for tracers, and standardized sieve and Sedigraph protocols for granulometry. Validity and reliability are addressed via standardized calibration with certified reference materials, replicate analyses (n = 3–5 per sample), and cross-method validation among tracers. The analytical framework comprises (i) exploratory data analysis to identify discriminant tracers, (ii) a hierarchical Bayesian mixing model to integrate endmember signatures with measurement uncertainties, and (iii) a machine learning-augmented variance decomposition to quantify unknown-source contributions. Model specification follows a probabilistic graphical approach, incorporating prior information from regional geology and lithologic maps, with posterior inference conducted using Markov chain Monte Carlo (MCMC) sampling and convergence diagnostics (Gelman-Rubin statistic < 1.05). A complementary deterministic component uses a mass-balance approach to validate probabilistic outputs, while sensitivity analyses quantify the influence of tracer selection and endmember uncertainty. Geographic information system (GIS) tools are employed to map provenance fractions across spatial scales and to visualize source-to-sink sediment pathways. Expected findings include (i) a robust, transferrable tracer combination that distinguishes sources with overlapping geochemical signatures; (ii) quantified uncertainty bounds for endmember contributions, enabling risk-based interpretation of sediment budgets; (iii) improved accuracy of sediment provenance estimates compared with single-tracer or non-probabilistic approaches; and (iv) clear relationships between land-use change, hydrological regimes, and provenance signatures, revealing how anthropogenic and climatic drivers modulate sediment routing. The study contributes to knowledge by integrating tracers, Bayesian inference, and data-fusion techniques into a single coherent framework that can be adapted to other river systems, enhancing predictability of sediment provenance under future scenarios. The main conclusion is that quantitative riverine sediment provenance modeling benefits from a probabilistic, multi-tracer approach that accommodates endmember variability and measurement uncertainty, yielding more reliable source attribution and sediment budget assessments. Recommendations include standardizing tracer panels for regional programs, expanding the endmember database to capture seasonal variation, and applying the framework to coupled sediment-transport and morphodynamic models to inform watershed management, reservoir sedimentation planning, and flood-risk mitigation strategies.